debrief
Skill2 starsChanged 4 months agoMITView on GitHub
Post-dispatch verification. After a dispatched sub-agent, workflow step, or dispatch-charter (FR-NNNN) returns, verify it actually achieved the intent, that its findings were persisted to disk (not left only in the dispatch's conversation), and that production-reachability holds (IMPL vs WIRED). Updates the traceability ledger + log.md. Use after EVERY dispatch before the orchestrator treats the work as done. Catches the #1 recurring failure — code that compiles and passes tests but is never wired.
30 other projects use a version of this.
Adopted by
Most recently taken up by
- srizzon/skillstheir own version, first seen Nov 8, 2026
- kiloloop/oacp-skillstheir own version, first seen Nov 1, 2026
- Works-HQ/ai-brain-starter-packtheir own version, first seen Nov 1, 2026
- piotrmechlinski/executive-braintheir own version, first seen Nov 1, 2026
- pedroromeroluna/peter-braintheir own version, first seen Nov 1, 2026
- bks-lab/open-bridgetheir own version, first seen Oct 30, 2026
Used with
Projects with it also use brief and review.
Other files people use alongside this one.
Files found in the same public projects as a version of this one, more often than they turn up anywhere else. Each project counts once, large collections count for less, and this file's own repository is left out.
AnthropicSkillbrief
Generate contextual briefings for legal work — daily summary, topic research, or incident response. Use when starting your day and need a scan of legal-relevant items across email, calendar, and contracts, when researching a specific legal question across internal sources, or when a developing situation (data breach, litigation threat, regulatory inquiry) needs rapid context.
Also in 3 projects that use this
SentrySkillreview
Fixture review skill.
Also in 3 projects that use this
Other versions30 projects use a version of this.
Word for word · 0
None found.
Edited · 0
None found.
Wrote their own version · 30
- suboss87/FDEOps★ 940same name
- spacedock-dev/spacedock★ 85same name
- bks-lab/open-bridge★ 13same name
- pedroromeroluna/peter-brain★ 9same name
- trydebrief/debrief-skills★ 4same name
- uttambharadwaj/kb-graph★ 4same name
- adamlinscott/claude-skills★ 3same name
- Works-HQ/ai-brain-starter-pack★ 3same name
- siemendev/release-process★ 2same name
- aaronwskinner/claude-skills★ 1same name
20 more not listed
How matches are found. Word for word: the exact same text. Edited: shares passages with this file (“99% same” means 99% of its five-word phrases also appear here; 10% or more counts). Their own version: similar meaning, or, weakest, the same name. We can’t tell which project had it first. Projects are counted among the 200 most-starred candidates, and mentions are links to the repository we found, up to 50 per site, so both counts are a floor.
In this repository11 other files it gives its agents.
- CLAUDE.mdCLAUDE.md
- distill-lessonsSkill
- explain-modelmuxSkill
- flushSkill
- getting-startedSkill
- handoffSkill
- kit-doctorSkill
- kit-upgradeSkill
- orientSkill
- sitrepSkill
- switch-modelsSkill
The file
Show its rules and text
We split this skill into the rules it gives an agent
13 rules. 0 are common, 7 are distinctive, and 6 are very specific to this project.
0
Common
Most of the 27 edited versions we read keep these.
7
Distinctive
This author's own choices. Worth a look.
6
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Distinctive: line 85
If the intent was NOT achieved → the work is BLOCKED, not done.
- Distinctive: line 99
If a finding is only in the conversation, FILE IT NOW (or note it as a debrief gap).
- Very specific: line 130
Write/update the work-unit's row in the traceability/ ledger (one row per work-unit, UPDATE-IN-PLACE).
- Very specific: line 131
UPDATE-IN-PLACE). Columns: WU-NNNN · FR-id · deliverable · IMPL/WIRED/DEFER · the reachability evidence (the prover-menu query output ref, …
- Very specific: line 133
If traceability/index.md does not yet list the ledger, add it in the same change (index completeness).
- Very specific: line 150
Use the dispatch op for the debrief line.
- Very specific: line 158
If gaps: the specific remediation (re-dispatch with corrected charter / file the orphaned finding / escalate).
- Distinctive: line 165
Do NOT accept "tests pass" as WIRED — trace production-reachability (the #1 recurring trap).
- Distinctive: line 166
Do NOT let the executing agent debrief its own dispatch — separation of duties (CONVENTIONS.md §7.2).
- Distinctive: line 168
Do NOT leave a dispatch's findings in its returned text — persist or lose them.
- Very specific: line 169
Do NOT mark a WU done on an IMPL-only row without an explicit DEFER rationale.
- Distinctive: line 170
Do NOT skip the traceability update — the ledger is the durable record that the loop was closed.
- Distinctive: line 171
Do NOT include secrets, credentials, or regulated / user data — reference paths.
“Common” is measured: kept by at least 6 in 10 of the 27 edited versions of this file we could read. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
/debrief — verify a dispatch achieved intent and the findings survived
Standard / Full — only relevant if you dispatch sub-agents or run fan-out workflows. If you work solo in a single session, you don't need this skill. It is the orchestrator-side gate you run after handing work to another agent.
A dispatch (a dispatch-charter FR-NNNN, a sub-agent call, a workflow step) returns a claim of done.
A returned claim is not a verified outcome. /debrief is the gate that closes the loop: did the
dispatched work achieve the intent, did its findings reach disk, and is the result actually reachable
from a production path?
This back-ports three durable disciplines the orchestrator must not skip after fan-out:
- IMPL→WIRED — verification by production-reachability, not test-pass (the #1 recurring failure).
- Findings-to-disk-or-they-don't-exist — a sub-agent's findings that live only in ITS conversation are dead the moment the dispatch returns.
- The executor never audits its own work —
/debriefis run by the dispatching orchestrator (or a clean-context verifier), never by the agent that did the work.
Foreground in the main agent / orchestrator. The orchestrator holds the dispatch intent and the brief; it is the right place to compare returned-against-intended.
When to invoke
- After EVERY dispatch returns — a dispatch-charter, a sub-agent, a workflow step. Before marking the work-unit done.
- At a phase gate — when a wave of charters has returned and you are about to merge / proceed. Debrief each leg.
- When a sub-agent reports "done" but you have not seen the artifact on disk — the most common failure shape.
- After a research-pipeline track or a workflow fan-out — confirm each leg's findings landed in its tier file / owning surface.
When NOT to invoke
- For a read-only sub-agent call whose only output is a verbal answer you have already captured (no claimed artifact, no wiring claim) — there is nothing to verify against disk.
- Mid-dispatch — wait for the leg to return.
- As a substitute for the research-pipeline's adversarial gate — that gate checks CLAIMS against
primary sources;
/debriefchecks DISPATCH OUTCOMES against intent + disk + reachability. Different jobs; a research investigation gets both.
Steps
1. Recover the dispatch intent (the ground truth to verify against)
The persisted brief is ground truth, not the conversation. For each dispatch being debriefed, pull:
- the dispatch-charter (
.agent-docs/dispatch-charters/<FR-id>.md, or the dispatch prompt itself in the Standard profile) — its Task, Acceptance criteria, and wiring-proof target (the production entrypoint the work must be reachable from). - the work-unit it serves (
WU-NNNN).
pwd && git rev-parse --abbrev-ref HEAD # cwd-check before any verification ops
cat .agent-docs/dispatch-charters/<FR-id>.md 2>/dev/null | sed -n '1,60p' # the charter's intent + acceptance2. Did it achieve the intent? (returned-vs-intended)
Compare what the dispatch CLAIMS it did against the charter's Task + Acceptance. For each acceptance criterion, find the on-disk evidence — do not take the dispatch's word:
git status -s && git log --oneline -5 # what actually landed in the tree
git diff --stat HEAD~1 2>/dev/null # diff shape vs the charter's expected diff shape- Deliverables present? The files the charter said it would produce — do they exist, with real content (no stubs / TODOs / not-implemented placeholders)?
- Gates green? The gates are the acceptance bar:
bun run build/bun run lint/bun test test/(+ ``). Confirm the dispatch actually ran them and they passed — cite the result, don't assume. A behavior change OWES a test that fails without it. - Scope honored? Did it stay inside its one-file-one-owner boundary, or did it touch files another leg owns (a fan-out collision)?
If the intent was NOT achieved → the work is BLOCKED, not done. Record the gap (step 4), and either re-dispatch with a corrected charter or escalate to the operator.
3. Did the findings reach disk? AND is it WIRED? (the two recurring traps)
3a — Findings-to-disk. Walk what the dispatch discovered. Anything load-bearing that exists only in the dispatch's returned text — an investigation result, a dead-end, a decision with rejected alternatives, an anti-assumption — must be persisted to its owning surface, or it is lost:
- a research finding → the
research/<R-id>/tier file - a decision with alternatives → an ADR (
decisions/) or a checkpoint point 5 - a gotcha / anti-assumption → a
memories/claim ornow/status.md - a self-audit finding → the self-audit corpus (
dogfood/, if you run one)
If a finding is only in the conversation, FILE IT NOW (or note it as a debrief gap). (Findings only in conversation are DEAD at compaction.)
3b — IMPL→WIRED (production-reachability, not test-pass). This is the gate that kills the #1 failure. "It compiles and tests pass" does NOT mean it is wired to a production path. Verify reachability from the charter's wiring-proof target using the prover menu — first available wins, grep-the-callers is the honest floor:
the TypeScript language serverreachability / call-path query, if the project ships a code-intel tool. (Rebuild its index first if the tree changed.)- Else an LSP find-references on the new symbol — does anything on a production path call it?
- Else an LSP call-hierarchy from the production entrypoint down to the new code.
- Else a language call-graph tool (whatever node-ts offers).
- Else grep-the-callers as an explicit floor — record the query + the result, not just "looks wired".
- Else a manual-trace note in the row — name the entrypoint and the hop chain by hand.
Does a real path trace from a production entrypoint (a main(), a served command, a public API
surface) to the new code? If the query comes back EMPTY, the code is orphaned — IMPL, not WIRED.
Dead code the linter flags is this failure surfacing, not noise.
- Cross-layer lockstep surfaces (only if you run the revisit-ledger module) → confirm the
twin:anchors were set on the surfaces that must move in lockstep (a schema ↔ its parser ↔ its consumers; an interface ↔ its implementations) and the ledger row exists.
Classify the work-unit's state: IMPL (built, not yet reachable) / WIRED (traced from a
production entrypoint) / DEFER (intentionally not wired yet, with a written reason). WIRED
requires the traced path — an assertion is not enough.
4. Update the traceability ledger (the IMPL→WIRED record)
Write/update the work-unit's row in the traceability/ ledger (one row per work-unit,
UPDATE-IN-PLACE). Columns: WU-NNNN · FR-id · deliverable · IMPL/WIRED/DEFER · the
reachability evidence (the prover-menu query output ref, or the twin: anchor + ledger row) · date.
If traceability/index.md does not yet list the ledger, add it in the same change (index
completeness). (If you are on the Standard profile without traceability/, record the state in the
WU's now/work-plan.md row and in checkpoint point 8.)
A work-unit cannot be marked done in now/work-plan.md while its row is IMPL (not WIRED) without
an explicit DEFER rationale.
5. Journal the debrief
.agent-docs/log.md (newest at top):
## [YYYY-MM-DD] dispatch | debrief FR-NNNN (WU-NNNN) — <verdict: achieved | gaps | blocked>
Intent: <one line>. Reachability: IMPL|WIRED|DEFER (<evidence ref>). Findings persisted: <where>. <gap, if any>.Use the dispatch op for the debrief line; add memory / incident / dogfood op lines if the
debrief surfaced one of those.
6. Report
- "/debrief FR-NNNN (WU-NNNN): <achieved | gaps found | BLOCKED>"
- Reachability verdict:
IMPL/WIRED/DEFER+ the evidence ref. - Findings persisted to:
<paths>(or "all findings already on disk"). - If gaps: the specific remediation (re-dispatch with corrected charter / file the orphaned finding /
escalate). If it is a recurring failure shape: "candidate lesson — surface at
/distill-lessons."
DO NOT auto-commit.
Anti-patterns
- Do NOT accept "tests pass" as WIRED — trace production-reachability (the #1 recurring trap).
- Do NOT let the executing agent debrief its own dispatch — separation of duties (CONVENTIONS.md §7.2); run it from the orchestrator or a clean-context verifier.
- Do NOT leave a dispatch's findings in its returned text — persist or lose them.
- Do NOT mark a WU done on an
IMPL-only row without an explicitDEFERrationale. - Do NOT skip the traceability update — the ledger is the durable record that the loop was closed.
- Do NOT include secrets, credentials, or regulated / user data — reference paths.
Relationship to the other lifecycle skills
| Skill | Verifies | Against |
/debrief | a dispatch outcome | the charter's intent + on-disk artifacts + production-reachability |
/research-pipeline adversarial gate | a research claim | PRIMARY current-year sources |
/sitrep | (captures, does not verify) | freezes the forensic state incl. each touched file's wiring status |
/handoff | (consolidates) | folds open debrief gaps into the next-actions queue + the detour map |
/debrief is the per-dispatch gate; /sitrep and /handoff consolidate its outcomes into the durable
record.
Design rationale
IMPL→WIRED + production-reachability and the dispatch/verifier discipline: CONVENTIONS-full-addendum.md
§C (IMPL→WIRED) and §D (dispatch / wave / verifier). Separation of duties (the executor never
self-audits): CONVENTIONS.md §7.2. The recurring built-but-not-wired trap — the single most
expensive failure this gate exists to catch — and findings-to-disk are the reasons the loop is closed
by the orchestrator, not the builder.
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

