| name | agent-notes |
| description | Write succinct agent notes to minimise context. Use when: reporting on work done, summarising completed tasks, reviewing another agent's output, flagging what was good, what needs fixing, what is incomplete. Triggers: report, summarise work, note findings, review agent output, what was good, what needs improving. |
| argument-hint | self-report | peer-review |
Agent Notes
Write minimal, scannable notes. Every byte counts — other agents will read this.
When to Use
- Self-report: After completing a task, log what you did and any issues
- Peer-review: After reading another agent's output, score and flag it
Rules
- One bullet per item — no prose, no padding
- Max 5 bullets per section
- Omit sections that are empty
- No markdown headings inside bullets
- File path references use short relative paths
Procedure
Self-Report
- Open (or create)
.notes/agent-report.md
- Append a block using self-report template
- Replace
{{DATE}} with today's date (ISO 8601: YYYY-MM-DD)
- Replace
{{AGENT_NAME}} with your agent name or filename (e.g. test-plan-writer)
- Fill only non-empty sections; remove empty ones
- Validate before saving: confirm
date and agent are set — if either is {{...}}, replace before writing
Peer-Review
- Read the target output (file or chat)
- Open (or create)
.notes/agent-report.md
- Append a block using peer-review template
- Replace
{{DATE}} with today's date (ISO 8601: YYYY-MM-DD)
- Replace
{{AGENT_NAME}} with your agent name or filename
- Replace
{{REVIEWED_AGENT_OR_FILE}} with the name or path of what you reviewed
- Fill only non-empty sections; remove empty ones
- Validate before saving: confirm
date, reviewer, and subject are set — if any is {{...}}, replace before writing
Decision Line (peer-review only)
Every peer-review block must end with a machine-scannable decision line in exactly this format:
DECISION: APPROVED | REQUIRES REVISION — <one-line reason>
Rules:
- Must be the last line of the block (after all bullets)
- Use
APPROVED or REQUIRES REVISION — no other values
- The reason after
— is required; never leave it blank
- Downstream agents scan for
DECISION: APPROVED or DECISION: REQUIRES REVISION to gate their work
Output Format
Appended block in .notes/agent-report.md. Each block is self-contained and timestamped.
Writing Output Documents
This applies to all primary output files — design plans, review reports, and test reports. Verbose documents cost every downstream agent that reads them.
Rules
- Lead with the verdict or result — never bury it
- Tables over prose: decisions, findings, and results belong in rows, not paragraphs
- One item per table cell — no paragraph explanations inside cells
- Bullets over sentences, sentences over paragraphs
- No preamble ("This document describes…") and no conclusion ("In summary…")
- Omit empty sections entirely — do not write "Not applicable" or "N/A" paragraphs
- No repetition: if something was stated in the summary, do not restate it in the body
Format by document type
Design plan (*-design.md):
- Summary: 2–3 sentences only — what it does, why, who benefits
- 9-layer matrix: one decision per row, max 3 sentences per cell; no paragraph blocks
- Technical spec: flat bullet list; no sub-prose
- Acceptance criteria: table rows only; pass condition is one testable line
Review report (*-review.md):
- Verdict: one sentence — cite the single most important finding
- Verification table: one row per claim, result in one word (CONFIRMED / UNVERIFIED / INCORRECT)
- 9-layer scores: one key finding per layer, one sentence max
- Strengths / weaknesses: one bullet per item, no nested prose
- Recommendations table: recommendation is an imperative sentence; rationale is one phrase
Test report (*-test-report.md):
- Results: one table row per criterion; evidence in one line
- Failures: fill structured fields only; one line per field
- Summary: counts + one-line action required