| name | grok-tool-augmented-research |
| description | Use Grok Build's current web_search/web_fetch tools as a bounded external-evidence layer inside coding, debugging, dependency evaluation, and autoresearch loops. Use for "Grok์ผ๋ก ๊ฒ์ํด", fresh release/API/CVE evidence, recent community signal, or search-grounded hypotheses. Treat dedicated X corpus search as optional and only claim it when an installed MCP/tool explicitly provides it; never leak private repo data. |
Grok Tool-Augmented Research
Use fresh external evidence only when it can change the next engineering decision. Search
belongs inside observe โ hypothesize โ verify โ act, not in a disconnected research dump.
Version-matched Grok docs are under ~/.grok/docs/. In current Grok Build, the documented
built-in research tools are web_search and web_fetch; --disable-web-search removes
both. A dedicated X/Twitter corpus search is not guaranteed by those tool names. Use one
only when grok inspect shows a configured MCP/plugin tool that explicitly offers it.
Decision gate
Search when the answer is volatile or external:
- current provider/API behavior, releases, deprecations, or compatibility;
- recent CVEs and maintainer advisories;
- unresolved errors tied to a specific current version;
- dependency activity, adoption friction, or a claim that materially affects a hypothesis.
Do not search for stable language rules, facts already proven by the checked-out source, or
questions that can be answered by running a local test.
Procedure
- Name the decision. Write the exact choice the evidence will influence.
- Sanitize the query. Remove private code, customer data, secrets, internal hostnames,
unpublished product names, incident details, and unique stack traces.
- Choose sources. Prefer official docs/release notes/advisories and upstream issues.
Use community posts only as secondary evidence of real-user friction.
- Bound the run. Use a small query budget, read-only tools, no subagents unless needed,
and a max-turn limit.
- Record evidence. Keep query, retrieval date, URL/post ID, relevant version/date, and
the exact claim supported. A search snippet alone is not final evidence; fetch the source.
- Resolve contradictions. Compare publication dates, versions, primary-source status,
and whether claims describe the same environment.
- Close the loop. State
decision โ evidence โ confidence โ next local verification.
- Verify locally. External evidence informs the hypothesis; tests, source inspection,
and runtime behavior decide whether it applies here.
Bounded headless recipe
grok -p '
Research the current documented behavior relevant to this decision.
Prefer primary sources, cite every material claim, state contradictions,
and end with decision, confidence, and the next local verification.
' \
--tools 'web_search,web_fetch,read_file,grep,list_dir' \
--disallowed-tools 'run_terminal_cmd,search_replace,write_file,Agent' \
--max-turns 20 \
--output-format json
This recipe permits external disclosure of the prompt text. Keep the prompt generic and
public-safe. For confidential repositories, describe only the abstract technology question.
X/social signal
X can be useful for current DX pain, outages, adoption friction, and statements by official
maintainers. But distinguish these cases:
web_search happens to return an x.com URL โ ordinary web evidence;
- an installed MCP/plugin exposes a dedicated X search tool โ corpus-specific evidence;
- no such tool exists โ do not claim an X search was performed.
For social evidence, preserve the post URL/ID, author, timestamp, and whether it is firsthand.
Do not treat likes, reposts, or a single viral post as technical validation.
Evidence note template
Decision: <what changes if the claim is true>
Query: <sanitized exact query>
Retrieved: <UTC date>
Primary sources:
- <URL> โ <supported claim, version/date>
Secondary signals:
- <URL/post ID> โ <what users report; not authoritative>
Contradictions/limits: <missing, stale, conflicting, inaccessible>
Conclusion: <decision and confidence>
Local verification: <test/source/runtime probe>
Autoresearch composition
- autoresearch owns the fixed experiment budget, baseline, one-change discipline, and
metric honesty.
- This skill may spend a bounded evidence call during hypothesis formation.
- Search findings never count as an experiment result. Log the query/URLs, then measure the
candidate change against the unchanged evaluation procedure.
Guardrails
- Never send confidential source, credentials, private incidents, or proprietary strategy
logic to web/X search.
- Never cite an AI summary as the primary source when the underlying document is available.
- Never silently use the current year as a proxy for recency; apply an explicit date window.
- If evidence is weak, contradictory, paywalled, or inaccessible, say so and lower confidence.
- Keep URLs in notes/docs when they materially justify a code or architecture decision; avoid
noisy permanent comments for disposable search context.