Read the manual
Use when the user wants an answer sourced directly from the indexed document corpus, not from model knowledge. Triggers on "/read-the-manual", "read the manual", "check the docs", "what does the doc/manual say about X", "find in the corpus", "answer from the docs".From its SKILL.md
npx -y skills add canlin-zhang/rtfm --skill read-the-manualAssembled from the repository path, not quoted from the project. Check it against their README if it does not work.
One thing to look at
- 1 stars1 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.
SKILL.md
3.0 KB, as published. Nobody here has run it
Read the Manual
Answer the user's question using only text found in the indexed corpus via rtfm. Do not answer from training knowledge. Every claim must be backed by a direct quote with a source, file, and format-native locator (page for PDFs, line for text).
Trigger
- User invokes
/read-the-manual. - User says "read the manual", "check the docs", "what does the manual say about X", "find in the corpus", or similar phrasing signalling they want corpus-sourced truth rather than model knowledge.
Routine
1. Search
Call mcp__plugin_rtfm_rtfm__search(query) with the key terms from the question. Note sources_searched to stay aware of the full corpus. Small new or edited files are auto-indexed on search; if the response carries a STALE SOURCE warning (a source has more new/changed files than the auto-reindex budget), build it yourself with mcp__plugin_rtfm_rtfm__reindex('<source>'), then search again.
If the first query returns nothing useful, try narrower or alternative terms — don't give up after one miss.
2. Read
For each promising hit, call mcp__plugin_rtfm_rtfm__read(source, relpath, start, end) to retrieve the actual text around the hit's locator: read ±1–2 pages (PDF) or a surrounding line range (text) for full context. A hit's locations lists up to a few of the paths its content lives at (total_locations is the true count; find_duplicates gives the full list) — any one reads identically.
3. Answer
- Quote the corpus text verbatim for key claims, in block quotes.
- Cite every quote with source, file, and locator —
— *relpath* (source), p.Nfor PDFs,— *relpath* (source), line Nfor text. - If several documents cover the topic, synthesize but attribute each point separately.
- If the corpus has no clear answer, say so explicitly: "The indexed corpus doesn't cover this." Do NOT fill the gap with model knowledge.
4. Report missing references
After answering, scan the text you read for documents referenced by name that are not among the sources you searched. Report them compactly:
Referenced but not in corpus: <Document Title>, <Document Title>
Include only documents cited directly by name, not vague phrases ("see the documentation"). If nothing is missing, skip this step silently.
Rules
- No knowledge fallback. If it isn't in the corpus text you read this session, don't state it as fact.
- No paraphrasing without citation. If you summarize, immediately follow with the locator.
- Prefer verbatim quotes for precise technical claims (field widths, timing constraints, opcodes, state transitions, …).
- If the question is ambiguous, ask one clarifying question before searching.