research-report
Research a technical question with probes on the local system, a web fan-out through deep-research, a grounding pass on every load-bearing claim, and a report rendered as one HTML page. Use when the user asks for a research report, a design study or an evidence-backed recommendation about their own system.
- License 许可
- ISC
- 最近更新
- 2026-11-15
- Stars
- ★ 3
- Forks
- 0
01 前言 Front Matter
- name
- research-report
- description
- Research a technical question with probes on the local system, a web fan-out through deep-research, a grounding pass on every load-bearing claim, and a report rendered as one HTML page. Use when the user asks for a research report, a design study or an evidence-backed recommendation about their own system.
02 资源文件 1
- 参考文档reference/report-core.md1.6 KB
03 SKILL.md 正文
research-report — local probes, web evidence, grounded claims, one readable page
anthropic-skills:deep-research runs the web fan-out and writes notes. This skill adds what it lacks: probes on the operator's own system, advisor checkpoints, a grounding pass, a fixed report shape and the HTML render.
Before you start
The user invoked this skill by name: go on. Claude routed it: ask once, because the run spawns several agents and a workflow, then go on after a yes.
Work in .scratchpad/research/<YYYY-MM-DD>-<slug>/, the research directory. Every file this skill names goes there.
Steps
- Probe the real system first. Read the installed source, the schema, the config, the data volumes. Save each probe as
probes/<n>-<name>.md: the exact command, its output, the date. Probes are read-only; confirm isolation before a probe that writes. Cap a long run at 5 minutes and record the cap. Done when every fact the question depends on locally has a probe file. - Plan with the advisor. Call
advisorwhen the session has it. Writeplan.md: the question in one sentence, the facets (3 to 6, no overlap), the deliverable shape and the stop rule (every facet reported, at most one gap round). - Fan out to the web. Run
anthropic-skills:deep-researchon the question and the facets fromplan.md. It writesresearch_notes/<title>/andreports/<title>.mdin the directory it used for research earlier in the conversation, else in the current working directory; when it is done, move both into the research directory. Its narrative report is an input toreport.md, not the deliverable. Without deep-research, spawn onewhetstone:whetstone-scoutper facet, all in one message, and save each claim table tonotes/<facet>.md. - Check the stop rule with the advisor. Deep-research's own gap round is the one round; here, only check that each facet has claims and record the uncovered ones as open items. Without deep-research, a facet with no claims gets one gap round.
- Ground every load-bearing claim. List each claim the answer or a decision rests on, with its source (URL,
path:lineor command), about 20 to 40 claims. Run thewhetstone:verify-claimsworkflow withargs: { question, claims, batch: 5 }. For each claim that is notCONFIRMED, correct it from the source or mark itUNVERIFIED:in the report. Save the result asgrounding.md. - Write
report.mdto the shape in report-core.md. Every number cites a probe, a[key]source orUNVERIFIED:. - Render. Run the
whetstone:html-reportskill onreport.mdtoreport.html, then read every figure it lists. Re-render after the last edit toreport.md. - Check with the advisor, then account for every agent you spawned: arrived or pending.
Hand over: the HTML path, the answer in 3 to 5 bullets, and the count of UNVERIFIED: claims.
Honesty labels
| claim | enforced by |
|---|---|
| each claim gets a verdict | code — verify-claims marks a claim with no ruling UNRULED |
| a verdict is correct | model-judgment — a grounder reads the source |
| a broken diagram stops the render | code — html-report exit 1 |
| probes are read-only | model-judgment — this prompt; no hook checks a probe |
| deep-research output ends in the research directory | model-judgment — deep-research writes to the conversation's research directory or the cwd; step 3 moves it |
| the advisor runs at steps 2, 4 and 8 | model-judgment — no hook checks the calls |
04 同仓库更多 Skills
ship
Review the dossier branch once, fold the fixup commits into the planned commits, hand the push and PR to the operator, then watch CI. Invoke when the user says "ds:ship", "ship the dossier", or when ds next exits 4.
plan
Open a dossier and write its plan with the operator, then get it checked by a fresh agent before any build. Invoke when the user says "ds:plan", "start a dossier", "plan this", "grill me" in a repo with `.scratchpad/dossier/`, or when work needs more than one commit.
suite-review
Run the monthly evidence review of this repo's plugin skills. Count each skill's opportunities and uses with the census, read what happened where a skill was not used, put each skill in one bucket, act, and write the actions as predictions with a check date. Use when the operator says "suite review" or "monthly review", or when the suite-review line on LATER.md is due.
html-report
Render a markdown report as one self-contained HTML page for a human reader, in plain English, with the answer first, a table of contents and Mermaid diagrams drawn as images. Use when the user asks for an HTML report, or for a readable page made from a markdown document.
domain-modeling
Build and sharpen a project's domain model. Use when discussing codebase terminology, writing or editing a GLOSSARY.md, or recording or editing an ADR.
codebase-design
Shared vocabulary for designing deep modules. Use when the user wants to design or improve a module's interface, find deepening opportunities, decide where a seam goes, make code more testable or AI-navigable, or when another skill needs the deep-module vocabulary.