| name | eng-retro |
| description | Retrospective system for agent sessions — structured collection of mistakes, tooling gaps, and positives, plus cross-session pattern analysis. Use when running a retro, reviewing session quality, collecting observations about agent behavior, analyzing patterns across multiple retros, or producing a retro findings doc. |
Retrospective System
Part of the .eng/ documentation system. Retro files live in .eng/retros/ alongside objectives and findings. For file format conventions, frontmatter schemas, and naming patterns, see the eng-docs skill — it is the authority on .eng/ structure.
Two workflows: collection (per-session) and analysis (cross-session). Never mix them in the same document.
Collection Workflow
Collect structured observations from a single chat session. Output: one retro file in .eng/retros/.
- Review the full conversation history.
- Skim 2–3 recent retros in
.eng/retros/ if they exist — note recurring patterns but don't analyze.
- Collect observations honestly. Empty sections are fine if the session was clean.
- Write the retro file using the template in references/retro-template.md.
- Print a 2–3 line summary: session type, observation count, severity max.
Severity Rubric
| Level | Criteria |
|---|
| minor | User corrected with a short remark. No wasted work beyond the immediate undo. |
| moderate | Required user to explain what was wrong. Some wasted effort or a digression. |
| major | User had to intervene multiple times, cancel actions, or lost meaningful work/visibility. |
If unsure, pick the lower severity. Inflation makes retros useless for trend detection.
Category Taxonomy
See references/categories.md for definitions and real examples of each mistake and gap category.
Rules
- Be concrete — quote or paraphrase actual moments. Vague observations have no aggregation value.
- Don't editorialize or suggest fixes. State what happened and why it matters, then move on.
- Don't manufacture findings for clean sessions.
- One retro per session. Don't append to existing retro files.
Analysis Workflow
Analyze patterns across multiple retro files. Output: a findings doc in .eng/findings/.
- Read all retro files in scope (usually a date range or project).
- List every individual mistake and gap with session attribution.
- Group into patterns by root cause, not surface symptom.
- Compare against prior findings docs — note what's improving, stable, or worsening.
- Categorize by actionability: hook-solvable, instruction update, new convention, new skill, project-specific, behavioral-only.
- Write to
.eng/findings/finding-retro-patterns-YYYY-MM-DD.md.
Analysis Rules
- Every claim traces to specific retro entries. No unsourced generalizations.
- Patterns need 3+ instances across 2+ sessions to be worth naming.
- Tag frequency and severity distribution are required — they're the trend signal.
- Include "What's Working Well" — reinforcement matters as much as correction.