agent-briefing
Use when writing a prompt for a subagent or teammate, deciding how much context another agent needs, writing or reading a handoff report between agents, or sending a message to another agent mid-task.
NoobyGains · 已有静态检查记录
安装与使用
运行命令前先阅读来源与依赖说明。复制命令不会自动安装或执行。
npx skills add NoobyGains/godmode --skill agent-briefing技能说明
name: agent-briefing description: "Use when writing a prompt for a subagent or teammate, deciding how much context another agent needs, writing or reading a handoff report between agents, or sending a message to another agent mid-task."
Agent Briefing
Most multi-agent failures are briefing failures: the worker lacked one fact, drowned in irrelevant history, guessed at a constraint, or returned a report nobody could check. A dispatched agent knows only what its brief says, and you know only what its report says. Both handoffs lose detail unless they are written on purpose.
Core principle: write every brief for a capable stranger with no memory of this session, and require every report as a file whose claims can be checked. Give exactly the context the task needs, point to files instead of pasting them, and ask for evidence, not reassurance.
The Prime Directive
NO DISPATCH WITHOUT A COMPLETE BRIEF, NO HANDOFF WITHOUT A REPORT FILE
No exceptions. No workarounds. No shortcuts.
When to Use
Required:
- Dispatching any subagent, teammate or other-model reviewer.
- Resuming an agent with findings or new information.
- Messaging a teammate in a team.
- Reading a worker's report before acting on it.
Skip:
- A one-line lookup you can do yourself in a couple of actions: do it, do not delegate it.
The Brief
Every brief has these eight parts, in this order. Leave a part out only when it truly does not apply, and write "none" rather than omitting it silently.
| Part | What it says | Example |
|---|---|---|
| 1. Role | Who the agent is in this run, in one line | "You are the implementer for Task 3 of the export plan." |
| 2. Task | The one job and its deliverable | "Add CSV export to the report page." |
| 3. Context | Where this fits; decisions already made; what was ruled out | "Task 2 added toRows(); use it. JSON export was rejected." |
| 4. Files | Paths to read first, paths in scope, paths out of scope | "Read task-3-brief.md. Change only src/export/*." |
| 5. Constraints | Rules that bind it | "No new dependencies. Do not dispatch subagents." |
| 6. Verification | The exact command and what passing looks like | "npm test -- export passes, output clean." |
| 7. Deliverable | The report file path and the reply format | "Write task-3-report.md; reply with status, commits, one-line test result." |
| 8. Questions | Whether and when it may ask | "Ask before starting if anything is unclear; stop with NEEDS_CONTEXT if you get stuck." |
Name the model and effort on the dispatch itself, chosen per godmode:model-routing. A dispatch without them inherits the session's model.
Lean Versus Full Context
Lean (the default). The brief names the task, the interfaces it touches, the global constraints and the files to read. The agent reads the files itself. Use this for implementers working from a task brief file, reviewers working from a review package, and investigators with a clear problem.
Full. The brief also carries the reasoning behind decisions, the alternatives rejected and why, and the relevant history. Use it when the agent must make judgment calls the plan did not settle, when a previous agent failed and the next must not repeat the attempt, or for a whole-branch or cross-model review that needs the intent.
Never in a brief, either way:
- The session's history ("state after Tasks 1 to 4" pasted in): a fresh agent needs its task, its interfaces and its constraints.
- Content you can point to: plan text, diffs, logs, long files. Write them to a file and give the path; anything pasted into a dispatch also stays in your own context for the rest of the session.
- Your conclusion, when you want an independent check. Hand a reviewer the code and the question, not your hypothesis; an agent handed a hypothesis tends to return it confirmed.
- Instructions that pre-judge a finding ("do not flag X", "at most Minor").
The Report-File Rule
Every worker writes its full report to a file you named in the brief, and replies with a short contract only:
Status: DONE | DONE_WITH_CONCERNS | BLOCKED | NEEDS_CONTEXT
Commits: <short sha> <subject>, one per line
Tests: <one line, for example "14/14 passing, output clean">
Concerns: <one line each, or "none">
Report: <path>
You read the reply first and open the report only when the reply raises a concern, a reviewer needs it, or a fix round needs its history. The report file is also the persistent memory: a fresh agent taking over a task reads it to see what was tried. Report files live in the run's git-ignored workspace (see godmode:delegated-execution).
Handoff Report Formats
Implementer to controller or reviewer
## What was built what the brief asked for, and what was done
## Files changed path: one-line reason
## Tests command, output (including the failing run when test-first applied)
## Self-review what was checked and what changed because of it
## Concerns deviations from the brief and why; open doubts
Reviewer to controller
## Spec compliance PASS or FAIL with file:line; "cannot verify" items
## Strengths specific
## Issues Critical / Important / Minor, each with file:line, why, fix
## Verdict APPROVED or NEEDS FIXES, one-line reason
Controller to implementer (fix request)
## Findings to fix verbatim from the review, one per bullet
## Covering tests which tests to re-run for each finding
## Out of scope what not to touch
## Append to the same report file, as a fix report
Investigator to lead
## Hypothesis what was tested
## Evidence commands, output, file:line, for and against
## Confidence high / medium / low, and what would change it
## Recommendation next step
## Cross-cutting anything another investigator or teammate should know
阅读完整来源 ↗