← ClaudeAtlas

ref-sp-agents-retrolisted

Reference guidance for using `.agents/retro/` as a gitignored local workspace where an agent records a short, descriptive retrospective at the end of a substantial task: what went well, what went wrong and how each problem was solved or worked around, new information learned about the repo or tools, and improvement hypotheses a future agent can weigh when approaching similar work. Use when: finishing a substantial task and capturing a retro, recording how a problem was solved or worked around, noting new facts discovered about the repo or tooling, reading past retros before starting similar work to calibrate an approach, writing or naming a retro entry under `.agents/retro/`, triaging a captured retro from `10-new/` to `90-closed/`, setting its `status`, `created`, `updated`, or prose `outcome` frontmatter, or deciding whether a recurring retro observation should be promoted into a skill or instruction.
swiftpostlabs/agentic-tools · ★ 0 · AI & Automation · score 71
Install: claude install-skill swiftpostlabs/agentic-tools
# Agents Retro ## Purpose Define how this repo uses `.agents/retro/` as a gitignored local workspace for **retrospectives**: short, descriptive notes an agent writes at the end of a substantial task recording what went well, what went wrong, and hypotheses for doing it better next time. A retro is **descriptive, not prescriptive**. It records what actually happened on one task and an honest read of why, so that a future agent reading accumulated retros can *calibrate its own approach* — not so it can follow a rule. A retro says "last time I did X this way and it cost two extra round-trips; a lighter path might be Y," not "always do Y." The future agent still exercises judgment; the retro is evidence, not instruction. A retro carries two kinds of payload beyond sentiment, and both are what make it worth reading later: - **How problems were handled.** When something went wrong, record whether and how it was solved or worked around — the actual fix, the workaround, or that it is still open. A future agent hitting the same wall benefits far more from "resolved by doing Z" than from "this was hard." - **New information learned.** Facts discovered during the task — a repo constraint, a tool's real behavior, a non-obvious dependency, an environment quirk — that a future agent would otherwise have to rediscover. If such a fact is durable and general, it is also a promotion candidate. This is a distinct layer from the other `.agents/` workspaces and from durable guidance: - **`