{
  "tool": "list_pack_skills",
  "slug": "scholastic-research",
  "kind": "agent",
  "name": "Scholastic Research",
  "format": "mybot.farm/agent-pack",
  "skills": [
    {
      "name": "Scholastic research router",
      "description": "Use for any research request: scope the question, plan the search strategy, then dispatch to source-specific skills.",
      "content": "Scholastic Research is a research-process agent for students and researchers. On a research request: (1) clarify the question, scope, and academic level; (2) build a multi-faceted search strategy — use the arxiv procedure for STEM, cross-reference with JSTOR/Google Scholar/PubMed; (3) evaluate every source on peer-review status, venue, author credentials, date, and conflicts of interest; (4) synthesize thematic connections across 3+ credible publications (convergent findings vs. contradictory evidence, frameworks, methods, gaps) rather than summarizing each source in isolation; (5) cite in the requested style (APA/MLA/Chicago) with full bibliographic detail and access date. Academic-but-accessible tone; organize by research phase (exploration → evaluation → synthesis). Mark source-derived content vs. original analysis."
    },
    {
      "name": "arxiv",
      "description": "Search arXiv papers by keyword, author, category, or ID.",
      "content": "# arXiv Research\n\nSearch and retrieve academic papers from arXiv via their free REST API. No API key, no dependencies — just curl.\n\n## Quick Reference\n\n| Action | Command |\n|--------|---------|\n| Search papers | `curl \"https://export.arxiv.org/api/query?search_query=all:QUERY&max_results=5\"` |\n| Get specific paper | `curl \"https://export.arxiv.org/api/query?id_list=2402.03300\"` |\n| Read abstract (web) | `web_extract(urls=[\"https://arxiv.org/abs/2402.03300\"])` |\n| Read full paper (PDF) | `web_extract(urls=[\"https://arxiv.org/pdf/2402.03300\"])` |\n\n## Searching Papers\n\nThe API returns Atom XML. Parse with `grep`/`sed` or pipe through `python` for clean output.\n\n### Basic search\n\n```bash\ncurl -s \"https://export.arxiv.org/api/query?search_query=all:GRPO+reinforcement+learning&max_results=5\"\n```\n\n### Clean output (parse XML to readable format)\n\n```bash\ncurl -s \"https://export.arxiv.org/api/query?search_query=all:GRPO+reinforcement+learning&max_results=5&sortBy=submittedDate&sortOrder=descending\" | python -c \"\nimport sys, xml.etree.ElementTree as ET\nns = {'a': 'http://www.w3.org/2005/Atom'}\nroot = ET.parse(sys.stdin).getroot()\nfor i, entry in enumerate(root.findall('a:entry', ns)):\n    title = entry.find('a:title', ns).text.strip().replace('\\n', ' ')\n    arxiv_id = entry.find('a:id', ns).text.strip().split('/abs/')[-1]\n    published = entry.find('a:published', ns).text[:10]\n    authors = ', '.join(a.find('a:name', ns).text for a in entry.findall('a:author', ns))\n    summary = entry.find('a:summary', ns).text.strip()[:200]\n    cats = ', '.join(c.get('term') for c in entry.findall('a:category', ns))\n    print(f'{i+1}. [{arxiv_id}] {title}')\n    print(f'   Authors: {authors}')\n    print(f'   Published: {published} | Categories: {cats}')\n    print(f'   Abstract: {summary}...')\n    print(f'   PDF: https://arxiv.org/pdf/{arxiv_id}')\n    print()\n\"\n```\n\n## Search Query Syntax\n\n| Prefix | Searches | Example |\n|--------|----------|---------|\n| `all:` | All fields | `all:transformer+attention` |\n| `ti:` | Title | `ti:large+language+models` |\n| `au:` | Author | `au:vaswani` |\n| `abs:` | Abstract | `abs:reinforcement+learning` |\n| `cat:` | Category | `cat:cs.AI` |\n| `co:` | Comment | `co:accepted+NeurIPS` |\n\n### Boolean operators\n\n```\n# AND (default when using +)\nsearch_query=all:transformer+attention\n\n# OR\nsearch_query=all:GPT+OR+all:BERT\n\n# AND NOT\nsearch_query=all:language+model+ANDNOT+all:vision\n\n# Exact phrase\nsearch_query=ti:\"chain+of+thought\"\n\n# Combined\nsearch_query=au:hinton+AND+cat:cs.LG\n```\n\n## Sort and Pagination\n\n| Parameter | Options |\n|-----------|---------|\n| `sortBy` | `relevance`, `lastUpdatedDate`, `submittedDate` |\n| `sortOrder` | `ascending`, `descending` |\n| `start` | Result offset (0-based) |\n| `max_results` | Number of results (default 10, max 30000) |\n\n```bash\n# Latest 10 papers in cs.AI\ncurl -s \"https://export.arxiv.org/api/query?search_query=cat:cs.AI&sortBy=submittedDate&sortOrder=descending&max_results=10\"\n```\n\n## Fetching Specific Papers\n\n```bash\n# By arXiv ID\ncurl -s \"https://export.arxiv.org/api/query?id_list=2402.03300\"\n\n# Multiple papers\ncurl -s \"https://export.arxiv.org/api/query?id_list=2402.03300,2401.12345,2403.00001\"\n```\n\n## BibTeX Generation\n\nAfter fetching metadata for a paper, generate a BibTeX entry:\n\n{% raw %}\n```bash\ncurl -s \"https://export.arxiv.org/api/query?id_list=1706.03762\" | python -c \"\nimport sys, xml.etree.ElementTree as ET\nns = {'a': 'http://www.w3.org/2005/Atom', 'arxiv': 'http://arxiv.org/schemas/atom'}\nroot = ET.parse(sys.stdin).getroot()\nentry = root.find('a:entry', ns)\nif entry is None: sys.exit('Paper not found')\ntitle = entry.find('a:title', ns).text.strip().replace('\\n', ' ')\nauthors = ' and '.join(a.find('a:name', ns).text for a in entry.findall('a:author', ns))\nyear = entry.find('a:published', ns).text[:4]\nraw_id = entry.find('a:id', ns).text.strip().split('/abs/')[-1]\ncat = entry.find('arxiv:primary_category', ns)\nprimary = cat.get('term') if cat is not None else 'cs.LG'\nlast_name = entry.find('a:author', ns).find('a:name', ns).text.split()[-1]\nprint(f'@article{{{last_name}{year}_{raw_id.replace(\\\".\\\", \\\"\\\")},')\nprint(f'  title     = {{{title}}},')\nprint(f'  author    = {{{authors}}},')\nprint(f'  year      = {{{year}}},')\nprint(f'  eprint    = {{{raw_id}}},')\nprint(f'  archivePrefix = {{arXiv}},')\nprint(f'  primaryClass  = {{{primary}}},')\nprint(f'  url       = {{https://arxiv.org/abs/{raw_id}}}')\nprint('}')\n\"\n```\n{% endraw %}\n\n## Reading Paper Content\n\nAfter finding a paper, read it:\n\n```\n# Abstract page (fast, metadata + abstract)\nweb_extract(urls=[\"https://arxiv.org/abs/2402.03300\"])\n\n# Full paper (PDF → markdown via Firecrawl)\nweb_extract(urls=[\"https://arxiv.org/pdf/2402.03300\"])\n```\n\nFor local PDF processing, see the `ocr-and-documents` skill.\n\n## Common Categories\n\n| Category | Field |\n|----------|-------|\n| `cs.AI` | Artificial Intelligence |\n| `cs.CL` | Computation and Language (NLP) |\n| `cs.CV` | Computer Vision |\n| `cs.LG` | Machine Learning |\n| `cs.CR` | Cryptography and Security |\n| `stat.ML` | Machine Learning (Statistics) |\n| `math.OC` | Optimization and Control |\n| `physics.comp-ph` | Computational Physics |\n\nFull list: https://arxiv.org/category_taxonomy\n\n## Helper Script\n\nThe `scripts/search_arxiv.py` script handles XML parsing and provides clean output:\n\n```bash\npython scripts/search_arxiv.py \"GRPO reinforcement learning\"\npython scripts/search_arxiv.py \"transformer attention\" --max 10 --sort date\npython scripts/search_arxiv.py --author \"Yann LeCun\" --max 5\npython scripts/search_arxiv.py --category cs.AI --sort date\npython scripts/search_arxiv.py --id 2402.03300\npython scripts/search_arxiv.py --id 2402.03300,2401.12345\n```\n\nNo dependencies — uses only Python stdlib.\n\n---\n\n## Semantic Scholar (Citations, Related Papers, Author Profiles)\n\narXiv doesn't provide citation data or recommendations. Use the **Semantic Scholar API** for that — free, no key needed for basic use (1 req/sec), returns JSON.\n\n### Get paper details + citations\n\n```bash\n# By arXiv ID\ncurl -s \"https://api.semanticscholar.org/graph/v1/paper/arXiv:2402.03300?fields=title,authors,citationCount,referenceCount,influentialCitationCount,year,abstract\" | python -m json.tool\n\n# By Semantic Scholar paper ID or DOI\ncurl -s \"https://api.semanticscholar.org/graph/v1/paper/DOI:10.1234/example?fields=title,citationCount\"\n```\n\n### Get citations OF a paper (who cited it)\n\n```bash\ncurl -s \"https://api.semanticscholar.org/graph/v1/paper/arXiv:2402.03300/citations?fields=title,authors,year,citationCount&limit=10\" | python -m json.tool\n```\n\n### Get references FROM a paper (what it cites)\n\n```bash\ncurl -s \"https://api.semanticscholar.org/graph/v1/paper/arXiv:2402.03300/references?fields=title,authors,year,citationCount&limit=10\" | python -m json.tool\n```\n\n### Search papers (alternative to arXiv search, returns JSON)\n\n```bash\ncurl -s \"https://api.semanticscholar.org/graph/v1/paper/search?query=GRPO+reinforcement+learning&limit=5&fields=title,authors,year,citationCount,externalIds\" | python -m json.tool\n```\n\n### Get paper recommendations\n\n```bash\ncurl -s -X POST \"https://api.semanticscholar.org/recommendations/v1/papers/\" \\\n  -H \"Content-Type: application/json\" \\\n  -d '{\"positivePaperIds\": [\"arXiv:2402.03300\"], \"negativePaperIds\": []}' | python -m json.tool\n```\n\n### Author profile\n\n```bash\ncurl -s \"https://api.semanticscholar.org/graph/v1/author/search?query=Yann+LeCun&fields=name,hIndex,citationCount,paperCount\" | python -m json.tool\n```\n\n### Useful Semantic Scholar fields\n\n`title`, `authors`, `year`, `abstract`, `citationCount`, `referenceCount`, `influentialCitationCount`, `isOpenAccess`, `openAccessPdf`, `fieldsOfStudy`, `publicationVenue`, `externalIds` (contains arXiv ID, DOI, etc.)\n\n---\n\n## Complete Research Workflow\n\n1. **Discover**: `python scripts/search_arxiv.py \"your topic\" --sort date --max 10`\n2. **Assess impact**: `curl -s \"https://api.semanticscholar.org/graph/v1/paper/arXiv:ID?fields=citationCount,influentialCitationCount\"`\n3. **Read abstract**: `web_extract(urls=[\"https://arxiv.org/abs/ID\"])`\n4. **Read full paper**: `web_extract(urls=[\"https://arxiv.org/pdf/ID\"])`\n5. **Find related work**: `curl -s \"https://api.semanticscholar.org/graph/v1/paper/arXiv:ID/references?fields=title,citationCount&limit=20\"`\n6. **Get recommendations**: POST to Semantic Scholar recommendations endpoint\n7. **Track authors**: `curl -s \"https://api.semanticscholar.org/graph/v1/author/search?query=NAME\"`\n\n## Rate Limits\n\n| API | Rate | Auth |\n|-----|------|------|\n| arXiv | ~1 req / 3 seconds | None needed |\n| Semantic Scholar | 1 req / second | None (100/sec with API key) |\n\n## Notes\n\n- arXiv returns Atom XML — use the helper script or parsing snippet for clean output\n- Semantic Scholar returns JSON — pipe through `python -m json.tool` for readability\n- arXiv IDs: old format (`hep-th/0601001`) vs new (`2402.03300`)\n- PDF: `https://arxiv.org/pdf/{id}` — Abstract: `https://arxiv.org/abs/{id}`\n- HTML (when available): `https://arxiv.org/html/{id}`\n- For local PDF processing, see the `ocr-and-documents` skill\n\n## ID Versioning\n\n- `arxiv.org/abs/1706.03762` always resolves to the **latest** version\n- `arxiv.org/abs/1706.03762v1` points to a **specific** immutable version\n- When generating citations, preserve the version suffix you actually read to prevent citation drift (a later version may substantially change content)\n- The API `<id>` field returns the versioned URL (e.g., `http://arxiv.org/abs/1706.03762v7`)\n\n## Withdrawn Papers\n\nPapers can be withdrawn after submission. When this happens:\n- The `<summary>` field contains a withdrawal notice (look for \"withdrawn\" or \"retracted\")\n- Metadata fields may be incomplete\n- Always check the summary before treating a result as a valid paper"
    },
    {
      "name": "grounded-citations",
      "description": "Ground answers and documents in cited, verifiable sources.",
      "content": "# Grounded Citations\n\nEvery claim taken from an outside source gets an inline numbered citation and a\n`Sources:` list, Perplexity-style. A ledger script owns the `url → [n]` mapping\nso the numbers and URLs come from retrieval, never from memory — the model only\never emits small integers it was handed.\n\nFor high-stakes work the same ledger doubles as a fact-checking chain: verbatim\nquotes are attached to each source (rejected unless they literally appear in\nthe fetched page text), claims from model knowledge are flagged `[unverified]`,\nand `verify --evidence` fails any draft whose cited sources carry no evidence.\n\nThis skill covers answers in chat, written documents (markdown, PDF, docx,\nslides), and research reports. It does not cover academic BibTeX pipelines —\nfor conference papers use the `arxiv` skill, which this skill\nfeeds (see `references/citation-formats.md`).\n\n## When to Use\n\nUse whenever an answer or artifact rests on information you fetched rather than\nknew:\n\n- Research, comparisons, news summaries, \"what is the current state of X\"\n- Any deliverable you write to disk that quotes, paraphrases, or reports\n  outside facts — reports, briefs, docs, decks, wiki pages\n- Fact-finding where the user will want to check your work\n- Multi-source synthesis where conflicting sources must be attributed\n\nSkip inline citations when the retrieval is incidental to another task — a\nquick syntax/version lookup mid-coding, casual conversation, creative writing.\nMention a URL only if the user would plausibly want the link.\n\n## Prerequisites\n\nNone beyond the standard toolset. `scripts/sources.py` is stdlib-only Python 3.\nRetrieval comes from whatever is configured: `web_search`, `web_extract`,\n`browser_navigate`, or `terminal` (curl, CLIs).\n\nLedger location: `$HERMES_HOME/cache/citations/ledger.json` (profile-aware).\nOverride per task with `--ledger <path>` or `HERMES_CITATION_LEDGER`.\n\n## How to Run\n\n```bash\nS=~/.hermes/skills/research/grounded-citations/scripts/sources.py\n\npython \"$S\" reset                                  # start a clean ledger\npython \"$S\" add https://example.com/a --title \"A\"  # prints: [1]\npython \"$S\" add https://example.com/b --title \"B\"  # prints: [2]\npython \"$S\" list                                   # ledger table\npython \"$S\" render                                 # Sources: block\npython \"$S\" verify draft.md                        # catch bad citations\n```\n\n`add` is idempotent and URL-normalized: the same page always returns the same\nid within a ledger, so ids stay stable across many search/extract rounds.\n\n## Quick Reference\n\n| Action | Command |\n|---|---|\n| Fresh ledger for a new task | `sources.py reset` |\n| Register a source, get its id | `sources.py add <url> [--title T]` |\n| Register several at once | `sources.py add <url1> <url2> ...` |\n| Register from JSON tool output | `sources.py ingest results.json` |\n| Attach verbatim evidence to a source | `sources.py quote <id> --text \"exact wording\" --from page.txt` |\n| Show ledger | `sources.py list [--json]` |\n| Render the Sources block | `sources.py render [--style markdown\\|plain\\|footnotes\\|bibtex\\|evidence] [--only 1,3]` |\n| Render only what a draft cites | `sources.py render --cited-in draft.md` |\n| Rewrite a draft's Sources block in place | `sources.py render --replace-in draft.md` |\n| Check a draft's citations | `sources.py verify draft.md [--strict] [--min-coverage 0.6] [--evidence]` |\n\n## Procedure\n\n① **Reset the ledger** at the start of a task that will produce a grounded\nanswer or document. Skip the reset when continuing work whose ids are already\nin a draft — reusing the ledger keeps the numbering stable.\n\n② **Register every source at retrieval time.** After each `web_search` /\n`web_extract` / `browser_navigate` / fetch, pass the URLs to `sources.py add`\n(or pipe the raw JSON through `sources.py ingest`). Do this *before* writing\nprose. Registering later, from memory, is the failure mode this skill exists to\nprevent.\n\n③ **Write cite-while-drafting.** Place the bracketed id(s) immediately after\neach sentence the source supports:\n\n```\nIce floats because it is less dense than liquid water.[1][2]\n```\n\n- No space before the bracket; each id in its own brackets.\n- Max 3 ids per sentence. Cite per sentence, not one dump at the end.\n- Only ids the ledger returned. Never invent an id or a URL.\n- Claims from your own knowledge get no citation.\n- Conflicting sources: present both readings, each with its own id.\n- Quote exact figures, dates, and names as the source states them; flag gaps\n  explicitly (\"no source found for X\") instead of smoothing them over.\n\n④ **Append the Sources block** with `sources.py render --cited-in <draft>` so\nthe id → URL mapping is generated mechanically from the ledger, not retyped.\nFor non-markdown targets pick the matching `--style` and follow\n`references/citation-formats.md` for placement (footnotes in docx, endnotes in\nPDF/LaTeX, a Sources slide in decks, per-page source lists in wiki output).\n\n⑤ **Verify before delivering** — `sources.py verify <draft>` exits non-zero on\nunknown ids, on a Sources block that disagrees with the ledger, or (with\n`--min-coverage`) on prose that is too thinly cited. Fix and re-run.\n\n⑥ **Chat answers** follow the same steps with the draft in your reply: register\nsources, cite inline, end with the rendered `Sources:` list. For a short answer\nyou may render the block from `sources.py render --only <ids>` instead of\nwriting to a file.\n\n## Multi-Platform Sweeps\n\n\"What are people saying about X\" / \"research X across the web\" is not one\n`web_search`. Fan out across source types, collect in parallel, then synthesise\nwith every claim attributed to the platform it came from:\n\n| Source type | Route | What it adds |\n|---|---|---|\n| Open web | `web_search` → `web_extract` | official docs, articles, announcements |\n| Community discussion | `reddit-reading` (`search`, `thread`) | real user experience, complaints, workarounds |\n| Blogs / releases / changelogs | `rss-feeds` (`read`, `discover`) | dated primary posts, version history |\n| Video | `youtube-content` | walkthroughs, demos, talks |\n| Code | `terminal` with `gh search repos` / `gh search issues` | implementations, open bugs |\n| X/Twitter | `xurl` (needs API access) | announcements, developer chatter |\n\nThe `reddit-reading` and `rss-feeds` skills are optional. If absent, install with\n`hermes skills install official/social-media/reddit-reading` or\n`hermes skills install official/research/rss-feeds` before using them.\n\nRegister every URL from every route in the ledger as it arrives (step ②). Keep\nopinion and measurement apart: a Reddit thread is evidence that users *report*\nsomething, not that it is true; pair it with a primary source or label it as\nsentiment. Report per-platform coverage gaps (\"Reddit search returned nothing\nnewer than March\") rather than silently narrowing to what worked.\n\n## Fact-Checking Mode\n\nFor work where the reader must be able to check the chain — medical, legal,\nfinancial, safety, disputed claims, or when the user asks for fact-checking —\nupgrade from citations to evidence:\n\n① **Attach a verbatim quote per source.** After extracting a page, save its\ntext to a file and attach the sentence(s) that carry each claim:\n\n```bash\npython \"$S\" quote 1 --text \"Ice is about 9% less dense than liquid water.\" --from page1.txt\n```\n\nThe quote is rejected unless it appears verbatim in the evidence text\n(insensitive to whitespace, case, and markdown markup — inline links like\n`_[ERAP1](https://…)_` in extracted text match the plain prose a reader sees),\nso a paraphrase or misremembered figure cannot masquerade as evidence.\nCopy-paste from the fetched text; never retype. Quote the sentence as the\nreader sees it — the matcher sees through the extractor's markup for you, so\nyou don't have to reproduce link syntax or escaped asterisks in your quote.\n\n② **Flag model-knowledge claims with `[unverified]`.** A load-bearing claim\nyou could not source gets an explicit marker instead of a citation:\n\n```\nThe refactor likely predates the 2.0 release.[unverified]\n```\n\n`verify --min-coverage` counts `[unverified]` sentences as covered — the goal\nis declared provenance for every claim, not a citation on every sentence.\nIf a key claim can be checked, check it; `[unverified]` is for what genuinely\ncannot be, and a fact-check deliverable dominated by `[unverified]` markers\nshould say so in its summary.\n\n③ **Cross-check disputed facts against a second independent source.** When two\nsources disagree, cite both readings with their own ids and quotes, and say\nwhich you weight and why. One source is reporting; two independent sources are\ncorroboration.\n\n④ **Verify with the evidence gate and render the evidence block:**\n\n```bash\npython \"$S\" verify report.md --evidence --min-coverage 0.5\npython \"$S\" render --style evidence --replace-in report.md\n```\n\n`--evidence` fails the draft if any cited source has no attached quote. The\n`evidence` render style prints each source's quotes beneath its URL, so the\ndeliverable shows claim → source → exact supporting text with nothing taken on\nfaith. Use `--replace-in <draft>` to rewrite an existing Sources block in place\n(idempotent — safe to re-run after attaching more quotes); `--cited-in` prints\nto stdout instead. Both emit the heading `## Sources` (`--style plain` emits\n`Sources:`).\n\n**What `--min-coverage` counts.** Coverage is\n`sentences with declared provenance / prose sentences`. A prose sentence is a\nnon-empty line fragment of 4+ words after the Sources block, headings (`#`),\ntable rows (`|`), and fenced code are dropped; blockquote markers are stripped.\nProvenance is declared by either a `[n]` citation or an `[unverified]` marker,\nso a sentence carrying both counts once. Run `verify` without a threshold first\nand read the `info: stats:` line to see the counts before picking a number.\n\n## Pitfalls\n\n- **Registering after writing.** The ledger must be populated from tool output,\n  not reconstructed from the draft — that reintroduces exactly the hallucinated\n  -URL risk the numbering removes.\n- **Renumbering mid-task.** Never hand-edit ids in a draft. Ids are ledger\n  identities; if a draft cites `[4]`, `[4]` must stay that source. Run `reset`\n  only between tasks.\n- **Retyping URLs into the Sources block.** Always `render`. A hand-typed URL\n  is an unverified claim.\n- **Citing a search snippet as if you read the page.** A `web_search`\n  description supports only what it literally says. Cite the extracted page\n  when the claim needs the body — `web_extract` it first.\n- **Over-citing.** Three ids on a sentence is the ceiling; a citation on every\n  clause makes text unreadable and hides which source carries the load.\n- **Citing the ledger in code/config artifacts.** Source comments belong in\n  prose deliverables and doc headers, not inside generated code.\n- **Parallel subagents.** Each subagent has its own working directory; point\n  them all at one ledger with `--ledger` (or `HERMES_CITATION_LEDGER`) if their\n  outputs get merged, otherwise their ids will collide.\n- **Quoting from a snippet instead of the page.** Evidence quotes must come\n  from the extracted page text, not a search-result description — `web_extract`\n  first, save the text, then `quote --from` that file.\n- **Paraphrasing into `quote --text`.** The verbatim check will reject it; the\n  fix is to find the actual sentence, not to reword until something matches.\n- **Using `[unverified]` as an escape hatch.** It marks the rare claim that\n  genuinely cannot be sourced; if most sentences carry it, the task needed more\n  retrieval, not more markers.\n- **Hand-editing the Sources block.** Use `render --replace-in <draft>`; slicing\n  the file yourself risks a stale or duplicated block that `verify` then flags.\n\n## Verification\n\n```bash\npython \"$S\" verify report.md --strict --min-coverage 0.5\n```\n\nGreen means: every `[n]` in the draft exists in the ledger, the Sources block\nlists exactly the cited ids with the ledger's URLs, and the cited share of\nsource-bearing sentences meets the threshold. Read the warnings even when the\nexit code is 0 — uncited registered sources usually mean a claim lost its\nattribution during editing."
    },
    {
      "name": "llm-wiki",
      "description": "Karpathy's LLM Wiki: build/query interlinked markdown KB.",
      "content": "# Karpathy's LLM Wiki\n\nBuild and maintain a persistent, compounding knowledge base as interlinked markdown files.\nBased on [Andrej Karpathy's LLM Wiki pattern](https://gist.github.com/karpathy/442a6bf555914893e9891c11519de94f).\n\nUnlike traditional RAG (which rediscovers knowledge from scratch per query), the wiki\ncompiles knowledge once and keeps it current. Cross-references are already there.\nContradictions have already been flagged. Synthesis reflects everything ingested.\n\n**Division of labor:** The human curates sources and directs analysis. The agent\nsummarizes, cross-references, files, and maintains consistency.\n\n## When This Skill Activates\n\nUse this skill when the user:\n- Asks to create, build, or start a wiki or knowledge base\n- Asks to ingest, add, or process a source into their wiki\n- Asks a question and an existing wiki is present at the configured path\n- Asks to lint, audit, or health-check their wiki\n- References their wiki, knowledge base, or \"notes\" in a research context\n\n## Wiki Location\n\n**Location:** Set via `WIKI_PATH` environment variable (e.g. in `${HERMES_HOME:-~/.hermes}/.env`).\n\nIf unset, defaults to `~/wiki`.\n\n```bash\nWIKI=\"${WIKI_PATH:-$HOME/wiki}\"\n```\n\nThe wiki is just a directory of markdown files — open it in Obsidian, VS Code, or\nany editor. No database, no special tooling required.\n\n## Architecture: Three Layers\n\n```\nwiki/\n├── SCHEMA.md           # Conventions, structure rules, domain config\n├── index.md            # Sectioned content catalog with one-line summaries\n├── log.md              # Chronological action log (append-only, rotated yearly)\n├── raw/                # Layer 1: Immutable source material\n│   ├── articles/       # Web articles, clippings\n│   ├── papers/         # PDFs, arxiv papers\n│   ├── transcripts/    # Meeting notes, interviews\n│   └── assets/         # Images, diagrams referenced by sources\n├── entities/           # Layer 2: Entity pages (people, orgs, products, models)\n├── concepts/           # Layer 2: Concept/topic pages\n├── comparisons/        # Layer 2: Side-by-side analyses\n└── queries/            # Layer 2: Filed query results worth keeping\n```\n\n**Layer 1 — Raw Sources:** Immutable. The agent reads but never modifies these.\n**Layer 2 — The Wiki:** Agent-owned markdown files. Created, updated, and\ncross-referenced by the agent.\n**Layer 3 — The Schema:** `SCHEMA.md` defines structure, conventions, and tag taxonomy.\n\n## Resuming an Existing Wiki (CRITICAL — do this every session)\n\nWhen the user has an existing wiki, **always orient yourself before doing anything**:\n\n① **Read `SCHEMA.md`** — understand the domain, conventions, and tag taxonomy.\n② **Read `index.md`** — learn what pages exist and their summaries.\n③ **Scan recent `log.md`** — read the last 20-30 entries to understand recent activity.\n\n```bash\nWIKI=\"${WIKI_PATH:-$HOME/wiki}\"\n# Orientation reads at session start\nread_file \"$WIKI/SCHEMA.md\"\nread_file \"$WIKI/index.md\"\nread_file \"$WIKI/log.md\" offset=<last 30 lines>\n```\n\nOnly after orientation should you ingest, query, or lint. This prevents:\n- Creating duplicate pages for entities that already exist\n- Missing cross-references to existing content\n- Contradicting the schema's conventions\n- Repeating work already logged\n\nFor large wikis (100+ pages), also run a quick `search_files` for the topic\nat hand before creating anything new.\n\n## Initializing a New Wiki\n\nWhen the user asks to create or start a wiki:\n\n1. Determine the wiki path (from `$WIKI_PATH` env var, or ask the user; default `~/wiki`)\n2. Create the directory structure above\n3. Ask the user what domain the wiki covers — be specific\n4. Write `SCHEMA.md` customized to the domain (see template below)\n5. Write initial `index.md` with sectioned header\n6. Write initial `log.md` with creation entry\n7. Confirm the wiki is ready and suggest first sources to ingest\n\n### SCHEMA.md Template\n\nAdapt to the user's domain. The schema constrains agent behavior and ensures consistency:\n\n```markdown\n# Wiki Schema\n\n## Domain\n[What this wiki covers — e.g., \"AI/ML research\", \"personal health\", \"startup intelligence\"]\n\n## Conventions\n- File names: lowercase, hyphens, no spaces (e.g., `transformer-architecture.md`)\n- Every wiki page starts with YAML frontmatter (see below)\n- Use `[[wikilinks]]` to link between pages (minimum 2 outbound links per page)\n- When updating a page, always bump the `updated` date\n- Every new page must be added to `index.md` under the correct section\n- Every action must be appended to `log.md`\n- **Provenance markers:** On pages that synthesize 3+ sources, append `^[raw/articles/source-file.md]`\n  at the end of paragraphs whose claims come from a specific source. This lets a reader trace each\n  claim back without re-reading the whole raw file. Optional on single-source pages where the\n  `sources:` frontmatter is enough.\n\n## Frontmatter\n  ```yaml\n  ---\n  title: Page Title\n  created: YYYY-MM-DD\n  updated: YYYY-MM-DD\n  type: entity | concept | comparison | query | summary\n  tags: [from taxonomy below]\n  sources: [raw/articles/source-name.md]\n  # Optional quality signals:\n  confidence: high | medium | low        # how well-supported the claims are\n  contested: true                        # set when the page has unresolved contradictions\n  contradictions: [other-page-slug]      # pages this one conflicts with\n  ---\n  ```\n\n`confidence` and `contested` are optional but recommended for opinion-heavy or fast-moving\ntopics. Lint surfaces `contested: true` and `confidence: low` pages for review so weak claims\ndon't silently harden into accepted wiki fact.\n\n### raw/ Frontmatter\n\nRaw sources ALSO get a small frontmatter block so re-ingests can detect drift:\n\n```yaml\n---\nsource_url: https://example.com/article   # original URL, if applicable\ningested: YYYY-MM-DD\nsha256: <hex digest of the raw content below the frontmatter>\n---\n```\n\nThe `sha256:` lets a future re-ingest of the same URL skip processing when content is unchanged,\nand flag drift when it has changed. Compute over the body only (everything after the closing\n`---`), not the frontmatter itself.\n\n## Tag Taxonomy\n[Define 10-20 top-level tags for the domain. Add new tags here BEFORE using them.]\n\nExample for AI/ML:\n- Models: model, architecture, benchmark, training\n- People/Orgs: person, company, lab, open-source\n- Techniques: optimization, fine-tuning, inference, alignment, data\n- Meta: comparison, timeline, controversy, prediction\n\nRule: every tag on a page must appear in this taxonomy. If a new tag is needed,\nadd it here first, then use it. This prevents tag sprawl.\n\n## Page Thresholds\n- **Create a page** when an entity/concept appears in 2+ sources OR is central to one source\n- **Add to existing page** when a source mentions something already covered\n- **DON'T create a page** for passing mentions, minor details, or things outside the domain\n- **Split a page** when it exceeds ~200 lines — break into sub-topics with cross-links\n- **Archive a page** when its content is fully superseded — move to `_archive/`, remove from index\n\n## Entity Pages\nOne page per notable entity. Include:\n- Overview / what it is\n- Key facts and dates\n- Relationships to other entities ([[wikilinks]])\n- Source references\n\n## Concept Pages\nOne page per concept or topic. Include:\n- Definition / explanation\n- Current state of knowledge\n- Open questions or debates\n- Related concepts ([[wikilinks]])\n\n## Comparison Pages\nSide-by-side analyses. Include:\n- What is being compared and why\n- Dimensions of comparison (table format preferred)\n- Verdict or synthesis\n- Sources\n\n## Update Policy\nWhen new information conflicts with existing content:\n1. Check the dates — newer sources generally supersede older ones\n2. If genuinely contradictory, note both positions with dates and sources\n3. Mark the contradiction in frontmatter: `contradictions: [page-name]`\n4. Flag for user review in the lint report\n```\n\n### index.md Template\n\nThe index is sectioned by type. Each entry is one line: wikilink + summary.\n\n```markdown\n# Wiki Index\n\n> Content catalog. Every wiki page listed under its type with a one-line summary.\n> Read this first to find relevant pages for any query.\n> Last updated: YYYY-MM-DD | Total pages: N\n\n## Entities\n<!-- Alphabetical within section -->\n\n## Concepts\n\n## Comparisons\n\n## Queries\n```\n\n**Scaling rule:** When any section exceeds 50 entries, split it into sub-sections\nby first letter or sub-domain. When the index exceeds 200 entries total, create\na `_meta/topic-map.md` that groups pages by theme for faster navigation.\n\n### log.md Template\n\n```markdown\n# Wiki Log\n\n> Chronological record of all wiki actions. Append-only.\n> Format: `## [YYYY-MM-DD] action | subject`\n> Actions: ingest, update, query, lint, create, archive, delete\n> When this file exceeds 500 entries, rotate: rename to log-YYYY.md, start fresh.\n\n## [YYYY-MM-DD] create | Wiki initialized\n- Domain: [domain]\n- Structure created with SCHEMA.md, index.md, log.md\n```\n\n## Core Operations\n\n### 1. Ingest\n\nWhen the user provides a source (URL, file, paste), integrate it into the wiki:\n\n① **Capture the raw source:**\n   - URL → use `web_extract` to get markdown, save to `raw/articles/`\n   - PDF → use `web_extract` (handles PDFs), save to `raw/papers/`\n   - Pasted text → save to appropriate `raw/` subdirectory\n   - Name the file descriptively: `raw/articles/karpathy-llm-wiki-2026.md`\n   - **Add raw frontmatter** (`source_url`, `ingested`, `sha256` of the body).\n     On re-ingest of the same URL: recompute the sha256, compare to the stored value —\n     skip if identical, flag drift and update if different. This is cheap enough to\n     do on every re-ingest and catches silent source changes.\n\n② **Discuss takeaways** with the user — what's interesting, what matters for\n   the domain. (Skip this in automated/cron contexts — proceed directly.)\n\n③ **Check what already exists** — search index.md and use `search_files` to find\n   existing pages for mentioned entities/concepts. This is the difference between\n   a growing wiki and a pile of duplicates.\n\n④ **Write or update wiki pages:**\n   - **New entities/concepts:** Create pages only if they meet the Page Thresholds\n     in SCHEMA.md (2+ source mentions, or central to one source)\n   - **Existing pages:** Add new information, update facts, bump `updated` date.\n     When new info contradicts existing content, follow the Update Policy.\n   - **Cross-reference:** Every new or updated page must link to at least 2 other\n     pages via `[[wikilinks]]`. Check that existing pages link back.\n   - **Tags:** Only use tags from the taxonomy in SCHEMA.md\n   - **Provenance:** On pages synthesizing 3+ sources, append `^[raw/articles/source.md]`\n     markers to paragraphs whose claims trace to a specific source.\n   - **Confidence:** For opinion-heavy, fast-moving, or single-source claims, set\n     `confidence: medium` or `low` in frontmatter. Don't mark `high` unless the\n     claim is well-supported across multiple sources.\n\n⑤ **Update navigation:**\n   - Add new pages to `index.md` under the correct section, alphabetically\n   - Update the \"Total pages\" count and \"Last updated\" date in index header\n   - Append to `log.md`: `## [YYYY-MM-DD] ingest | Source Title`\n   - List every file created or updated in the log entry\n\n⑥ **Report what changed** — list every file created or updated to the user.\n\nA single source can trigger updates across 5-15 wiki pages. This is normal\nand desired — it's the compounding effect.\n\n### 2. Query\n\nWhen the user asks a question about the wiki's domain:\n\n① **Read `index.md`** to identify relevant pages.\n② **For wikis with 100+ pages**, also `search_files` across all `.md` files\n   for key terms — the index alone may miss relevant content.\n③ **Read the relevant pages** using `read_file`.\n④ **Synthesize an answer** from the compiled knowledge. Cite the wiki pages\n   you drew from: \"Based on [[page-a]] and [[page-b]]...\"\n⑤ **File valuable answers back** — if the answer is a substantial comparison,\n   deep dive, or novel synthesis, create a page in `queries/` or `comparisons/`.\n   Don't file trivial lookups — only answers that would be painful to re-derive.\n⑥ **Update log.md** with the query and whether it was filed.\n\n### 3. Lint\n\nWhen the user asks to lint, health-check, or audit the wiki:\n\n① **Orphan pages:** Find pages with no inbound `[[wikilinks]]` from other pages.\n```python\n# Use execute_code for this — programmatic scan across all wiki pages\nimport os, re\nfrom collections import defaultdict\nwiki = \"<WIKI_PATH>\"\n# Scan all .md files in entities/, concepts/, comparisons/, queries/\n# Extract all [[wikilinks]] — build inbound link map\n# Pages with zero inbound links are orphans\n```\n\n② **Broken wikilinks:** Find `[[links]]` that point to pages that don't exist.\n\n③ **Index completeness:** Every wiki page should appear in `index.md`. Compare\n   the filesystem against index entries.\n\n④ **Frontmatter validation:** Every wiki page must have all required fields\n   (title, created, updated, type, tags, sources). Tags must be in the taxonomy.\n\n⑤ **Stale content:** Pages whose `updated` date is >90 days older than the most\n   recent source that mentions the same entities.\n\n⑥ **Contradictions:** Pages on the same topic with conflicting claims. Look for\n   pages that share tags/entities but state different facts. Surface all pages\n   with `contested: true` or `contradictions:` frontmatter for user review.\n\n⑦ **Quality signals:** List pages with `confidence: low` and any page that cites\n   only a single source but has no confidence field set — these are candidates\n   for either finding corroboration or demoting to `confidence: medium`.\n\n⑧ **Source drift:** For each file in `raw/` with a `sha256:` frontmatter, recompute\n   the hash and flag mismatches. Mismatches indicate the raw file was edited\n   (shouldn't happen — raw/ is immutable) or ingested from a URL that has since\n   changed. Not a hard error, but worth reporting.\n\n⑨ **Page size:** Flag pages over 200 lines — candidates for splitting.\n\n⑩ **Tag audit:** List all tags in use, flag any not in the SCHEMA.md taxonomy.\n\n⑪ **Log rotation:** If log.md exceeds 500 entries, rotate it.\n\n⑫ **Report findings** with specific file paths and suggested actions, grouped by\n   severity (broken links > orphans > source drift > contested pages > stale content > style issues).\n\n⑬ **Append to log.md:** `## [YYYY-MM-DD] lint | N issues found`\n\n## Working with the Wiki\n\n### Searching\n\n```bash\n# Find pages by content\nsearch_files \"transformer\" path=\"$WIKI\" file_glob=\"*.md\"\n\n# Find pages by filename\nsearch_files \"*.md\" target=\"files\" path=\"$WIKI\"\n\n# Find pages by tag\nsearch_files \"tags:.*alignment\" path=\"$WIKI\" file_glob=\"*.md\"\n\n# Recent activity\nread_file \"$WIKI/log.md\" offset=<last 20 lines>\n```\n\n### Bulk Ingest\n\nWhen ingesting multiple sources at once, batch the updates:\n1. Read all sources first\n2. Identify all entities and concepts across all sources\n3. Check existing pages for all of them (one search pass, not N)\n4. Create/update pages in one pass (avoids redundant updates)\n5. Update index.md once at the end\n6. Write a single log entry covering the batch\n\n### Archiving\n\nWhen content is fully superseded or the domain scope changes:\n1. Create `_archive/` directory if it doesn't exist\n2. Move the page to `_archive/` with its original path (e.g., `_archive/entities/old-page.md`)\n3. Remove from `index.md`\n4. Update any pages that linked to it — replace wikilink with plain text + \"(archived)\"\n5. Log the archive action\n\n### Obsidian Integration\n\nThe wiki directory works as an Obsidian vault out of the box:\n- `[[wikilinks]]` render as clickable links\n- Graph View visualizes the knowledge network\n- YAML frontmatter powers Dataview queries\n- The `raw/assets/` folder holds images referenced via `![[image.png]]`\n\nFor best results:\n- Set Obsidian's attachment folder to `raw/assets/`\n- Enable \"Wikilinks\" in Obsidian settings (usually on by default)\n- Install Dataview plugin for queries like `TABLE tags FROM \"entities\" WHERE contains(tags, \"company\")`\n\nIf using the Obsidian skill alongside this one, set `OBSIDIAN_VAULT_PATH` to the\nsame directory as the wiki path.\n\n### Obsidian Headless (servers and headless machines)\n\nOn machines without a display, use `obsidian-headless` instead of the desktop app.\nIt syncs vaults via Obsidian Sync without a GUI — perfect for agents running on\nservers that write to the wiki while Obsidian desktop reads it on another device.\n\n**Setup:**\n```bash\n# Requires Node.js 22+\nnpm install -g obsidian-headless\n\n# Login (requires Obsidian account with Sync subscription)\nob login --email <email> --password '<password>'\n\n# Create a remote vault for the wiki\nob sync-create-remote --name \"LLM Wiki\"\n\n# Connect the wiki directory to the vault\ncd ~/wiki\nob sync-setup --vault \"<vault-id>\"\n\n# Initial sync\nob sync\n\n# Continuous sync (foreground — use systemd for background)\nob sync --continuous\n```\n\n**Continuous background sync via systemd:**\n```ini\n# ~/.config/systemd/user/obsidian-wiki-sync.service\n[Unit]\nDescription=Obsidian LLM Wiki Sync\nAfter=network-online.target\nWants=network-online.target\n\n[Service]\nExecStart=/path/to/ob sync --continuous\nWorkingDirectory=%h/wiki\nRestart=on-failure\nRestartSec=10\n\n[Install]\nWantedBy=default.target\n```\n\n```bash\nsystemctl --user daemon-reload\nsystemctl --user enable --now obsidian-wiki-sync\n# Enable linger so sync survives logout:\nsudo loginctl enable-linger $USER\n```\n\nThis lets the agent write to `~/wiki` on a server while you browse the same\nvault in Obsidian on your laptop/phone — changes appear within seconds.\n\n## Pitfalls\n\n- **Never modify files in `raw/`** — sources are immutable. Corrections go in wiki pages.\n- **Always orient first** — read SCHEMA + index + recent log before any operation in a new session.\n  Skipping this causes duplicates and missed cross-references.\n- **Always update index.md and log.md** — skipping this makes the wiki degrade. These are the\n  navigational backbone.\n- **Don't create pages for passing mentions** — follow the Page Thresholds in SCHEMA.md. A name\n  appearing once in a footnote doesn't warrant an entity page.\n- **Don't create pages without cross-references** — isolated pages are invisible. Every page must\n  link to at least 2 other pages.\n- **Frontmatter is required** — it enables search, filtering, and staleness detection.\n- **Tags must come from the taxonomy** — freeform tags decay into noise. Add new tags to SCHEMA.md\n  first, then use them.\n- **Keep pages scannable** — a wiki page should be readable in 30 seconds. Split pages over\n  200 lines. Move detailed analysis to dedicated deep-dive pages.\n- **Ask before mass-updating** — if an ingest would touch 10+ existing pages, confirm\n  the scope with the user first.\n- **Rotate the log** — when log.md exceeds 500 entries, rename it `log-YYYY.md` and start fresh.\n  The agent should check log size during lint.\n- **Handle contradictions explicitly** — don't silently overwrite. Note both claims with dates,\n  mark in frontmatter, flag for user review.\n\n## Related Tools\n\n[llm-wiki-compiler](https://github.com/atomicmemory/llm-wiki-compiler) is a Node.js CLI that\ncompiles sources into a concept wiki with the same Karpathy inspiration. It's Obsidian-compatible,\nso users who want a scheduled/CLI-driven compile pipeline can point it at the same vault this\nskill maintains. Trade-offs: it owns page generation (replaces the agent's judgment on page\ncreation) and is tuned for small corpora. Use this skill when you want agent-in-the-loop curation;\nuse llmwiki when you want batch compile of a source directory."
    },
    {
      "name": "rss-feeds",
      "description": "Read RSS, Atom, JSON feeds; discover feeds behind a page.",
      "content": "# RSS Feeds Skill\n\nReads any RSS 2.0, RSS 1.0/RDF, Atom, or JSON Feed URL into a clean, date-sorted list of\nentries, and discovers the feed behind an ordinary page URL (`<link rel=\"alternate\">` or\nthe usual `/feed`, `/rss.xml`, `/atom.xml` paths). Standard library only, nothing to\ninstall. It does not fetch full article bodies — pass an entry's link to `web_extract` for\nthat.\n\n## When to Use\n\n- \"What's new on <blog/site>\", \"latest releases of <GitHub repo>\", \"recent posts in\n  <subreddit>\", \"read this feed\", \"does this site have an RSS feed\".\n- Building a recurring digest with `cronjob_manage` (feeds are cheaper and more stable than\n  scraping the HTML front page every run). For a persistent read/unread database across\n  many feeds install the optional `blogwatcher` skill; this skill is the zero-install read.\n- Anything where a structured list of `title / link / date / author / summary` beats a\n  rendered page: podcasts, changelogs, YouTube channels, newsrooms, forum categories.\n\n## Prerequisites\n\nNone. Python 3.10+, network access to the feed host.\n\n## How to Run\n\nRun through `terminal` with the skill-relative script path:\n\n```bash\npython3 scripts/feed.py read https://hnrss.org/frontpage --limit 10\npython3 scripts/feed.py read https://simonwillison.net/            # page URL → discovers the feed\npython3 scripts/feed.py read URL --since 2026-09-01 --json          # only newer entries, machine-readable\npython3 scripts/feed.py discover https://example.com/               # list candidate feed URLs\n```\n\n## Quick Reference\n\n| Source | Feed URL pattern |\n|---|---|\n| GitHub releases / commits / tags | `https://github.com/OWNER/REPO/releases.atom`, `…/commits/BRANCH.atom`, `…/tags.atom` |\n| Subreddit / Reddit search | `https://www.reddit.com/r/NAME/.rss`, `https://www.reddit.com/search.rss?q=…` (1 req/min anon; see `reddit-reading`) |\n| YouTube channel | `https://www.youtube.com/feeds/videos.xml?channel_id=UC…` |\n| Hacker News | `https://hnrss.org/frontpage`, `https://hnrss.org/newest?q=TERM` |\n| arXiv category | `https://rss.arxiv.org/rss/cs.CL` |\n| Substack / Medium / WordPress / Ghost | `SITE/feed`, `medium.com/feed/@user`, `SITE/rss/` |\n| Podcasts | the show's RSS URL from its hosting page (`discover` finds it) |\n\nOutput fields per entry: `title`, `link`, `published` (UTC ISO 8601), `author`, `summary`\n(HTML stripped, ≤ 2000 chars). Entries are sorted newest-first.\n\n## Procedure\n\n① If you only have a site URL, run `read` on it directly; the script discovers the feed\nand reports which URL it used (`discovered_from`). Use `discover` when you want to choose\nbetween several advertised feeds (comments feed vs posts feed, per-category feeds).\n\n② Bound the request: `--limit` for \"latest N\", `--since YYYY-MM-DD` for \"since last\ncheck\". For a cron digest persist the last-seen `published` value and pass it as\n`--since` next run.\n\n③ For full text, hand the entry `link` to `web_extract`; feed summaries are frequently\ntruncated or the first paragraph only.\n\n④ Cite the entry `link`, not the feed URL, when the result feeds a report\n(`grounded-citations`).\n\n## Pitfalls\n\n- A 200 response with HTML means the URL is a page, not a feed; the script falls through\n  to discovery automatically, but a site with no `<link rel=\"alternate\">` and none of the\n  common paths reports `no feed found` — check the site's footer or `/sitemap.xml` before\n  concluding there is none.\n- Reddit feeds share Reddit's anonymous throttle (about one request per minute per IP).\n  Chain them through `reddit-reading`, which waits out the window, when you need more\n  than one Reddit call.\n- Dates: RSS `pubDate` is RFC 822 and Atom uses ISO 8601; the script normalises both to\n  UTC. Feeds that omit dates sort to the bottom and are dropped by `--since`.\n- Some feeds are Cloudflare-fronted and 403 non-browser clients; `blocked-page-recovery`\n  handles that class.\n\n## Verification\n\n`python3 scripts/feed.py read https://github.com/NousResearch/hermes-agent/releases.atom\n--limit 1` prints one entry with a `releases/tag/` link and a `[atom]` format tag;\n`discover https://simonwillison.net/` prints an `/atom/` URL."
    }
  ],
  "memory": [
    {
      "kind": "profile",
      "content": "Scholastic Research focuses exclusively on the research process — finding, evaluating, and synthesizing scholarly sources. It is not a writing tutor and does not revise user drafts or provide discipline-specific subject-matter expertise."
    },
    {
      "kind": "profile",
      "content": "Honesty rules (hard): never fabricate citations or publication details; never write original paragraphs >2-3 sentences without clear attribution; never recommend non-peer-reviewed blogs, Wikipedia, or commercial content as primary evidence; always flag when evidence has significant limitations."
    },
    {
      "kind": "profile",
      "content": "Scope boundaries: no real-time access to subscription library systems; recommendations are based on publicly available information and standard research practice. Direct deep domain questions to discipline-specific resources."
    }
  ],
  "sharedMemory": [],
  "members": []
}