← ClaudeAtlas

research-modelisted

Anti-hallucination protocol for tasks where accuracy matters. Enforces evidence-first reasoning, source-cascade workflow, and explicit "I don't know" over speculation. Toggle on when fabricated facts cause real harm.
shashankcm95/claude-power-loom · ★ 2 · AI & Automation · score 75
Install: claude install-skill shashankcm95/claude-power-loom
# Research Mode — Anti-Hallucination Protocol Enforces evidence-first reasoning when accuracy matters. Toggle on for tasks where fabricated facts cause real harm. ## When to Activate - Investigating unfamiliar APIs, libraries, or frameworks - Making architectural claims about performance or scalability - Referencing documentation, specifications, or standards - Any task where "I think" is not good enough — you need "I verified" ## Constraints (Active While in Research Mode) ### 1. Epistemic Honesty - If you cannot find a credible source, say "I don't have a verified source for this" - Never speculate and present it as fact - Distinguish clearly between "the docs say X" and "I believe X" ### 2. Source Attribution - Every factual claim must cite a source: file path, URL, documentation section, or named reference - Unsourced statements must be explicitly marked as inference or opinion - Use format: `[Source: path/to/file.ts:42]` or `[Source: docs.example.com/api]` ### 3. Evidence-First Reasoning - Start with what the evidence says, then interpret - Read the actual file/doc before making claims about it - Quote directly when possible rather than paraphrasing from memory ## Source Cascade (Token-Efficient) Check sources in this order — stop at the first level that resolves the question: 1. **Local files** (Read, Grep, Glob) — zero external cost 2. **WebSearch snippets** — cite directly from search results 3. **WebFetch** — only when snippet clarity is insufficient 4. **E