AGENT SKILL

research

Multi-source dev research with parallel scouts. Scales effort to the question's shape, writes a plan and per-angle notes to a run folder, then synthesizes a cited report with a recommendation first. Every claim traces to a note finding with a URL. Inferred triggers are confirmed before a deep run.

License 许可
MIT
最近更新
2026-11-03
Stars
★ 2
Forks
0

01 前言 Front Matter

name
research
description
Multi-source dev research with parallel scouts. Scales effort to the question's shape, writes a plan and per-angle notes to a run folder, then synthesizes a cited report with a recommendation first. Every claim traces to a note finding with a URL. Inferred triggers are confirmed before a deep run.
when_to_use
Comparing libraries, frameworks or approaches; surveying how others solve X; "research X"; ecosystem or standards surveys that need several sources. Not for a single API or signature lookup (researcher agent, or answer directly); not for implementation tasks (/workflow); not for explaining code (answer directly).
argument-hint
[question to research]

02 资源文件 3

03 SKILL.md 正文

/research

Break a question into angles, send scouts at them in parallel, and return a report you can trust.

1. Clarify (only if it changes the research)

Ask only when the answer would change what gets researched: which options to compare, a version or time bound, the deciding criterion. Max 3 questions, all at once. Otherwise state your assumptions in plan.md and go.

2. Classify the shape

Shape Example Scouts
Single fact or API "what does useId return" 0 — answer directly or hand to researcher
Comparison "Zustand vs Jotai vs Valtio" 2–4 — one per option, plus one for criteria if needed
Survey "how do teams handle offline sync in 2026" 5–6 — one per distinct angle

More than 6 needs the user's yes first. Name the cost when you ask: multi-agent research ran at about 15x the tokens of a chat turn in Anthropic's own measurement.

Depth gate. How the skill was reached decides whether to fan out:

Reached by What happens
Explicit — the user named the skill (/condux:research, "use the research skill", "condux:research") or asked for depth in words ("deep research", "thorough", "comprehensive", "cited report", "dig into") Proceed with the classified scout count. Do not ask.
Inferred — the skill loaded because a question sounded like research (routing hook or trigger description) Single fact or API: answer directly, never ask. Comparison or survey: ask ONCE, recommendation marked: "Quick answer (one pass, about 1x cost) or deep research (N scouts, about 15x cost)?" N is the classified count. Recommend quick for a comparison of well-known options, deep for a survey or a comparison where the user's constraints decide it.

Quick answer means: the main session answers in one pass, with a few searches and fetches through the host's search and fetch tools, sources cited inline, no run folder and no scouts. Same honesty rules: a version or date on dev claims, and say what could not be verified. If the user picks quick, answer and stop; offer deep research in one line only if the answer is visibly thin.

3. Open the run folder

.condux/research/<YYYY-MM-DD>-<slug>/
  plan.md
  notes/<angle>.md
  report.md

Never overwrite a run folder. A follow-up opens a new one and names the earlier run as background in its plan.md.

Bootstrap. .condux/ is created on demand at the git root. Before the first write, check it is ignored:

git check-ignore -q .condux/ || echo "not ignored"

If not, offer once: "condux keeps its working files in .condux/ — add it to .gitignore so they stay out of your commits?" On yes, append .condux/ to .gitignore; if the user would rather not touch a tracked file, use .git/info/exclude. Never edit either without asking. Not a git repo: fall back to CWD and say so once.

4. Write plan.md

Before any dispatch, write: the sub-questions, the scout count with the reason, and every brief verbatim. Recording the notes paths here also makes a stray write by a scout visible later.

5. Dispatch — all scouts in ONE message

Parallel only works if every call goes out together. Dispatch the named scout agent and nothing else — never a general-purpose or custom-prompt agent.

A vague brief makes scouts repeat each other or wander. Each brief carries:

Part Rule
Objective one angle, with a clear boundary against its siblings
Key questions the specific things the notes must answer
Suggested sources starting points (official docs, changelogs, repos)
Constraints the user's version and time bounds, carried through unchanged
Notes path absolute: <git-root>/.condux/research/<run>/notes/<angle>.md

Wait for every scout to return before step 6. Their summaries say what was covered; the evidence is in the notes files.

6. Gap round (at most one)

Read every notes file. If a critical gap remains — one that blocks the recommendation — dispatch a single follow-up round of scouts aimed at only those gaps. Minor gaps go into the report as unknowns. A second round needs the user's consent.

7. Synthesize report.md

Read the notes from disk, not from scout summaries.

Part Contains
Recommendation / answer first, decision-ready, with the deciding reason
Comparison table when the question has options
Evidence sections claims with inline ([source](url))
Risks and unknowns unresolved conflicts and gaps, stated plainly
Sources the full list, each with its version or date

When notes conflict, present both sides — don't pick one quietly.

8. Citation check

Walk the report claim by claim. Each must trace to a note finding that has a URL. Drop any claim that doesn't, or keep it visibly flagged as unverified. No re-fetching — the notes are the evidence.

9. Deliver

In chat: 3–5 lines (the answer, the confidence, the biggest unknown) and the report path.

Offer promotion once: "Does this feed a spec? If so, which one?" On a named spec:

  1. Copy the report to specs/<that-spec>/research-<YYYY-MM-DD>-<topic>.md.
  2. Add a link row to that spec's index.md.
  3. The copy keeps external URLs only. It must never cite a .condux/ path.

No fitting spec: the report stays in .condux/.

Sequential fallback

If scout cannot be spawned — no agent tool, or no scout agent (Cursor, or Codex before the agent installer is re-run) — say so once: "no scout agent here — researching sequentially, slower, same output."

Then do the work yourself: take each angle in turn, using the host's search and fetch tools under the same rules (10–15 calls per angle, primary sources, version or date on every finding, conflicts recorded). Write the same notes/<angle>.md files in the scout's shape, one block per key question, then continue from step 7:

## <key question>
**Answer**: <1–2 lines>
**Findings**
- <claim> — <source URL> — <source version or date>
**Conflicts**
- <source A says X; source B says Y — both cited>
**Gaps**
- <what could not be found, and why>

A finding without a URL goes under Gaps.

What does NOT happen

  • No scouts for a single fact — answer it or use researcher
  • No fan-out from an inferred trigger without asking
  • No more than 6 scouts without the user's yes and the cost named
  • No dispatch before plan.md holds the briefs
  • No general-purpose or custom-prompt agent standing in for scout
  • No overwriting a run folder
  • No second gap round without consent
  • No claim in the report without a URL in the notes
  • No .condux/ path cited from a promoted copy
  • No promotion without asking which spec

04 同仓库更多 Skills

jabworks
jabworks/agentic-toolkit

orphan

暂无简介

★ 2
jabworks
jabworks/agentic-toolkit

no-blurb

暂无简介

★ 2
jabworks
jabworks/agentic-toolkit

workflow

Routes any dev task into the right execution tier (Small/Medium/Large) and carries condux's operating rules — proportional effort, lazy skill loading, soft gates, implement-yourself-by-default.

★ 2
jabworks
jabworks/agentic-toolkit

usage-trace

Checks and manages durable, local usage capture — ingest, health status, and the opt-in limit snapshot — for Claude Code and Codex. Capture itself runs automatically at session start.

★ 2
jabworks
jabworks/agentic-toolkit

toolkit-skill-standards

Use when writing or reviewing a SKILL.md for jabworks/agentic-toolkit — frontmatter budgets, the trigger contract ("Use when…" description or when_to_use field), progressive disclosure into references/, and keeping a new description out of existing skills' trigger space. Triggers include "review this skill description", "write the frontmatter", "does this collide with an existing skill", "where should this content live". Not for scaffold/sync mechanics; use toolkit-foundry.

★ 2
jabworks
jabworks/agentic-toolkit

toolkit-research-frontier

Use when planning improvement work on jabworks/agentic-toolkit itself — open problems (trigger-collision automation, YAML-strict frontmatter checks, docs-catalog enforcement), existing assets, the next three concrete steps, and the four-front library-health campaign with commands and success criteria. Triggers include "let's improve our toolkit", "let's improve this repo", "what should we improve next", "toolkit roadmap", "open problems", "run the health campaign".

★ 2