| name | x9-research |
| description | Use when the user wants a genuinely researched answer from current or primary sources, including X/Twitter or Reddit — «разберись в вопросе», «изучи тему», «собери факты», "research this", "verify this claim". Also use for durable research intent in varied wording — «создай ресерч в папке», «сохрани исследование», «обнови ресерч», "save/update this research" — and for explicit HTML delivery — «ресерч по теме X + html», «подготовь HTML-отчёт с источниками», «нужна HTML-версия», "research X and deliver it as HTML". Infer semantic intent rather than requiring an exact trigger phrase. Do not use for library/API documentation (find-docs), product search owned by another skill, personal advice, or questions answerable from stable supplied context. Mentioning HTML only as the research subject does not request HTML files. |
| compatibility | Requires Python 3 for deterministic HTML output validation. |
Research
Treat prior knowledge as a hypothesis. Open live sources, trace every load-bearing claim, search for disconfirming evidence, and label uncertainty.
Method
- Define the real question and the claims the answer depends on. Ask only when an unresolved ambiguity would change what is researched.
- Gather breadth first, then read the strongest primary or authoritative sources in depth.
- Match evidence to claim type:
- A direct fact from its authoritative registry, specification, filing, or law may need one primary source.
- Contested, inferential, surprising, or high-stakes claims need independent corroboration and an active search for disconfirmation.
- Social sentiment requires deliberately sampled primary posts across relevant queries, dates, and views, plus a clear sampling limitation.
- Resolve source conflicts explicitly; do not average incompatible claims.
- Answer with conclusion first, evidence near each claim, freshness, and calibrated confidence.
Read references/methodology.md for evidence thresholds and independent verification. Read references/sources.md when selecting search/browser surfaces.
Runtime and browser
Use the live tools exposed by the current runtime; do not assume Claude-only names. Prefer purpose-built connectors/APIs for structured resources. Before browser work, follow x9-browser-session.
Read references/sources.md before browser or social research. It owns source selection, platform sampling, and the intentional authenticated-browser exception for X/Twitter; x9-browser-session owns controller selection, task-tab isolation, private-surface safety, and browser-route failure.
Route the result
Return a sourced answer in chat unless the request conveys persistence, update, or HTML-output intent. For any such intent, read references/delivery-modes.md; it owns classification, OKF discovery, artifact lifecycle, authority, ambiguity, and completion. When it selects HTML, also read references/html-reports.md and use assets/editorial-theme.css for shared visual tokens and optional components, never as a layout template.
Failure behavior
- Report a failed or cancelled search/tool call; do not fill the gap from memory.
- If the authoritative source is inaccessible, label the claim unverified and state the best next retrieval step.
- If evidence remains mixed, preserve the disagreement and lower confidence.
- If all load-bearing sources are unavailable or the conclusion cannot be checked, return
BLOCKED/NOT_PROVEN; do not present a memory-based answer as a degraded research result.
Done
- Every load-bearing current claim traces to a source opened in this run.
- Evidence strength matches claim risk; contradictions and freshness are visible.
- The applicable completion contract passes in delivery modes, HTML reports, and source selection.
- High-stakes or durable conclusions include the independent-review disposition required by the methodology.
