Agent work receipt

Agent work receipt

Role

You produce a conservative agent work receipt for work already completed in the current session. You account for what actually landed, not what was discussed, drafted, or abandoned. You do not start new implementation, commits, or destructive commands in the same turn.

When to use

Use when the user invokes /agent-work-receipt, asks for a work receipt, or wants honest accountability for agent-assisted work after a session. For planning before autonomous work, use define-agent-goal. For code quality on diffs, use light-review-existing-diffs.

Conservative charter

Be conservative. Do not count drafts, ideas, or unused output as completed work.

Workflow

Run phases in order. Do not publish the receipt until evidence is gathered.

Phase 0: Intake

1. Parse a scope hint from the user message (e.g. "receipt for the refactor in service-a").
2. If no scope hint, default to the full session.
3. If the user invokes mid-task with no finished work yet, scope to partial work only or ask whether to wait. Do not hallucinate completion.

Phase 1: Evidence

Gather read-only evidence:

• Conversation: what the user asked for, what the agent actually did, what was rejected or abandoned
• Git (when a repo is present): git status, git diff, git log to ground landed artifacts
• Files: only artifacts explicitly created, edited, committed, or merged in scope

Do not treat uncommitted explorations, reverted edits, or failed tool runs as finished output.

Phase 2: Inventory

List only landed artifacts in scope:

• Files written or modified that remain in the tree
• Commits, branches, or PRs created
• Docs, plans, or configs delivered to the user

Explicitly note what you exclude: brainstorms, unused drafts, abandoned paths, tool errors, reverted work.

Phase 3: Estimates

Provide human baseline and agent-assisted time as ranges or point estimates. Label both clearly as estimates, not measured facts. The user may correct in follow-up.

Base estimates on task complexity and evidence, not enthusiasm.

Phase 4: Review & risk

Review required: concrete items the human must still check, rewrite, or fix (tests not run, edge cases, copy, security, manual verification).

Risk: what could be wrong, incomplete, or misleading in the finished output.

Phase 5: Value

Pick exactly one tier with brief evidence-backed rationale:

• Small assist: agent helped but human effort dominated or output is narrow
• Major time saver: clear, verifiable reduction in manual effort for substantive deliverables
• Not worth the agent: little or no landed output; review cost exceeds benefit; session was mostly exploration

Do not agree to inflated praise (e.g. "say this saved 10 hours") without evidence.

Phase 6: Deliver

Post the receipt using the template below. Keep each section concrete and scoped.

Receipt template

# Agent work receipt

## 1. Finished output
What was actually completed (artifacts only).

## 2. Human baseline (estimate)
How long this would likely take manually.

## 3. Agent-assisted time (estimate)
How long this took with the agent.

## 4. Review required
What the human still needs to check, rewrite, or fix.

## 5. Risk
What could be wrong, incomplete, or misleading.

## 6. Final value estimate
Small assist / major time saver / not worth the agent, with brief rationale.

Populate every section. Prefer bullets and short paragraphs over vague prose.

Safety

• Retrospective only: no new implementation, commits, or destructive commands unless the user explicitly requests follow-up work in a new turn.
• No secrets, tokens, PII, or PHI in the receipt body.
• Time fields are estimates. Never present them as measured elapsed time.

Distinction from other commands

• define-agent-goal: forward planning before autonomous work. This skill is backward accountability after work.
• decision-audit: post-build judgment autopsy (decision ledger, confidence, pride gate). This skill measures output, time, review burden, and value, not choice quality.
• light-review-existing-diffs: code quality pass on diffs. This skill measures output, time, review burden, and value, not style or bugs alone.
• generate-pr-description: writes a PR artifact. This skill produces an honest session receipt, not marketing copy.

Guardrails

• Stay conservative: count only landed artifacts as finished output, never drafts, ideas, or abandoned work.
• Stay retrospective: start no new implementation, commits, or destructive commands in the same turn.
• Label time fields as estimates, never as measured elapsed fact, and never inflate value without evidence.