Deep research ✔ security-verified · 2 models
Deep research A traffic-light research method with three tiers of quality checks. Every task runs the RED mandatory elements. Higher-stakes work adds YELLOW conditional checks. Urgent, high-risk decisions trigger GREEN elements. The output is a one-page executive summary on top of fully cited research. How it works - Break the question into atomic sub-questions - List every h
Security verification log
| Model | Verdict | Notes | Date |
|---|---|---|---|
| claude-fable-5 | secure | Read in full: research quality methodology (triangulation, contrarian search, convergence). Pure method; no network writes, secrets, or cost. | 2026-07-11 |
| claude-sonnet-5 | secure | Reviewed full markdown: pure research methodology (decompose, hypothesize, triangulate, contrarian search). No exfiltration, no secrets, no network/download calls, no prompt-injection or hidden direct | 2026-07-11 |
Weekly usage scores
No weekly usage scores reported yet.
Skill preview
Show raw markdown (2627 bytes)
# Deep Research Traffic-light research system with 37 quality elements across three tiers. Every research task gets 14 mandatory RED elements. Higher-stakes work adds 15 conditional YELLOW elements. Life-safety and major decisions trigger 8 GREEN elements. **When to use this** Any research task where the answer matters — product decisions, market analysis, competitive intelligence, investment research, fact verification. If you are about to make a decision based on what you found, use this. **The process** 1. **Decompose** the question into atomic sub-questions. Rank by importance. Research each independently. 2. **Generate hypotheses** before researching. List all plausible explanations. Rate evidence against each, not just your favourite. 3. **Map stakeholders.** Who benefits? Who loses? Who funds? Who is incentivised to mislead? 4. **Triangulate.** Three independent sources with different biases per key claim. If sources that disagree on everything else agree on THIS, it is probably true. 5. **Check freshness.** AI news: 48 hours. Academic papers: 2 years. Market data: 30 days. Historical facts: no expiry but verify originals. 6. **Contrarian search.** Spend 20% of your time finding evidence AGAINST your emerging conclusion. If you cannot find any, you have not looked hard enough. 7. **Branch.** When a search leads somewhere unexpected, follow that thread. Maintain parallel paths. Merge results. Prune dead ends. 8. **Minimum 10 cycles** of search-refine before declaring research complete. Detect convergence (when new results stop adding new information) and stop. **Output format** - Executive summary: 1 page, easy to scan, bold key words, bullet points, confidence level, what is still unknown - Full research underneath: as long as needed, source citations throughout, insight stacking (each section builds on the previous, no circling) **Pre-prepared sub-agents** **1. Research agent** — searches sources, gathers facts, writes findings with provenance **2. Hypothesis analyst** — lists all plausible explanations, rates evidence against each using competing-hypotheses methodology **3. Synthesis agent** — triangulates across sources, identifies convergence, writes the executive summary **Tips** - The first hypothesis that feels right is usually the most biased. Always generate alternatives before researching. - "Statistically significant" with tiny effect size is meaningless. Always ask "how much?" not just "is there an effect?" - An unverified URL claim is a false claim. Always check that links resolve before presenting them. **Tags:** Analysis & research, Education