| name | show-me-your-work |
| description | Keep an auditable decision trail for long-running, autonomous, delegated, or multi-phase work. Use whenever a human will review the result after stepping away and needs decisions, evidence, pivots, and verification without rereading the full transcript. |
Show me your work
Maintain one append-only TSV decision log so a reviewer can reconstruct what was decided, why, and on what evidence. Log decisions and checkpoints, not keystrokes.
Start the trail
Copy references/decision-log-template.tsv to decisions.tsv, or use .audit/<task-slug>.tsv when several efforts share a worktree. Keep it local by default. Commit it only when a reviewer needs the trail to trust a large migration, port, or unattended run.
Append safely with:
scripts/log.sh <logfile> <phase> <decision> <why> <evidence> <result>
The helper timestamps the row, preserves single-line cells, and neutralizes spreadsheet formulas. Resolve scripts/log.sh relative to this installed skill.
| Column | Contract |
|---|
ts | UTC ISO-8601 timestamp |
phase | Current phase or workstream |
decision | One concrete choice or checkpoint |
why | Plain-language reason |
evidence | A resolvable pointer: commit, file:line, test output, trace, screenshot, or artifact path |
result | Observable state such as tests green, reverted, INCONCLUSIVE, or open |
Write one row for a fork chosen, a verified unit completed, a pivot/revert, a blocker, or a gate changing state. For iterative runs, write one row per meaningful iteration. A wrong call gets a later superseding row; never rewrite history.
Harness availability
| Need | Native path | Fallback |
|---|
| Locate this run's transcript | Use the active transcript path supplied by Cursor, Claude Code, Codex, or T3 | Match workspace metadata and opening prompt in that harness's scoped history; otherwise use the visible conversation |
| Fresh-eye audit | Use an authorized subagent; request a different model family only when the harness exposes valid model selection | Use an independent inherited-model reviewer; if delegation is unavailable, perform the audit yourself and disclose that limitation |
| Evidence | Use repository/runtime/browser/connector tools available in the harness | Mark the row INCONCLUSIVE; never manufacture a pointer |
Only delegate when the current user and system policy permit it.
End-of-run audit
- Walk the log against the active transcript or visible conversation.
- Confirm every row maps to a real action and every evidence pointer resolves.
- Add omitted forks, abandoned approaches, and verification failures that shaped the result.
- Remove padding; keep only rows a reviewer would audit.
- Ask the fresh-eye reviewer to flag weak evidence, skipped verification, risky choices, and gaps. The reviewer inspects; it does not redo or edit the work.
End the handoff with an Attention section naming the reviewer configuration and its flags. No flags is valid. If model diversity or delegation was unavailable, say so plainly.
