start-research
Fan-out internet research into a verified, cited brief via persona research subagents plus an adversarial reviewer. Use to research a topic or gather prior art before a design.
- License 许可
- MIT
- 最近更新
- 2026-10-26
- Stars
- ★ 0
- Forks
- 0
01 前言 Front Matter
- name
- start-research
- description
- Fan-out internet research into a verified, cited brief via persona research subagents plus an adversarial reviewer. Use to research a topic or gather prior art before a design.
02 资源文件 8
- 脚本scripts/brief_path.sh555 B
- 脚本scripts/doc_path.sh1.1 KB
- 其他tests/test_brief_path.bats1.9 KB
- 脚本tests/test_brief_path.sh2.5 KB
- 其他tests/test_doc_path.bats2.5 KB
- 脚本tests/test_doc_path.sh3.4 KB
- 脚本tests/test_research_agents.py2.2 KB
- 脚本tests/test_start_skills_structure.py18.5 KB
03 SKILL.md 正文
start-research
Researches a topic across the internet using several persona research subagents
in parallel and one adversarial reviewer subagent, then synthesizes a verified,
cited brief at docs/research/<date>-<slug>.md. Spawns subagents via
the Agent tool - never the Workflow tool.
Procedure
-
Scope. Restate the topic in one line. Pick the researchers from the
lensesargument (any subset ofdocs-purist,practitioner,prior-art-scout,paper-scout):- Default
lenses: the three core personas (docs-purist,practitioner,prior-art-scout). Addpaper-scoutwhen the topic has a research-paper dimension. - Broad topic: split one lens across two researchers with different
sub-scopes. A topic focus goes to
workbench:research-practitioneras a focus line in its prompt; no custom persona is run. Hard cap of 5 researchers.
- Default
-
Research (parallel). Get each lens's output path first:
bash ${CLAUDE_PLUGIN_ROOT}/skills/start-research/scripts/doc_path.sh research "<topic>-<persona>". Spawn all researchers in a SINGLE message (multipleAgentcalls) so they run concurrently. Each issubagent_type: workbench:research-<persona>(e.g.workbench:research-docs-purist) withmodel: sonnetset explicitly in the call. The persona's method and output contract live in its agent file; the prompt carries only the topic and the output path for that lens's brief. -
Review (adversarial). Get the reviewer's output path with
doc_path.sh research "<topic>-review". Spawn onesubagent_type: workbench:research-reviewerwithmodel: sonnetset explicitly. Its prompt carries only the topic, every lens brief path, and its own output path. It marks each claim supported / unsupported / contradicted and lists gaps and missing angles. -
Follow-up (bounded). If the reviewer reports material gaps, run ONE more round of targeted researcher(s) scoped to those gaps, then a quick reviewer re-check. Do not loop further.
-
Synthesize and write. Drop reviewer-killed claims. Get the brief path:
mise exec -- bash ${CLAUDE_PLUGIN_ROOT}/skills/start-research/scripts/doc_path.sh research "<topic>"(
${CLAUDE_PLUGIN_ROOT}/skills/start-research/scripts/brief_path.sh "<topic>"is the equivalent path helper.) Write the brief there and return its path plus an inline executive summary.
Brief format
Markdown: title + date + topic; executive summary; themed sections with inline
[n] citations; explicit "Best practices" and "Prior art" sections; an "Open
questions / gaps" section from the reviewer; a numbered "Sources" list. Every
non-trivial claim cites a source; killed claims do not appear.
The brief ends with a mandatory final ## Plan inputs section, defined once in
phase 1 of ${CLAUDE_PLUGIN_ROOT}/skills/start-development/SKILL.md (lists: reuse,
do-not-build, fix-as-ac, ideas).
Notes
- Runnable standalone (
/start-research <topic>) and called bystart-developmentduring its discovery and design phases. - The four
research-*skills (research-docs,research-code-external,research-code-internal,research-papers) are the single-lens entry points; use this skill when several lenses are wanted at once. - Always set each subagent's model explicitly (do not inherit the session default).
docs/research/is gitignored; briefs stay local.
04 同仓库更多 Skills
claudy
d claudy
gamma
Literal line one. Literal line two.
beta
Quoted: description with a colon.
alpha
Folded description that spans two lines and ends here.
worktree-merge
Rebases one subagent's worktree onto the base branch, optionally verifies the rebased tree, fast-forwards it into the worktree that has the base checked out (an epic integration tree or the main checkout), verifies every commit landed, and only then removes the worktree and its branch. Keeps the worktree intact on any conflict, dirty tree, or failed check. Use after a subagent returns work created with worktree-create.
worktree-create
Creates one isolated git worktree for one subagent, at <workbench>/.claude/worktrees/<name>, branched from main with the newest origin/main and local main merged in. Refuses a name already in use so two agents can never share a tree. Use once per implementation or review subagent before dispatching it; pair with worktree-merge to land the work and clean up.