Installs into .claude/skills of the current project.
Are you the author of Retro?
Add the live security badge to your README. It updates with every re-scan.
[](https://www.skillsdirectory.com/skills/snappedly-retro)
---
name: retro
description: "Review a coding session and recommend evidence-backed improvements to the agent's environment."
disable-model-invocation: true
license: MIT
---
# Retro
A retrospective produces prioritized recommendations. Implement them only when the user's request includes that work; that request is the approval. Otherwise, report them as recommendations.
## Inspect the session
Use the current session unless the user names another. Read the relevant transcript, tool results, and existing decision logs, and point to specific events. State what evidence is missing or incomplete, and draw conclusions only from the parts of the session you could inspect.
For each observed difficulty, inspect the relevant repository configuration before proposing a remedy. Existing check commands, CI, hooks, standards, and `docs/agents/workflow.md` may already address it but be disconnected, unclear, or broken.
For code changes, also check whether required feedback-loop commands run through CI or local hooks. Report missing or broken enforcement where those checks should protect the changed work. A documented command alone does not establish that it runs.
Done when every observed difficulty has an evidence pointer and the repository configuration bearing on it has been inspected.
## Choose improvements
Consider these remedies where the evidence supports them:
- **Navigation:** add a pointer where the agent looked, when needed information was difficult to find.
- **Checks:** repair existing enforcement first. For a mechanical failure with no adequate check, propose the cheapest reliable check through the repository's existing tools.
- **Standards:** clarify judgment calls in existing standards or workflow docs. Keep applicable standards available to both implementation and review, including work handled by one agent.
- **Instructions:** clarify or remove steering that caused confusion or added no useful direction. Load and apply `writing-for-agents` when editing agent documents; if it is not installed, say so and hold them to its core standard: each meaning in one place, positive instructions, and a clear completion criterion for every step.
- **Tooling and access:** reduce demonstrated tool cost or expose needed information through a scoped capability, such as readable development logs. A recommendation does not authorize credential changes or broader access.
Weigh each remedy's maintenance cost, runtime overhead, and false positives; a noisy or redundant check may need narrowing or removal. Extend an existing standards file when one fits.
## Return findings
Order candidates by observed consequence and likely recurrence. For each, report the event and evidence pointer, whether its cause is observed or inferred, the proposed change and destination, and how to verify improvement. No quota applies: report no actionable findings when the evidence supports none.
When implementation is authorized, apply the selected remedies within scope and run the affected checks; carry unresolved candidates forward explicitly.