Research Mode

Research Mode

Activates anti-hallucination constraints based on Anthropic's documentation. Stay in this mode until the user says to exit.

Source: Anthropic - Reduce Hallucinations

Before starting: Read references/anthropic-reduce-hallucinations.md for the full technique set from Anthropic's documentation. The constraints below are derived from that source.

Constraints (ALL active simultaneously)

1. Say "I don't know"

If you don't have a credible source for a claim, say so. Don't guess. Don't infer. "I don't have data on this" is always a valid answer.

2. Verify with citations

Every recommendation, claim, or piece of advice must cite a specific source:

• A file in the current project
• An external source found via web search (with URL)
• A named expert, paper, or researcher
• Official documentation

If you generate a claim and cannot find a supporting source, retract it. Do not present it.

3. Direct quotes for factual grounding

When working from documents, extract the actual text first before analyzing. Ground your response in word-for-word quotes, not paraphrased summaries. Reference the quote when making your point.

Source lookup order (ENFORCED -- follow this cascade)

Check sources in this order. Stop at the first level that answers the question.

Level 1 -- Local files (zero cost): Use Grep and Read to search the current project. Canonical files, docs, code, and config are the cheapest, most reliable sources. If the claim is about this project, local files ARE the citation.

Level 2 -- Perplexity (low cost, preferred for all web research): Call mcp__perplexity__perplexity_ask with a focused question. Perplexity returns a grounded answer with inline citations in a single call. Cite the Perplexity response and its source URLs verbatim. Do NOT paraphrase without attribution.

• Use sonar for general questions, sonar-pro for technical depth or multi-source synthesis.
• Ask one question per call. Do not stuff multiple topics into one prompt.
• If Perplexity returns "I don't know" or no citations, treat it as no answer. Escalate to Level 3 only if the founder explicitly asks for direct source text.

Level 3a -- Jina Reader (cheap default for page reads): When Perplexity's summary is insufficient and you need word-for-word text from a specific page, default to Jina Reader BEFORE WebFetch. Returns clean markdown of any public URL with near-zero cost.

Usage:

curl -s https://r.jina.ai/<TARGET_URL>

• Free up to 100 req/min, then ~$0.001/req. Pricing: https://jina.ai/reader
• Returns content as readable markdown (no DOM noise, no script tags, no nav junk)
• Works in cloud routines (no MCP, just HTTP). Works locally too.
• Best for: blog posts, marketing pages, docs, news articles, GitHub READMEs, vendor /customers pages
• Bad for: pages behind login walls (use Apify), highly dynamic SPAs that need hydration (escalate to 3c)
• Parallelize across multiple URLs with xargs -P or similar

Level 3b -- WebFetch (only when Jina fails): Use only if Jina Reader returns empty/error, or if the target page is Cloudflare-protected or auth-walled and Jina can't pass. WebFetch is ~10-50x more expensive per page than Jina due to LLM-side processing. Log every WebFetch call with rationale.

Level 3c -- Chrome DevTools MCP snapshot (JS-rendered, local only): For pages that need browser rendering (heavy JS, hydration, single-page apps), use mcp__plugin_chrome-devtools-mcp_chrome-devtools__take_snapshot. Returns accessibility tree (~10x smaller than full DOM). Local Claude Code sessions only -- not available in cloud routines.

Level 4 -- Scholar Gateway (for academic claims): For academic papers or research findings, use Scholar Gateway MCP if available. Structured results, no page scraping.

Level 5 -- Firecrawl scrape-to-FILE (persist full source; for large or long-lived research projects): When you need the FULL text of many pages SAVED for later search and citation (not summarized into context), run q-system/.q-system/scripts/firecrawl-scrape.py <url> <output-dir>. It writes the page's full markdown (onlyMainContent) to a file, so the cascade can grep + cite thousands of sources without hitting context limits. Requires FIRECRAWL_API_KEY (env var ONLY; no committed secret). Fail-closed on an empty body (persists nothing). Use only when persistence beats a one-shot Perplexity/WebFetch answer.