" }`, and `"account": "<account>"` in the tool entry when you chose one. Note the `spreadsheetId`. Note the first tab's `sheetId` if the response shows it (normally 0; check with `GOOGLESHEETS_GET_SPREADSHEET_INFO` if unsure). 3. `lr.py sheet-plan --run <dir> --spreadsheet-id <ID> [--sheet-id <first tab id>] [--account <account>]`. This writes numbered plan files to `<run>/export/gsheets/`. Without `--account`, it uses `LR_AI_GOOGLE_ACCOUNT` when that is set. 4. Execute the files in order. Read each file and call `COMPOSIO_MULTI_EXECUTE_TOOL` once with `tools` = the file's `tools` array, copied verbatim, plus `session_id` and `sync_response_to_workbench: false`. Calls within a file are independent. - If calls fail with rate limits (Sheets allows about 60 writes per minute), wait about 60 seconds and re-send only the failed items. - If a call fails on its arguments, check the schema with `COMPOSIO_GET_TOOL_SCHEMAS`, fix that call, and continue. - With many `values_*` files (large reviews), delegate executing all plan files to one general-purpose subagent. Give it the file list and these rules, to keep the main context small. 5. Give the user the sheet URL (also saved in `<run>/sheet_url.txt`). ## 3. Without Composio Tell the user to upload `papers.xlsx` to Google Drive and open it with Google Sheets (or use File → Import in Sheets). Colors, dropdowns, frozen panes and filters carry over. 22:o8,28:Te6b,22:oe6b,--- name: screen-papers description: Filter and screen literature-review candidates - apply deterministic filters (year, language, type, citations, keywords) and then title/abstract screening by parallel subagents that label each paper core, skim or exclude. Use when the user wants to filter, screen, or narrow down the collected papers against inclusion criteria. allowed-tools: - Bash(python "${CLAUDE_PLUGIN_ROOT}/scripts/lr.py" *) - Bash(python3 "${CLAUDE_PLUGIN_ROOT}/scripts/lr.py" *) - Edit(lr-runs/**) --- # Filtering and screening (step 4) CLI: `python "${CLAUDE_PLUGIN_ROOT}/scripts/lr.py" <command> --run <run dir>`. Run it with the Bash tool exactly as written, so it needs no permission prompt. `lr.py <command> --help` lists the options. If the path still contains a `$` placeholder, the plugin root is two folders above this skill's base directory. Automatic by default: make the choices below yourself and keep going. Lines marked **Checkpoint** apply only in step-by-step mode (the user asked to confirm steps, e.g. `/lr-ai:lit-review --step`): there, stop and wait for the user's reply. ## 1. Deterministic filters Run `lr.py filter`. It recomputes `excluded_reason` from the `filters` in `config.yaml` every time and prints how many papers await screening. Screening takes about one subagent per 25 papers, so keep it to roughly 400 papers unless the user gave another limit. If more remain, tighten the filters in this order, re-running `filter` after each change, until the rest fits: 1. `min_hits: 2`: snowballed papers must be linked from at least 2 start-set papers 2. `require_any`: 3-8 core terms of the topic, with `*` for word stems, e.g. `["citation*", "bibliometric*"]` 3. `min_citations` (e.g. 5); `citation_grace_years` keeps recent papers exempt Keep track of what you tightened for the final report. **Checkpoint:** show the counts and propose the stricter filters instead of applying them yourself. Papers the user added (`seed.papers`, `add`) are never filtered out. ## 2. Screening with subagents If the request adds or changes criteria, update `inclusion_criteria` / `exclusion_criteria` in `config.yaml` first. The batch files carry them to the screeners. To re-screen papers that were already screened, run `batches` with `--force`. 1. Run `lr.py batches --stage screen`. It writes batch files of `screening.batch_size` (default 25) papers each and prints their paths. 2. Launch one `lr-ai:paper-screener` subagent per batch file, in parallel. Send them in a single message, up to about 10 at a time. Prompt each one with just the batch file path. The criteria and research questions travel inside the file. 3. Run `lr.py apply --stage screen`. If it lists batches with missing results, re-launch agents for those batch files only, then `apply` again. Status meanings: **core** means it directly addresses a research question and meets the inclusion criteria. **skim** means it's related, background, or partially relevant. **exclude** means it's off-topic or hits an exclusion criterion. ## 3. Results Report the core/skim/exclude counts in one line. If no paper was included, stop and tell the user: the criteria or the queries are probably off. **Checkpoint:** show the core papers (`lr.py show --where status=core --fields year,citations,title,reason`) and a few borderline ones (`--where status=skim`). Fix disagreements with `lr.py set <id> status=core|skim|exclude status_reason="..."`. Manual edits are permanent: later `apply` runs never overwrite them. If the user wants more rounds, go back to the `snowball` skill with `--from included`. Next step: categorization and summaries (skill `categorize-papers`). 22:o8,29:Tcb7,22:ocb7,--- name: seed-search description: Start a literature review - set up the review (topic, research questions, inclusion/exclusion criteria) and find an initial set of seed papers via Semantic Scholar queries and known DOIs/arXiv ids. Use when the user wants to begin a literature review, find initial or seed papers for a topic, or set up LR-AI. allowed-tools: - Bash(python "${CLAUDE_PLUGIN_ROOT}/scripts/lr.py" *) - Bash(python3 "${CLAUDE_PLUGIN_ROOT}/scripts/lr.py" *) - Edit(lr-runs/**) --- # Seed search (step 1) CLI: `python "${CLAUDE_PLUGIN_ROOT}/scripts/lr.py" <command>`. Run it with the Bash tool exactly as written, so it needs no permission prompt (use `python3` if `python` is not Python 3). `lr.py <command> --help` lists the options. Always pass `--run <run dir>` once a run exists. If the path still contains a `$` placeholder, the plugin root is two folders above this skill's base directory. Automatic by default: make the choices below yourself and keep going. Lines marked **Checkpoint** apply only in step-by-step mode (the user asked to confirm steps, e.g. `/lr-ai:lit-review --step`): there, stop and wait for the user's reply. ## 1. Check the environment (once per session) Run `lr.py doctor`. If `yaml` or `openpyxl` is missing, install them with `python -m pip install -r "${CLAUDE_PLUGIN_ROOT}/requirements.txt"` (Claude Code asks the user to approve it). Mention missing API keys in one line and continue. They're optional: LR-AI ships a shared OpenAlex key. If any command prints an `OPENALEX RATE LIMIT` notice, show it to the user word for word. ## 2. Frame the review Take the topic, research questions, criteria, year range and known papers from the user's request, and decide the rest yourself: - 1-3 research questions (RQ1, RQ2, ...) that the topic implies - inclusion and exclusion criteria that follow from them, for example "proposes a method for X" or "not peer-reviewed" - no year limit unless the request implies one **Checkpoint:** ask the user for what the request left out instead of deciding it yourself. Then run `lr.py init "<topic>" --name <short-name>` and fill `config.yaml` in the new run folder: `research_questions`, `inclusion_criteria`, `exclusion_criteria`, `filters.year_min` / `year_max`. ## 3. Seed queries Draft 3-6 natural-language queries that cover the RQs from different angles: synonyms, the method side and the application side. Put them in `seed.queries`, and known papers in `seed.papers`. **Checkpoint:** show the queries and let the user edit them. If scholarly-search tools are available in this session (e.g. an alphaXiv / Consensus / Semantic Scholar MCP, or web search), use them too. Put the DOIs or arXiv ids of highly relevant papers they return into `seed.papers`. This often gives better seeds than keyword queries. ## 4. Run `lr.py seed --run <dir>`. Each query falls back from Semantic Scholar to OpenAlex to arXiv if a service is rate-limited. Look over the returned list and exclude clearly off-topic seeds: `lr.py set --run <dir> <id> status=exclude status_reason="off-topic seed"`. Excluded seeds are not snowballed. **Checkpoint:** show the seed papers, grouped by relevance, before moving on. Next step: database search (skill `database-search`). 22:o8,2a:T8c5,22:o8c5,--- name: setup description: Check or set up LR-AI, the literature-review plugin - run its doctor (Python packages, API reachability, API keys), install its Python dependencies, and explain the optional API keys. Use when the user asks to check, diagnose, test, install or configure LR-AI, or asks for the LR-AI doctor. allowed-tools: - Bash(python "${CLAUDE_PLUGIN_ROOT}/scripts/lr.py" *) - Bash(python3 "${CLAUDE_PLUGIN_ROOT}/scripts/lr.py" *) --- # LR-AI setup and health check The tool is `python "${CLAUDE_PLUGIN_ROOT}/scripts/lr.py"`. Use `python3` if `python` is not Python 3. Don't search the disk for it; this is the command. Run it with the Bash tool exactly as written, so it needs no permission prompt. If the path still contains a `$` placeholder, the plugin root is two folders above this skill's base directory. 1. Run `python "${CLAUDE_PLUGIN_ROOT}/scripts/lr.py" doctor`. 2. If `yaml` or `openpyxl` is missing, install them with `python -m pip install -r "${CLAUDE_PLUGIN_ROOT}/requirements.txt"` (Claude Code asks the user to approve it). 3. Report the result in a few lines. A ⚠ next to a database is not fatal: LR-AI falls back to the other databases. Mention the optional environment variables once: - `OPENALEX_API_KEY`: free at openalex.org. LR-AI ships a shared key; a personal key has its own daily budget, so set one if OpenAlex reports a rate limit. - `S2_API_KEY`: free on request from Semantic Scholar. Gives a dedicated rate limit. - `OPENALEX_EMAIL`: identifies the user to OpenAlex (polite pool). - `LR_AI_GOOGLE_ACCOUNT`: the alias or id of the Composio Google account that gets the sheets, when several are connected. Otherwise the first connected account is used. To set one permanently on Windows: `setx OPENALEX_API_KEY "<key>"`, then restart Claude Code. On macOS/Linux: add `export OPENALEX_API_KEY=<key>` to the shell profile. 4. Point to the next step: `/lr-ai:lit-review <topic>` runs a whole review without stopping, and `/lr-ai:lit-review <topic> --step` pauses for approval after each step. To run and check phases one at a time, see `${CLAUDE_PLUGIN_ROOT}/docs/phases.md`. The full command reference is in `${CLAUDE_PLUGIN_ROOT}/docs/cli.md`. 22:o8,2b:Tba6,22:oba6,--- name: snowball description: Forward and backward citation snowballing for a literature review - collect the references of each paper and the papers that cite it (via OpenAlex, falling back to Semantic Scholar), for one or more rounds. Use when the user asks to snowball, expand a paper set through citations/references, or find related work from a start set. allowed-tools: - Bash(python "${CLAUDE_PLUGIN_ROOT}/scripts/lr.py" *) - Bash(python3 "${CLAUDE_PLUGIN_ROOT}/scripts/lr.py" *) - Edit(lr-runs/**) --- # Snowballing (step 3) CLI: `python "${CLAUDE_PLUGIN_ROOT}/scripts/lr.py" <command> --run <run dir>`. Run it with the Bash tool exactly as written, so it needs no permission prompt. `lr.py <command> --help` lists the options. If the path still contains a `$` placeholder, the plugin root is two folders above this skill's base directory. Automatic by default: make the choices below yourself and keep going. Lines marked **Checkpoint** apply only in step-by-step mode (the user asked to confirm steps, e.g. `/lr-ai:lit-review --step`): there, stop and wait for the user's reply. ## No run yet If the user just names papers to expand (DOIs, arXiv ids or titles), create a run first: `lr.py init "<topic>" --name <short-name>`, then `lr.py add --run <dir> <paper> <paper> ...`. Papers added this way are never filtered out. ## Round 1 (usually enough) 1. Run `lr.py filter` first, so papers outside the year range or excluded types are not expanded. 2. Adjust `snowball.max_references` and `max_citations` if needed. Both keep the most-cited papers first. Defaults are 200 references and 50 citing papers per paper. 3. Run `lr.py snowball --run <dir>`. Each start paper costs about 2-3 API calls. The command reports references and citing papers found, new unique papers, and which backend served each paper. 4. If it refuses because more than 300 papers would be expanded, don't pass `--force`. Snowball from the included papers only, which is the standard Wohlin guideline: set `snowball.from: included`, screen first (skill `screen-papers`), run `snowball` again, then screen the new papers (`screen-papers` again; only unscreened papers go to the screeners). ## More rounds Screen the new papers first (`filter`, then the `screen-papers` skill). Then run `lr.py snowball --from included` to expand only the included papers from the previous round. The round number increments automatically. Stop when a round adds few new included papers (saturation). ## Notes - Every paper records who led to it (`found_via`: `backward:<parent>` / `forward:<parent>`). Papers linked from several start-set papers are strong candidates: `lr.py show --sort hits`. - If OpenAlex's daily budget runs out, the command switches to Semantic Scholar and prints an `OPENALEX RATE LIMIT` notice. Show it to the user word for word: it says what happened and how to set their own `OPENALEX_API_KEY`. Next step: filtering and screening (skill `screen-papers`). 22:o10b2,11:[["$","script",null,{"type":"application/ld+json","dangerouslySetInnerHTML":{"__html":"{\"@context\":\"https://schema.org\",\"@type\":\"SoftwareApplication\",\"name\":\"LR-AI\",\"description\":\"Automated literature reviews: seed search, database keyword search, forward/backward snowballing, deduplication, filtering, LLM screening and categorization, exported as a styled Google Sheet (via Composio) and Excel file.\",\"applicationCategory\":\"Research & strategy\",\"operatingSystem\":\"Claude\",\"url\":\"https://agentready.co/connectors/lr-ai\",\"sameAs\":[],\"publisher\":{\"@type\":\"Organization\",\"name\":\"yazangthb\",\"url\":\"https://github.com\"}}"}}],["$","div",null,{"className":"wrap","children":[["$","div",null,{"className":"pp-head","children":[["$","$L23",null,{"name":"LR-AI","url":null,"size":56}],["$","div",null,{"children":[["$","div",null,{"className":"crumbs","children":[["$","$L5",null,{"href":"/","children":"Connectors"}],[" / ",["$","$L5",null,{"href":"/categories/research","children":"Research & strategy"}]]]}],["$","h1",null,{"children":["LR-AI",null]}]]}]]}],["$","div",null,{"className":"pp","children":[["$","div",null,{"className":"ca","children":[["$","p",null,{"className":"tag","children":"Automated literature reviews: seed search, database keyword search, forward/backward snowballing, deduplication, filtering, LLM screening and categorization, exported as a styled Google Sheet (via Composio) and Excel file."}],null,["$","$L24",null,{"examples":[],"tools":[],"prefix":"","server":null,"instructions":[],"skills":[{"name":"categorize-papers","label":"Categorize papers","gist":"Build a two-level taxonomy for the included papers of a literature review and have subagents write summaries, research-question relations, categories and reading priorities for each paper.","source":"https://github.com/yazangthb/lr-ai/blob/47cd39316b710f0ada2b4f6302fa877f2e065ec8/skills/categorize-papers/SKILL.md","raw":"$25"},{"name":"database-search","label":"Database search","gist":"Run boolean keyword searches for a literature review across scholarly databases (OpenAlex, Semantic Scholar, arXiv) and merge the results with deduplication.","source":"https://github.com/yazangthb/lr-ai/blob/47cd39316b710f0ada2b4f6302fa877f2e065ec8/skills/database-search/SKILL.md","raw":"$26"},{"name":"export-sheet","label":"Export sheet","gist":"Export a literature review to a styled, categorized Google Sheet (via Composio's Google Sheets tools) and an Excel file - grouped category rows, color-coded status and priority, reading-status dropdow","source":"https://github.com/yazangthb/lr-ai/blob/9651c04c88fd2751c4652af7ef2dbab621e424b0/skills/export-sheet/SKILL.md","raw":"$27"},{"name":"screen-papers","label":"Screen papers","gist":"Filter and screen literature-review candidates - apply deterministic filters (year, language, type, citations, keywords) and then title/abstract screening by parallel subagents that label each paper c","source":"https://github.com/yazangthb/lr-ai/blob/47cd39316b710f0ada2b4f6302fa877f2e065ec8/skills/screen-papers/SKILL.md","raw":"$28"},{"name":"seed-search","label":"Seed search","gist":"Start a literature review - set up the review (topic, research questions, inclusion/exclusion criteria) and find an initial set of seed papers via Semantic Scholar queries and known DOIs/arXiv ids.","source":"https://github.com/yazangthb/lr-ai/blob/47cd39316b710f0ada2b4f6302fa877f2e065ec8/skills/seed-search/SKILL.md","raw":"$29"},{"name":"setup","label":"setup","gist":"Check or set up LR-AI, the literature-review plugin - run its doctor (Python packages, API reachability, API keys), install its Python dependencies, and explain the optional API keys.","source":"https://github.com/yazangthb/lr-ai/blob/47cd39316b710f0ada2b4f6302fa877f2e065ec8/skills/setup/SKILL.md","raw":"$2a"},{"name":"snowball","label":"snowball","gist":"Forward and backward citation snowballing for a literature review - collect the references of each paper and the papers that cite it (via OpenAlex, falling back to Semantic Scholar), for one or more r","source":"https://github.com/yazangthb/lr-ai/blob/47cd39316b710f0ada2b4f6302fa877f2e065ec8/skills/snowball/SKILL.md","raw":"$2b"}],"skillNote":"7 in the Claude listing"}],null]}],"$L2c"]}]]}]] 22:oa26,1b:[["$","title","0",{"children":"LR-AI by yazangthb: AI connector · Agent Ready"}],["$","meta","1",{"name":"description","content":"Automated literature reviews: seed search, database keyword search, forward/backward snowballing, deduplication, filtering, LLM screening and categorization…"}],["$","link","2",{"rel":"manifest","href":"/manifest.webmanifest","crossOrigin":"$undefined"}],["$","link","3",{"rel":"canonical","href":"https://agentready.co/connectors/lr-ai"}],["$","meta","4",{"property":"og:title","content":"LR-AI by yazangthb: AI connector"}],["$","meta","5",{"property":"og:description","content":"Automated literature reviews: seed search, database keyword search, forward/backward snowballing, deduplication, filtering, LLM screening and categorization…"}],["$","meta","6",{"property":"og:url","content":"https://agentready.co/connectors/lr-ai"}],["$","meta","7",{"property":"og:site_name","content":"Agent Ready"}],["$","meta","8",{"property":"og:image","content":"https://agentready.co/connectors/lr-ai/opengraph-image?3b61e23efdbb5225"}],["$","meta","9",{"property":"og:image:type","content":"image/png"}],["$","meta","10",{"property":"og:image:width","content":"1200"}],["$","meta","11",{"property":"og:image:height","content":"630"}],["$","meta","12",{"property":"og:image:alt","content":"AI connector on Agent Ready"}],["$","meta","13",{"property":"og:type","content":"website"}],["$","meta","14",{"name":"twitter:card","content":"summary_large_image"}],["$","meta","15",{"name":"twitter:title","content":"LR-AI by yazangthb: AI connector"}],["$","meta","16",{"name":"twitter:description","content":"Automated literature reviews: seed search, database keyword search, forward/backward snowballing, deduplication, filtering, LLM screening and categorization…"}],["$","meta","17",{"name":"twitter:image","content":"https://agentready.co/connectors/lr-ai/opengraph-image?3b61e23efdbb5225"}],["$","meta","18",{"name":"twitter:image:alt","content":"AI connector on Agent Ready"}],["$","meta","19",{"name":"twitter:image:type","content":"image/png"}],["$","meta","20",{"name":"twitter:image:width","content":"1200"}],["$","meta","21",{"name":"twitter:image:height","content":"630"}],["$","link","22",{"rel":"icon","href":"/favicon.ico?favicon.3iymezf4ycw3e.ico","sizes":"64x64","type":"image/x-icon"}],["$","link","23",{"rel":"icon","href":"/icon.svg?icon.1m07wgr9jqhm5.svg","sizes":"any","type":"image/svg+xml"}],["$","link","24",{"rel":"apple-touch-icon","href":"/apple-icon.png?apple-icon.3q8qrgufoflqu.png","sizes":"180x180","type":"image/png"}],["$","$L2d","25",{}]] 22:o10,1c:null 22:null 22:o22,:HL["/stores/claude.png","image"] 22:o12de,2c:["$","aside",null,{"className":"side","children":[["$","div",null,{"className":"card","children":[["$","h2",null,{"children":"Stores"}],["$","div",null,{"className":"st","children":[["$","div","claude",{"style":{"display":"contents"},"children":[["$","span",null,{"className":"s","children":[["$","img",null,{"className":"sm","src":"/stores/claude.png","alt":"","width":16,"height":16}],["$","b",null,{"children":"Claude"}],["$","a",null,{"className":"src","href":"https://github.com/yazangthb/lr-ai","target":"_blank","rel":"noopener nofollow","children":"Source ↗"}]]}],["$","span",null,{"className":"rk2","children":[["$","small",null,{"children":"Listed, not ranked"}],null]}]]}]]}]]}],["$","div",null,{"className":"card","children":[["$","h2",null,{"children":"Details"}],["$","dl",null,{"className":"kv","children":[[["$","dt",null,{"children":"Company"}],["$","dd",null,{"children":"yazangthb"}]],[["$","dt",null,{"children":"Publisher"}],["$","dd",null,{"children":"Community"}]],["$","dt",null,{"children":"Category"}],["$","dd",null,{"children":["$","$L5",null,{"href":"/categories/research","children":"Research & strategy"}]}],[["$","dt",null,{"children":"Audience"}],["$","dd",null,{"children":"For business"}]],null,["$","dt",null,{"children":"Sign-in"}],["$","dd",null,{"children":"None needed"}],["$","dt",null,{"children":"Interactive UI"}],["$","dd",null,{"children":"No"}],[["$","dt",null,{"children":"Checked"}],["$","dd",null,{"children":"Oct 27, 2026"}]],null]}]]}],["$","div",null,{"className":"card","children":[["$","h2",null,{"children":["More in ","Research & strategy"]}],["$","ul",null,{"className":"sim","children":[["$","li","4242",{"children":["$","$L5",null,{"href":"/connectors/firecrawl","prefetch":false,"children":[["$","$L23",null,{"name":"Firecrawl","url":"https://bodgzvrssuydrdmsrqsh.supabase.co/storage/v1/object/public/listing-media/v/44/443993e148098c1de53cedc510fc840dea11d8c8e9a3f39a25617a72c776b90d-64.webp","size":24}],["$","b",null,{"children":"Firecrawl"}],["$","small",null,{"children":"#67"}]]}]}],["$","li","4241",{"children":["$","$L5",null,{"href":"/connectors/exa","prefetch":false,"children":[["$","$L23",null,{"name":"Exa","url":"https://bodgzvrssuydrdmsrqsh.supabase.co/storage/v1/object/public/listing-media/v/b5/b5752032e01cebb5979424bab9bedff6398c953e8aa64b4d84b5ee42d0076968-64.webp","size":24}],["$","b",null,{"children":"Exa"}],["$","small",null,{"children":"#81"}]]}]}],["$","li","4453",{"children":["$","$L5",null,{"href":"/connectors/tinyfish","prefetch":false,"children":[["$","$L23",null,{"name":"TinyFish","url":"https://bodgzvrssuydrdmsrqsh.supabase.co/storage/v1/object/public/listing-media/v/31/3141625032019b3d3e5242716d243abb70fb2afa27f4f744a4ccb620013214d1-64.webp","size":24}],["$","b",null,{"children":"TinyFish"}],["$","small",null,{"children":"#252"}]]}]}],["$","li","4481",{"children":["$","$L5",null,{"href":"/connectors/parallel-search","prefetch":false,"children":[["$","$L23",null,{"name":"Parallel Search","url":"https://bodgzvrssuydrdmsrqsh.supabase.co/storage/v1/object/public/listing-media/v/cd/cd5750317526dcb14ac7fccc73730522f3afc39f59846edf8a355200c16c95cb-64.webp","size":24}],["$","b",null,{"children":"Parallel Search"}],["$","small",null,{"children":"#319"}]]}]}],["$","li","4192",{"children":["$","$L5",null,{"href":"/connectors/pubmed","prefetch":false,"children":[["$","$L23",null,{"name":"PubMed","url":"https://bodgzvrssuydrdmsrqsh.supabase.co/storage/v1/object/public/listing-media/v/a6/a677ecb9865f7a0dbfff9863672a05acdb642af1774da954145c83140ddc51cf-64.webp","size":24}],["$","b",null,{"children":"PubMed"}],["$","small",null,{"children":"#30"}]]}]}],["$","li","4211",{"children":["$","$L5",null,{"href":"/connectors/consensus","prefetch":false,"children":[["$","$L23",null,{"name":"Consensus","url":"https://bodgzvrssuydrdmsrqsh.supabase.co/storage/v1/object/public/listing-media/v/6d/6dbc8484cf9a35843860367fc78998c73afa678bf6acc0bbbca6ee803d2670bf-64.webp","size":24}],["$","b",null,{"children":"Consensus"}],["$","small",null,{"children":"#44"}]]}]}],["$","li","4223",{"children":["$","$L5",null,{"href":"/connectors/clinical-trials","prefetch":false,"children":[["$","$L23",null,{"name":"Clinical Trials","url":"https://bodgzvrssuydrdmsrqsh.supabase.co/storage/v1/object/public/listing-media/v/61/6136f7db817a75fd75a5ef4dfeb548733bfa7d875c2d7e917a2d1485943e0420-64.webp","size":24}],["$","b",null,{"children":"Clinical Trials"}],["$","small",null,{"children":"#61"}]]}]}],["$","li","4324",{"children":["$","$L5",null,{"href":"/connectors/tavily","prefetch":false,"children":[["$","$L23",null,{"name":"Tavily","url":"https://bodgzvrssuydrdmsrqsh.supabase.co/storage/v1/object/public/listing-media/v/41/41f7229586af69c9a1663e8cc036b84e6502c620080aa14f820c28be06209eb1-64.webp","size":24}],"$L2e","$L2f"]}]}]]}]]}]]}] 22:o52,2e:["$","b",null,{"children":"Tavily"}] 2f:["$","small",null,{"children":"#179"}] 22:o11,20:false 1f:7025 22:o17,1e:C 1d:C 1:C 15:C b:C 22:C </article></body></html>

22:o1,#22:o641,2:"$Sreact.fragment" 3:"/_next/static/immutable/chunks/12yk9g49l7-7r.js" 4:"/_next/static/immutable/chunks/2jrhxdtf12-uf.js" 5:I[7575,["$3","$4"],""] 6:I[10291,["$3"],"default"] 7:I[2968,["$3"],"default"] 8:I[4098,["$3"],"WebMCP"] 9:I[57507,["$3"],"ExternalLinks"] a:I[93137,["$3"],"Analytics"] 10:"$Sreact.suspense" 16:"ViewportBoundary" 17:I[82520,["$3"],"$16"] 19:"MetadataBoundary" 1a:I[82520,["$3"],"$19"] :HL["/_next/static/immutable/chunks/3g2_9hnyt6p64.css","style"] 1:X b:X d:X 15:X f:[["children",{"s":"__PAGE__","h":49314,"d":{"r":["$","$2","c",{"children":[["$","$10",null,{"fallback":["$","div",null,{"className":"wrap","style":{"minHeight":"60vh"}}],"children":"$L11"}],["$L12","$L13"],"$L14"]}],"p":false,"v":"$15"}}]] e:[["children",{"s":{"n":"slug","t":"d","k":"lr-ai","s":[]},"h":49250,"d":{"r":["$","$2","c",{"children":[null,["$","$L6",null,{"parallelRouterKey":"children","error":"$undefined","errorStyles":"$undefined","errorScripts":"$undefined","template":["$","$L7",null,{}],"templateStyles":"$undefined","templateScripts":"$undefined","notFound":"$undefined","forbidden":"$undefined","unauthorized":"$undefined"}]]}],"p":false,"v":"$d"},"c":"$Qf"}]] c:[["children",{"s":"connectors","h":49250,"d":{"r":["$","$2","c",{"children":[null,["$","$L6",null,{"parallelRouterKey":"children","error":"$undefined","errorStyles":"$undefined","errorScripts":"$undefined","template":["$","$L7",null,{}],"templateStyles":"$undefined","templateScripts":"$undefined","notFound":"$undefined","forbidden":"$undefined","unauthorized":"$undefined"}]]}],"p":false,"v":"$d"},"c":"$Qe"}]] 1d:X 1e:X 22:o1232,0:{"q":"","i":false,"S":true,"r":"$1","t":{"t":{"s":"","h":49234,"d":{"r":["$","$2","c",{"children":[[["$","link","0",{"rel":"stylesheet","href":"/_next/static/immutable/chunks/3g2_9hnyt6p64.css","precedence":"next","crossOrigin":"$undefined","nonce":"$undefined"}],["$","script","script-0",{"src":"/_next/static/immutable/chunks/12yk9g49l7-7r.js","async":true,"nonce":"$undefined"}]],["$","html",null,{"lang":"en","className":"ec5Qua_variable","children":["$","body",null,{"children":[["$","script",null,{"type":"application/ld+json","dangerouslySetInnerHTML":{"__html":"{\"@context\":\"https://schema.org\",\"@graph\":[{\"@type\":\"Organization\",\"@id\":\"https://agentready.co/#org\",\"name\":\"Agent Ready\",\"url\":\"https://agentready.co/\",\"logo\":\"https://agentready.co/icon-512.png\"},{\"@type\":\"WebSite\",\"@id\":\"https://agentready.co/#site\",\"name\":\"Agent Ready\",\"url\":\"https://agentready.co/\",\"publisher\":{\"@id\":\"https://agentready.co/#org\"}}]}"}}],["$","header",null,{"className":"site","children":["$","div",null,{"className":"wrap","children":[["$","$L5",null,{"href":"/","className":"brand","children":[["$","svg",null,{"className":"brand-mark","viewBox":"0 0 64 64","aria-hidden":"true","children":[["$","defs",null,{"children":["$","linearGradient",null,{"id":"arg","x1":"0","y1":"1","x2":"1","y2":"0","children":[["$","stop",null,{"offset":"0","stopColor":"#2563EB"}],["$","stop",null,{"offset":"1","stopColor":"#06B6D4"}]]}]}],["$","path",null,{"d":"M9 55 L32 9 L55 55","fill":"none","stroke":"url(#arg)","strokeWidth":"11","strokeLinecap":"round","strokeLinejoin":"round"}],["$","circle",null,{"cx":"32","cy":"45","r":"7","fill":"#FACC15"}]]}],"Agent Ready"]}],["$","nav",null,{"aria-label":"Main","children":[["$","$L5",null,{"href":"/","children":"Connectors"}],["$","$L5",null,{"href":"/categories","className":"nv-x","children":"Categories"}],["$","$L5",null,{"href":"/blog","children":"Blog"}],["$","$L5",null,{"href":"/mcp-plugin-usage","className":"nv-x","children":"Usage"}],["$","$L5",null,{"href":"/methodology","className":"nv-x","children":"Methodology"}],["$","$L5",null,{"href":"/work-with-us","className":"nv-cta","children":"Work with us"}]]}]]}]}],["$","main",null,{"children":["$","$L6",null,{"parallelRouterKey":"children","error":"$undefined","errorStyles":"$undefined","errorScripts":"$undefined","template":["$","$L7",null,{}],"templateStyles":"$undefined","templateScripts":"$undefined","notFound":[[["$","title",null,{"children":"404: This page could not be found."}],["$","div",null,{"style":{"fontFamily":"system-ui,\"Segoe UI\",Roboto,Helvetica,Arial,sans-serif,\"Apple Color Emoji\",\"Segoe UI Emoji\"","height":"100vh","textAlign":"center","display":"flex","flexDirection":"column","alignItems":"center","justifyContent":"center"},"children":["$","div",null,{"children":[["$","style",null,{"dangerouslySetInnerHTML":{"__html":"body{color:#000;background:#fff;margin:0}.next-error-h1{border-right:1px solid rgba(0,0,0,.3)}@media (prefers-color-scheme:dark){body{color:#fff;background:#000}.next-error-h1{border-right:1px solid rgba(255,255,255,.3)}}"}}],["$","h1",null,{"className":"next-error-h1","style":{"display":"inline-block","margin":"0 20px 0 0","padding":"0 23px 0 0","fontSize":24,"fontWeight":500,"verticalAlign":"top","lineHeight":"49px"},"children":404}],["$","div",null,{"style":{"display":"inline-block"},"children":["$","h2",null,{"style":{"fontSize":14,"fontWeight":400,"lineHeight":"49px","margin":0},"children":"This page could not be found."}]}]]}]}]],[]],"forbidden":"$undefined","unauthorized":"$undefined"}]}],["$","$L8",null,{}],["$","$L9",null,{}],["$","$La",null,{}],["$","footer",null,{"className":"foot","children":["$","div",null,{"className":"wrap","children":[["$","span",null,{"children":"Agent Ready"}],["$","nav",null,{"aria-label":"Footer","children":[["$","$L5",null,{"href":"/blog","children":"Blog"}],["$","$L5",null,{"href":"/mcp-plugin-usage","children":"Usage"}],["$","$L5",null,{"href":"/methodology","children":"Methodology"}],["$","$L5",null,{"href":"/categories","children":"Categories"}],["$","a",null,{"href":"/llms.txt","children":"llms.txt"}],["$","a",null,{"href":"/sitemap.xml","children":"Sitemap"}],["$","$L5",null,{"href":"/work-with-us","children":"Work with us"}]]}]]}]}]]}]}]]}],"p":false,"v":"$b"},"c":"$Qc"},"h":{"r":["$","$2","h",{"children":[null,["$","$L17","r-6_8D9P-WwNIMZP1DkjEv",{"children":"$L18"}],["$","$L1a","r-6_8D9P-WwNIMZP1DkjEm",{"children":[["$","div",null,{"hidden":true,"children":["$","$10",null,{"name":"Next.Metadata","children":"$L1b"}]}],"$L1c"]}]]}],"p":false,"v":"$1d"}},"s":"$1e","a":"$@1f","w":"$@20"} 22:o1b1,21:I[82520,["$3"],"OutletBoundary"] :HL["/_next/static/immutable/chunks/2lo4itc7piqdu.css","style"] d:C 12:["$","link","0",{"rel":"stylesheet","href":"/_next/static/immutable/chunks/2lo4itc7piqdu.css","precedence":"next","crossOrigin":"$undefined","nonce":"$undefined"}] 13:["$","script","script-0",{"src":"/_next/static/immutable/chunks/2jrhxdtf12-uf.js","async":true,"nonce":"$undefined"}] 14:["$","$L21",null,{"children":"$@22"}] 22:o14d,1e:300 18:[["$","meta","0",{"charSet":"utf-8"}],["$","meta","1",{"name":"viewport","content":"width=device-width, initial-scale=1"}],["$","meta","2",{"name":"theme-color","content":"#ffffff","media":"(prefers-color-scheme: light)"}],["$","meta","3",{"name":"theme-color","content":"#0d0d0f","media":"(prefers-color-scheme: dark)"}]] 22:o25,1d:0 15:0 15:T4,slug1d:T4,slug1e:300 22:o67,23:I[37521,["$3","$4"],"Logo"] 24:I[53913,["$3","$4"],"Explore"] 2d:I[60329,["$3"],"IconMark"] 25:Tae0,22:oae0,--- name: categorize-papers description: Build a two-level taxonomy for the included papers of a literature review and have subagents write summaries, research-question relations, categories and reading priorities for each paper. Use after screening, or when the user asks to categorize, classify, group or summarize the review papers. allowed-tools: - Bash(python "${CLAUDE_PLUGIN_ROOT}/scripts/lr.py" *) - Bash(python3 "${CLAUDE_PLUGIN_ROOT}/scripts/lr.py" *) - Edit(lr-runs/**) --- # Taxonomy and summaries (step 5) CLI: `python "${CLAUDE_PLUGIN_ROOT}/scripts/lr.py" --run `. Run it with the Bash tool exactly as written, so it needs no permission prompt. `lr.py --help` lists the options. If the path still contains a `$` placeholder, the plugin root is two folders above this skill's base directory. Automatic by default: make the choices below yourself and keep going. Lines marked **Checkpoint** apply only in step-by-step mode (the user asked to confirm steps, e.g. `/lr-ai:lit-review --step`): there, stop and wait for the user's reply. ## 1. Build the taxonomy Read the included papers compactly: `lr.py show --where included=true --fields status,title,short_summary --limit 500`. Design 3-7 first-level categories, each with 2-6 sub-categories. - Align first-level categories with the research questions where possible, e.g. "Graph-Based Forecasting (RQ1)". - Categories should be mutually exclusive enough that each paper has one obvious home. - Every category should hold at least 2 papers. Use a small "Other / background" category for leftovers. **Checkpoint:** show the taxonomy as a short tree with an estimated count per category, and let the user rename, merge or split before you write it. Write it to `config.yaml`: ```yaml taxonomy: - name: "Graph-Based Forecasting (RQ1)" rq: RQ1 children: ["Citation Analysis & Recommendations", "Link Prediction"] ``` ## 2. Enrich with subagents 1. Run `lr.py batches --stage enrich`. Each batch file carries the taxonomy and research questions. 2. Launch one `lr-ai:paper-screener` subagent per batch file, in parallel, with the batch path as the prompt. It writes a summary, short summary, RQ relation, category, sub-category, priority and code availability for each paper. 3. Run `lr.py apply --stage enrich`. Re-run agents for any batches it reports as missing. If it reports categories outside the taxonomy, fix them with `lr.py set first_level="..." second_level="..."` or re-run the batch. ## 3. Check **Checkpoint:** show the distribution (`lr.py export` prints the category tree with counts) and ask whether anything should move. Move papers with `lr.py set first_level="..." second_level="..."`. Next step: export (skill `export-sheet`). 22:o8,26:Ta14,22:oa14,--- name: database-search description: Run boolean keyword searches for a literature review across scholarly databases (OpenAlex, Semantic Scholar, arXiv) and merge the results with deduplication. Use after seed search, or when the user asks to search databases, index databases, or run keyword/boolean queries for papers. allowed-tools: - Bash(python "${CLAUDE_PLUGIN_ROOT}/scripts/lr.py" *) - Bash(python3 "${CLAUDE_PLUGIN_ROOT}/scripts/lr.py" *) - Edit(lr-runs/**) --- # Database keyword search (step 2) CLI: `python "${CLAUDE_PLUGIN_ROOT}/scripts/lr.py" --run `. Run it with the Bash tool exactly as written, so it needs no permission prompt. `lr.py --help` lists the options. If the path still contains a `$` placeholder, the plugin root is two folders above this skill's base directory. Automatic by default: make the choices below yourself and keep going. Lines marked **Checkpoint** apply only in step-by-step mode (the user asked to confirm steps, e.g. `/lr-ai:lit-review --step`): there, stop and wait for the user's reply. ## 1. Build the queries Read `config.yaml` (topic, RQs, criteria) and a sample of the seed papers (`lr.py show --fields title,short_summary --limit 30`). Write 1-4 boolean queries: - one OR-group per concept, joined with AND: `("citation network" OR "citation graph") AND (forecast* OR predict*)` - quote multi-word phrases; use `*` for word stems; add `NOT` only for clear noise (e.g. `NOT protein`) - use uppercase AND / OR / NOT. The tool translates the query for each database. Put them in `search.queries`. `search.per_query` (default 100) caps the results kept per query per database, most relevant first. **Checkpoint:** show the queries before running them. ## 2. Run `lr.py search --run `. The output reports how many matches each database has in total: - tens of thousands of matches: the query is too broad, so add a concept or tighten the phrases. - very few matches: loosen the query (more synonyms, fewer ANDs). Fix such a query in `config.yaml` and run `search` once more. Unchanged queries are answered from the cache. Errors from one database don't stop the others. If the output has an `OPENALEX RATE LIMIT` notice, show it to the user word for word (what happened and how to set their own `OPENALEX_API_KEY`), then continue with the other databases. ## 3. Report Run `lr.py status --run ` and summarize it in one line: records per source, duplicates merged, unique papers. **Checkpoint:** wait for the user before snowballing. Next step: snowballing (skill `snowball`). 22:o9,27:T1157,22:o1157,--- name: export-sheet description: Export a literature review to a styled, categorized Google Sheet (via Composio's Google Sheets tools) and an Excel file - grouped category rows, color-coded status and priority, reading-status dropdowns, filters, plus Overview and Method (PRISMA counts) tabs. Use when the user asks for the review as a Google Sheet, spreadsheet, Excel file, or table. allowed-tools: - Bash(python "${CLAUDE_PLUGIN_ROOT}/scripts/lr.py" *) - Bash(python3 "${CLAUDE_PLUGIN_ROOT}/scripts/lr.py" *) - Edit(lr-runs/**) --- # Export (step 6) CLI: `python "${CLAUDE_PLUGIN_ROOT}/scripts/lr.py" --run `. Run it with the Bash tool exactly as written, so it needs no permission prompt. `lr.py --help` lists the options. If the path still contains a `$` placeholder, the plugin root is two folders above this skill's base directory. Automatic by default: make the choices below yourself and keep going. Lines marked **Checkpoint** apply only in step-by-step mode (the user asked to confirm steps, e.g. `/lr-ai:lit-review --step`): there, stop and wait for the user's reply. ## 1. Excel file (always) `lr.py export --run ` writes `/papers.xlsx` and prints the Google Sheet title, plus the Google account if `LR_AI_GOOGLE_ACCOUNT` is set. The workbook has these tabs: - **Papers**: category and sub-category header rows, papers sorted core → skim, then priority, then citations. Status and priority are color-coded. Reading-status and priority columns have dropdowns. There is a filter, and the header row and title column are frozen. - **Overview**: counts by category, year and venue. - **Method**: queries, sources, filters and the PRISMA-style flow. - **1 Seed search, 2 Database search, 3 Snowballing, 4 Filtering, 5 Screening** (Excel only): the papers each phase produced, under a note on what was done, with each paper's outcome. The Google Sheet leaves them out to keep it small. ## 2. Google Sheet via Composio (if available) Composio tools are named like `mcp____COMPOSIO_*`. If none are present, skip to step 3. 1. Call `COMPOSIO_SEARCH_TOOLS` with the use case "create a google sheet and write values". This gives you a `session_id` and the Google Sheets connection status. - No active connection: don't wait for one. Go to step 3, and in the final report say that the sheet can be made later: connect Google Sheets in Composio, then ask to "export the review to Google Sheets". **Checkpoint:** offer to connect now with `COMPOSIO_MANAGE_CONNECTIONS`. - Several connected accounts: Composio needs the account in every call. Use the account the user named, else the one `lr.py export` printed (`LR_AI_GOOGLE_ACCOUNT`), else the first account listed. Name it in the final report. **Checkpoint:** if the user named none, ask which one. 2. Create the spreadsheet: `COMPOSIO_MULTI_EXECUTE_TOOL` with `GOOGLESHEETS_CREATE_GOOGLE_SHEET1` `{ "title": "" }`, and `"account": "<account>"` in the tool entry when you chose one. Note the `spreadsheetId`. Note the first tab's `sheetId` if the response shows it (normally 0; check with `GOOGLESHEETS_GET_SPREADSHEET_INFO` if unsure). 3. `lr.py sheet-plan --run <dir> --spreadsheet-id <ID> [--sheet-id <first tab id>] [--account <account>]`. This writes numbered plan files to `<run>/export/gsheets/`. Without `--account`, it uses `LR_AI_GOOGLE_ACCOUNT` when that is set. 4. Execute the files in order. Read each file and call `COMPOSIO_MULTI_EXECUTE_TOOL` once with `tools` = the file's `tools` array, copied verbatim, plus `session_id` and `sync_response_to_workbench: false`. Calls within a file are independent. - If calls fail with rate limits (Sheets allows about 60 writes per minute), wait about 60 seconds and re-send only the failed items. - If a call fails on its arguments, check the schema with `COMPOSIO_GET_TOOL_SCHEMAS`, fix that call, and continue. - With many `values_*` files (large reviews), delegate executing all plan files to one general-purpose subagent. Give it the file list and these rules, to keep the main context small. 5. Give the user the sheet URL (also saved in `<run>/sheet_url.txt`). ## 3. Without Composio Tell the user to upload `papers.xlsx` to Google Drive and open it with Google Sheets (or use File → Import in Sheets). Colors, dropdowns, frozen panes and filters carry over. 22:o8,28:Te6b,22:oe6b,--- name: screen-papers description: Filter and screen literature-review candidates - apply deterministic filters (year, language, type, citations, keywords) and then title/abstract screening by parallel subagents that label each paper core, skim or exclude. Use when the user wants to filter, screen, or narrow down the collected papers against inclusion criteria. allowed-tools: - Bash(python "${CLAUDE_PLUGIN_ROOT}/scripts/lr.py" *) - Bash(python3 "${CLAUDE_PLUGIN_ROOT}/scripts/lr.py" *) - Edit(lr-runs/**) --- # Filtering and screening (step 4) CLI: `python "${CLAUDE_PLUGIN_ROOT}/scripts/lr.py" <command> --run <run dir>`. Run it with the Bash tool exactly as written, so it needs no permission prompt. `lr.py <command> --help` lists the options. If the path still contains a `$` placeholder, the plugin root is two folders above this skill's base directory. Automatic by default: make the choices below yourself and keep going. Lines marked **Checkpoint** apply only in step-by-step mode (the user asked to confirm steps, e.g. `/lr-ai:lit-review --step`): there, stop and wait for the user's reply. ## 1. Deterministic filters Run `lr.py filter`. It recomputes `excluded_reason` from the `filters` in `config.yaml` every time and prints how many papers await screening. Screening takes about one subagent per 25 papers, so keep it to roughly 400 papers unless the user gave another limit. If more remain, tighten the filters in this order, re-running `filter` after each change, until the rest fits: 1. `min_hits: 2`: snowballed papers must be linked from at least 2 start-set papers 2. `require_any`: 3-8 core terms of the topic, with `*` for word stems, e.g. `["citation*", "bibliometric*"]` 3. `min_citations` (e.g. 5); `citation_grace_years` keeps recent papers exempt Keep track of what you tightened for the final report. **Checkpoint:** show the counts and propose the stricter filters instead of applying them yourself. Papers the user added (`seed.papers`, `add`) are never filtered out. ## 2. Screening with subagents If the request adds or changes criteria, update `inclusion_criteria` / `exclusion_criteria` in `config.yaml` first. The batch files carry them to the screeners. To re-screen papers that were already screened, run `batches` with `--force`. 1. Run `lr.py batches --stage screen`. It writes batch files of `screening.batch_size` (default 25) papers each and prints their paths. 2. Launch one `lr-ai:paper-screener` subagent per batch file, in parallel. Send them in a single message, up to about 10 at a time. Prompt each one with just the batch file path. The criteria and research questions travel inside the file. 3. Run `lr.py apply --stage screen`. If it lists batches with missing results, re-launch agents for those batch files only, then `apply` again. Status meanings: **core** means it directly addresses a research question and meets the inclusion criteria. **skim** means it's related, background, or partially relevant. **exclude** means it's off-topic or hits an exclusion criterion. ## 3. Results Report the core/skim/exclude counts in one line. If no paper was included, stop and tell the user: the criteria or the queries are probably off. **Checkpoint:** show the core papers (`lr.py show --where status=core --fields year,citations,title,reason`) and a few borderline ones (`--where status=skim`). Fix disagreements with `lr.py set <id> status=core|skim|exclude status_reason="..."`. Manual edits are permanent: later `apply` runs never overwrite them. If the user wants more rounds, go back to the `snowball` skill with `--from included`. Next step: categorization and summaries (skill `categorize-papers`). 22:o8,29:Tcb7,22:ocb7,--- name: seed-search description: Start a literature review - set up the review (topic, research questions, inclusion/exclusion criteria) and find an initial set of seed papers via Semantic Scholar queries and known DOIs/arXiv ids. Use when the user wants to begin a literature review, find initial or seed papers for a topic, or set up LR-AI. allowed-tools: - Bash(python "${CLAUDE_PLUGIN_ROOT}/scripts/lr.py" *) - Bash(python3 "${CLAUDE_PLUGIN_ROOT}/scripts/lr.py" *) - Edit(lr-runs/**) --- # Seed search (step 1) CLI: `python "${CLAUDE_PLUGIN_ROOT}/scripts/lr.py" <command>`. Run it with the Bash tool exactly as written, so it needs no permission prompt (use `python3` if `python` is not Python 3). `lr.py <command> --help` lists the options. Always pass `--run <run dir>` once a run exists. If the path still contains a `$` placeholder, the plugin root is two folders above this skill's base directory. Automatic by default: make the choices below yourself and keep going. Lines marked **Checkpoint** apply only in step-by-step mode (the user asked to confirm steps, e.g. `/lr-ai:lit-review --step`): there, stop and wait for the user's reply. ## 1. Check the environment (once per session) Run `lr.py doctor`. If `yaml` or `openpyxl` is missing, install them with `python -m pip install -r "${CLAUDE_PLUGIN_ROOT}/requirements.txt"` (Claude Code asks the user to approve it). Mention missing API keys in one line and continue. They're optional: LR-AI ships a shared OpenAlex key. If any command prints an `OPENALEX RATE LIMIT` notice, show it to the user word for word. ## 2. Frame the review Take the topic, research questions, criteria, year range and known papers from the user's request, and decide the rest yourself: - 1-3 research questions (RQ1, RQ2, ...) that the topic implies - inclusion and exclusion criteria that follow from them, for example "proposes a method for X" or "not peer-reviewed" - no year limit unless the request implies one **Checkpoint:** ask the user for what the request left out instead of deciding it yourself. Then run `lr.py init "<topic>" --name <short-name>` and fill `config.yaml` in the new run folder: `research_questions`, `inclusion_criteria`, `exclusion_criteria`, `filters.year_min` / `year_max`. ## 3. Seed queries Draft 3-6 natural-language queries that cover the RQs from different angles: synonyms, the method side and the application side. Put them in `seed.queries`, and known papers in `seed.papers`. **Checkpoint:** show the queries and let the user edit them. If scholarly-search tools are available in this session (e.g. an alphaXiv / Consensus / Semantic Scholar MCP, or web search), use them too. Put the DOIs or arXiv ids of highly relevant papers they return into `seed.papers`. This often gives better seeds than keyword queries. ## 4. Run `lr.py seed --run <dir>`. Each query falls back from Semantic Scholar to OpenAlex to arXiv if a service is rate-limited. Look over the returned list and exclude clearly off-topic seeds: `lr.py set --run <dir> <id> status=exclude status_reason="off-topic seed"`. Excluded seeds are not snowballed. **Checkpoint:** show the seed papers, grouped by relevance, before moving on. Next step: database search (skill `database-search`). 22:o8,2a:T8c5,22:o8c5,--- name: setup description: Check or set up LR-AI, the literature-review plugin - run its doctor (Python packages, API reachability, API keys), install its Python dependencies, and explain the optional API keys. Use when the user asks to check, diagnose, test, install or configure LR-AI, or asks for the LR-AI doctor. allowed-tools: - Bash(python "${CLAUDE_PLUGIN_ROOT}/scripts/lr.py" *) - Bash(python3 "${CLAUDE_PLUGIN_ROOT}/scripts/lr.py" *) --- # LR-AI setup and health check The tool is `python "${CLAUDE_PLUGIN_ROOT}/scripts/lr.py"`. Use `python3` if `python` is not Python 3. Don't search the disk for it; this is the command. Run it with the Bash tool exactly as written, so it needs no permission prompt. If the path still contains a `$` placeholder, the plugin root is two folders above this skill's base directory. 1. Run `python "${CLAUDE_PLUGIN_ROOT}/scripts/lr.py" doctor`. 2. If `yaml` or `openpyxl` is missing, install them with `python -m pip install -r "${CLAUDE_PLUGIN_ROOT}/requirements.txt"` (Claude Code asks the user to approve it). 3. Report the result in a few lines. A ⚠ next to a database is not fatal: LR-AI falls back to the other databases. Mention the optional environment variables once: - `OPENALEX_API_KEY`: free at openalex.org. LR-AI ships a shared key; a personal key has its own daily budget, so set one if OpenAlex reports a rate limit. - `S2_API_KEY`: free on request from Semantic Scholar. Gives a dedicated rate limit. - `OPENALEX_EMAIL`: identifies the user to OpenAlex (polite pool). - `LR_AI_GOOGLE_ACCOUNT`: the alias or id of the Composio Google account that gets the sheets, when several are connected. Otherwise the first connected account is used. To set one permanently on Windows: `setx OPENALEX_API_KEY "<key>"`, then restart Claude Code. On macOS/Linux: add `export OPENALEX_API_KEY=<key>` to the shell profile. 4. Point to the next step: `/lr-ai:lit-review <topic>` runs a whole review without stopping, and `/lr-ai:lit-review <topic> --step` pauses for approval after each step. To run and check phases one at a time, see `${CLAUDE_PLUGIN_ROOT}/docs/phases.md`. The full command reference is in `${CLAUDE_PLUGIN_ROOT}/docs/cli.md`. 22:o8,2b:Tba6,22:oba6,--- name: snowball description: Forward and backward citation snowballing for a literature review - collect the references of each paper and the papers that cite it (via OpenAlex, falling back to Semantic Scholar), for one or more rounds. Use when the user asks to snowball, expand a paper set through citations/references, or find related work from a start set. allowed-tools: - Bash(python "${CLAUDE_PLUGIN_ROOT}/scripts/lr.py" *) - Bash(python3 "${CLAUDE_PLUGIN_ROOT}/scripts/lr.py" *) - Edit(lr-runs/**) --- # Snowballing (step 3) CLI: `python "${CLAUDE_PLUGIN_ROOT}/scripts/lr.py" <command> --run <run dir>`. Run it with the Bash tool exactly as written, so it needs no permission prompt. `lr.py <command> --help` lists the options. If the path still contains a `$` placeholder, the plugin root is two folders above this skill's base directory. Automatic by default: make the choices below yourself and keep going. Lines marked **Checkpoint** apply only in step-by-step mode (the user asked to confirm steps, e.g. `/lr-ai:lit-review --step`): there, stop and wait for the user's reply. ## No run yet If the user just names papers to expand (DOIs, arXiv ids or titles), create a run first: `lr.py init "<topic>" --name <short-name>`, then `lr.py add --run <dir> <paper> <paper> ...`. Papers added this way are never filtered out. ## Round 1 (usually enough) 1. Run `lr.py filter` first, so papers outside the year range or excluded types are not expanded. 2. Adjust `snowball.max_references` and `max_citations` if needed. Both keep the most-cited papers first. Defaults are 200 references and 50 citing papers per paper. 3. Run `lr.py snowball --run <dir>`. Each start paper costs about 2-3 API calls. The command reports references and citing papers found, new unique papers, and which backend served each paper. 4. If it refuses because more than 300 papers would be expanded, don't pass `--force`. Snowball from the included papers only, which is the standard Wohlin guideline: set `snowball.from: included`, screen first (skill `screen-papers`), run `snowball` again, then screen the new papers (`screen-papers` again; only unscreened papers go to the screeners). ## More rounds Screen the new papers first (`filter`, then the `screen-papers` skill). Then run `lr.py snowball --from included` to expand only the included papers from the previous round. The round number increments automatically. Stop when a round adds few new included papers (saturation). ## Notes - Every paper records who led to it (`found_via`: `backward:<parent>` / `forward:<parent>`). Papers linked from several start-set papers are strong candidates: `lr.py show --sort hits`. - If OpenAlex's daily budget runs out, the command switches to Semantic Scholar and prints an `OPENALEX RATE LIMIT` notice. Show it to the user word for word: it says what happened and how to set their own `OPENALEX_API_KEY`. Next step: filtering and screening (skill `screen-papers`). 22:o10b2,11:[["$","script",null,{"type":"application/ld+json","dangerouslySetInnerHTML":{"__html":"{\"@context\":\"https://schema.org\",\"@type\":\"SoftwareApplication\",\"name\":\"LR-AI\",\"description\":\"Automated literature reviews: seed search, database keyword search, forward/backward snowballing, deduplication, filtering, LLM screening and categorization, exported as a styled Google Sheet (via Composio) and Excel file.\",\"applicationCategory\":\"Research & strategy\",\"operatingSystem\":\"Claude\",\"url\":\"https://agentready.co/connectors/lr-ai\",\"sameAs\":[],\"publisher\":{\"@type\":\"Organization\",\"name\":\"yazangthb\",\"url\":\"https://github.com\"}}"}}],["$","div",null,{"className":"wrap","children":[["$","div",null,{"className":"pp-head","children":[["$","$L23",null,{"name":"LR-AI","url":null,"size":56}],["$","div",null,{"children":[["$","div",null,{"className":"crumbs","children":[["$","$L5",null,{"href":"/","children":"Connectors"}],[" / ",["$","$L5",null,{"href":"/categories/research","children":"Research & strategy"}]]]}],["$","h1",null,{"children":["LR-AI",null]}]]}]]}],["$","div",null,{"className":"pp","children":[["$","div",null,{"className":"ca","children":[["$","p",null,{"className":"tag","children":"Automated literature reviews: seed search, database keyword search, forward/backward snowballing, deduplication, filtering, LLM screening and categorization, exported as a styled Google Sheet (via Composio) and Excel file."}],null,["$","$L24",null,{"examples":[],"tools":[],"prefix":"","server":null,"instructions":[],"skills":[{"name":"categorize-papers","label":"Categorize papers","gist":"Build a two-level taxonomy for the included papers of a literature review and have subagents write summaries, research-question relations, categories and reading priorities for each paper.","source":"https://github.com/yazangthb/lr-ai/blob/47cd39316b710f0ada2b4f6302fa877f2e065ec8/skills/categorize-papers/SKILL.md","raw":"$25"},{"name":"database-search","label":"Database search","gist":"Run boolean keyword searches for a literature review across scholarly databases (OpenAlex, Semantic Scholar, arXiv) and merge the results with deduplication.","source":"https://github.com/yazangthb/lr-ai/blob/47cd39316b710f0ada2b4f6302fa877f2e065ec8/skills/database-search/SKILL.md","raw":"$26"},{"name":"export-sheet","label":"Export sheet","gist":"Export a literature review to a styled, categorized Google Sheet (via Composio's Google Sheets tools) and an Excel file - grouped category rows, color-coded status and priority, reading-status dropdow","source":"https://github.com/yazangthb/lr-ai/blob/9651c04c88fd2751c4652af7ef2dbab621e424b0/skills/export-sheet/SKILL.md","raw":"$27"},{"name":"screen-papers","label":"Screen papers","gist":"Filter and screen literature-review candidates - apply deterministic filters (year, language, type, citations, keywords) and then title/abstract screening by parallel subagents that label each paper c","source":"https://github.com/yazangthb/lr-ai/blob/47cd39316b710f0ada2b4f6302fa877f2e065ec8/skills/screen-papers/SKILL.md","raw":"$28"},{"name":"seed-search","label":"Seed search","gist":"Start a literature review - set up the review (topic, research questions, inclusion/exclusion criteria) and find an initial set of seed papers via Semantic Scholar queries and known DOIs/arXiv ids.","source":"https://github.com/yazangthb/lr-ai/blob/47cd39316b710f0ada2b4f6302fa877f2e065ec8/skills/seed-search/SKILL.md","raw":"$29"},{"name":"setup","label":"setup","gist":"Check or set up LR-AI, the literature-review plugin - run its doctor (Python packages, API reachability, API keys), install its Python dependencies, and explain the optional API keys.","source":"https://github.com/yazangthb/lr-ai/blob/47cd39316b710f0ada2b4f6302fa877f2e065ec8/skills/setup/SKILL.md","raw":"$2a"},{"name":"snowball","label":"snowball","gist":"Forward and backward citation snowballing for a literature review - collect the references of each paper and the papers that cite it (via OpenAlex, falling back to Semantic Scholar), for one or more r","source":"https://github.com/yazangthb/lr-ai/blob/47cd39316b710f0ada2b4f6302fa877f2e065ec8/skills/snowball/SKILL.md","raw":"$2b"}],"skillNote":"7 in the Claude listing"}],null]}],"$L2c"]}]]}]] 22:oa26,1b:[["$","title","0",{"children":"LR-AI by yazangthb: AI connector · Agent Ready"}],["$","meta","1",{"name":"description","content":"Automated literature reviews: seed search, database keyword search, forward/backward snowballing, deduplication, filtering, LLM screening and categorization…"}],["$","link","2",{"rel":"manifest","href":"/manifest.webmanifest","crossOrigin":"$undefined"}],["$","link","3",{"rel":"canonical","href":"https://agentready.co/connectors/lr-ai"}],["$","meta","4",{"property":"og:title","content":"LR-AI by yazangthb: AI connector"}],["$","meta","5",{"property":"og:description","content":"Automated literature reviews: seed search, database keyword search, forward/backward snowballing, deduplication, filtering, LLM screening and categorization…"}],["$","meta","6",{"property":"og:url","content":"https://agentready.co/connectors/lr-ai"}],["$","meta","7",{"property":"og:site_name","content":"Agent Ready"}],["$","meta","8",{"property":"og:image","content":"https://agentready.co/connectors/lr-ai/opengraph-image?3b61e23efdbb5225"}],["$","meta","9",{"property":"og:image:type","content":"image/png"}],["$","meta","10",{"property":"og:image:width","content":"1200"}],["$","meta","11",{"property":"og:image:height","content":"630"}],["$","meta","12",{"property":"og:image:alt","content":"AI connector on Agent Ready"}],["$","meta","13",{"property":"og:type","content":"website"}],["$","meta","14",{"name":"twitter:card","content":"summary_large_image"}],["$","meta","15",{"name":"twitter:title","content":"LR-AI by yazangthb: AI connector"}],["$","meta","16",{"name":"twitter:description","content":"Automated literature reviews: seed search, database keyword search, forward/backward snowballing, deduplication, filtering, LLM screening and categorization…"}],["$","meta","17",{"name":"twitter:image","content":"https://agentready.co/connectors/lr-ai/opengraph-image?3b61e23efdbb5225"}],["$","meta","18",{"name":"twitter:image:alt","content":"AI connector on Agent Ready"}],["$","meta","19",{"name":"twitter:image:type","content":"image/png"}],["$","meta","20",{"name":"twitter:image:width","content":"1200"}],["$","meta","21",{"name":"twitter:image:height","content":"630"}],["$","link","22",{"rel":"icon","href":"/favicon.ico?favicon.3iymezf4ycw3e.ico","sizes":"64x64","type":"image/x-icon"}],["$","link","23",{"rel":"icon","href":"/icon.svg?icon.1m07wgr9jqhm5.svg","sizes":"any","type":"image/svg+xml"}],["$","link","24",{"rel":"apple-touch-icon","href":"/apple-icon.png?apple-icon.3q8qrgufoflqu.png","sizes":"180x180","type":"image/png"}],["$","$L2d","25",{}]] 22:o10,1c:null 22:null 22:o22,:HL["/stores/claude.png","image"] 22:o12de,2c:["$","aside",null,{"className":"side","children":[["$","div",null,{"className":"card","children":[["$","h2",null,{"children":"Stores"}],["$","div",null,{"className":"st","children":[["$","div","claude",{"style":{"display":"contents"},"children":[["$","span",null,{"className":"s","children":[["$","img",null,{"className":"sm","src":"/stores/claude.png","alt":"","width":16,"height":16}],["$","b",null,{"children":"Claude"}],["$","a",null,{"className":"src","href":"https://github.com/yazangthb/lr-ai","target":"_blank","rel":"noopener nofollow","children":"Source ↗"}]]}],["$","span",null,{"className":"rk2","children":[["$","small",null,{"children":"Listed, not ranked"}],null]}]]}]]}]]}],["$","div",null,{"className":"card","children":[["$","h2",null,{"children":"Details"}],["$","dl",null,{"className":"kv","children":[[["$","dt",null,{"children":"Company"}],["$","dd",null,{"children":"yazangthb"}]],[["$","dt",null,{"children":"Publisher"}],["$","dd",null,{"children":"Community"}]],["$","dt",null,{"children":"Category"}],["$","dd",null,{"children":["$","$L5",null,{"href":"/categories/research","children":"Research & strategy"}]}],[["$","dt",null,{"children":"Audience"}],["$","dd",null,{"children":"For business"}]],null,["$","dt",null,{"children":"Sign-in"}],["$","dd",null,{"children":"None needed"}],["$","dt",null,{"children":"Interactive UI"}],["$","dd",null,{"children":"No"}],[["$","dt",null,{"children":"Checked"}],["$","dd",null,{"children":"Oct 27, 2026"}]],null]}]]}],["$","div",null,{"className":"card","children":[["$","h2",null,{"children":["More in ","Research & strategy"]}],["$","ul",null,{"className":"sim","children":[["$","li","4242",{"children":["$","$L5",null,{"href":"/connectors/firecrawl","prefetch":false,"children":[["$","$L23",null,{"name":"Firecrawl","url":"https://bodgzvrssuydrdmsrqsh.supabase.co/storage/v1/object/public/listing-media/v/44/443993e148098c1de53cedc510fc840dea11d8c8e9a3f39a25617a72c776b90d-64.webp","size":24}],["$","b",null,{"children":"Firecrawl"}],["$","small",null,{"children":"#67"}]]}]}],["$","li","4241",{"children":["$","$L5",null,{"href":"/connectors/exa","prefetch":false,"children":[["$","$L23",null,{"name":"Exa","url":"https://bodgzvrssuydrdmsrqsh.supabase.co/storage/v1/object/public/listing-media/v/b5/b5752032e01cebb5979424bab9bedff6398c953e8aa64b4d84b5ee42d0076968-64.webp","size":24}],["$","b",null,{"children":"Exa"}],["$","small",null,{"children":"#81"}]]}]}],["$","li","4453",{"children":["$","$L5",null,{"href":"/connectors/tinyfish","prefetch":false,"children":[["$","$L23",null,{"name":"TinyFish","url":"https://bodgzvrssuydrdmsrqsh.supabase.co/storage/v1/object/public/listing-media/v/31/3141625032019b3d3e5242716d243abb70fb2afa27f4f744a4ccb620013214d1-64.webp","size":24}],["$","b",null,{"children":"TinyFish"}],["$","small",null,{"children":"#252"}]]}]}],["$","li","4481",{"children":["$","$L5",null,{"href":"/connectors/parallel-search","prefetch":false,"children":[["$","$L23",null,{"name":"Parallel Search","url":"https://bodgzvrssuydrdmsrqsh.supabase.co/storage/v1/object/public/listing-media/v/cd/cd5750317526dcb14ac7fccc73730522f3afc39f59846edf8a355200c16c95cb-64.webp","size":24}],["$","b",null,{"children":"Parallel Search"}],["$","small",null,{"children":"#319"}]]}]}],["$","li","4192",{"children":["$","$L5",null,{"href":"/connectors/pubmed","prefetch":false,"children":[["$","$L23",null,{"name":"PubMed","url":"https://bodgzvrssuydrdmsrqsh.supabase.co/storage/v1/object/public/listing-media/v/a6/a677ecb9865f7a0dbfff9863672a05acdb642af1774da954145c83140ddc51cf-64.webp","size":24}],["$","b",null,{"children":"PubMed"}],["$","small",null,{"children":"#30"}]]}]}],["$","li","4211",{"children":["$","$L5",null,{"href":"/connectors/consensus","prefetch":false,"children":[["$","$L23",null,{"name":"Consensus","url":"https://bodgzvrssuydrdmsrqsh.supabase.co/storage/v1/object/public/listing-media/v/6d/6dbc8484cf9a35843860367fc78998c73afa678bf6acc0bbbca6ee803d2670bf-64.webp","size":24}],["$","b",null,{"children":"Consensus"}],["$","small",null,{"children":"#44"}]]}]}],["$","li","4223",{"children":["$","$L5",null,{"href":"/connectors/clinical-trials","prefetch":false,"children":[["$","$L23",null,{"name":"Clinical Trials","url":"https://bodgzvrssuydrdmsrqsh.supabase.co/storage/v1/object/public/listing-media/v/61/6136f7db817a75fd75a5ef4dfeb548733bfa7d875c2d7e917a2d1485943e0420-64.webp","size":24}],["$","b",null,{"children":"Clinical Trials"}],["$","small",null,{"children":"#61"}]]}]}],["$","li","4324",{"children":["$","$L5",null,{"href":"/connectors/tavily","prefetch":false,"children":[["$","$L23",null,{"name":"Tavily","url":"https://bodgzvrssuydrdmsrqsh.supabase.co/storage/v1/object/public/listing-media/v/41/41f7229586af69c9a1663e8cc036b84e6502c620080aa14f820c28be06209eb1-64.webp","size":24}],"$L2e","$L2f"]}]}]]}]]}]]}] 22:o52,2e:["$","b",null,{"children":"Tavily"}] 2f:["$","small",null,{"children":"#179"}] 22:o11,20:false 1f:7025 22:o17,1e:C 1d:C 1:C 15:C b:C 22:C </article></body></html>

LR-AI

Skills

LR-AI

Skills