AGENT SKILL

reflect

Review a long session through independent lenses and propose edits to existing skills, packs or checks, applied only after they pass skill-tdd. Use at the end of a long or rocky session, or when the user asks to reflect on what the skills got wrong.

License 许可
MIT
最近更新
2026-11-04
Stars
★ 0
Forks
0

01 前言 Front Matter

name
reflect
metadata
{ "layer": "L3", "source": "cursor/plugins@1e5c90a:pstack/skills/reflect/SKILL.md", "license": "MIT" }
description
Review a long session through independent lenses and propose edits to existing skills, packs or checks, applied only after they pass skill-tdd. Use at the end of a long or rocky session, or when the user asks to reflect on what the skills got wrong.

02 资源文件 1

03 SKILL.md 正文

Reflect

Mine this session for durable learnings about how the work was done, and turn each one into a concrete change to an existing skill, pack or check. Skip trivial sessions, and sessions where the skills were followed and worked. One-offs are not learnings.

Nothing is applied to a skill in this session. Skill edits become proposals that skill-tdd applies in a later session, so a skill never edits itself in the session that proposed the change.

1. Find the transcript

The transcript is your agent's log of this session, when it keeps one on disk (Claude Code: the newest *.jsonl in <config dir>/projects/<repo path with every non-alphanumeric character replaced by ->/, where <config dir> is $CLAUDE_CONFIG_DIR or ~/.claude). Confirm it by matching its first user message to this session's opening request. Stay in this repo's project directory; never read other projects' transcripts.

No match: write a digest instead (the requests, corrections, skills used, check refusals, outcomes).

Count corrections from the decisions log first: ven-decisions list --step correction shows what the user objected to this session and earlier. The transcript is the fallback and the source for everything else.

2. Run three blind lenses

delegate to three reviewers at once, each with the brief in references/lenses.md (shared part plus its lens) and the transcript path or digest:

Lens Looks for model
judgment The durable principle behind an incident orchestrator_model
tooling Commands, paths and context the agent rediscovered or the user had to hand over coder_model
divergent What the other two miss: second-order effects, skipped proof, skills that never fired orchestrator_model

Reviewers never see each other's output. On the lean budget tier, run the judgment lens only.

3. Synthesize

Judge every finding yourself against these tests, in order:

  1. Structural. If a check could enforce it, route it to ratchet instead of skill text (apply encode-lessons-in-structure, which also sets the promotion threshold).
  2. Durable. Still true after paths, versions and code shapes change.
  3. Decision-changing. A future agent acts differently, not just reads more.
  4. Used. The target skill ran in this session, or should have fired and didn't (then the change is to its description).
  5. Not already covered. Read the target skill. Already said clearly: the failure was execution, reject it. Said weakly or buried: propose moving or sharpening that line, not adding another.
  6. Convergent. A finding from two lenses ranks highest; a singleton must clear every test on its own.
  7. Agnostic. Tech, vendor or domain specifics go to a pack skill, never to core.

A finding about the repo rather than about how to work goes to compound.

4. Present and propose

Show the result and wait for the user to pick rows:

## Accepted
| Problem | Proposed change | Target (skill and section, or description) |
## Checks
| Rule | Rung | Original mistake it must catch |
## Rejected
- <finding> — <structural | durability | decision-changing | not used | already covered | one-off>

For each picked Accepted row, write a proposal file in the format skill-tdd defines. For each picked Checks row, run ratchet now (a check is not a skill edit) or queue it as a proposal of kind check.

5. Summarize

  • Proposals written: path and one line each.
  • Checks: added (with path) or queued.
  • Solution entries written.
  • Dropped: one line each, with the reason.
ven-decisions add --step reflect --decision "<n> proposals, <n> checks, <n> dropped" --evidence "<proposal paths>"

04 同仓库更多 Skills

Ventorius
Ventorius/ven-stack

yeet

Take every local change green through the repo's own checks, slice it into conventional commits, push, and with --pr open a pull request on the code host. Use only when the user explicitly asks to yeet, or to commit and push everything; never on your own initiative.

★ 0
Ventorius
Ventorius/ven-stack

write-docs

Write or update documentation from the code: gather facts, pick the shape for the reader, write it plain, and prove every command and path in it. Use when asked to document a repo, module or feature, or to bring stale docs up to date.

★ 0
Ventorius
Ventorius/ven-stack

why

Find why code is shaped the way it is (design rationale, trade-offs, regressions, odd thresholds) from history, tickets, docs and chat, citing every claim and stating its confidence. Use for 'why does X work this way', 'why did we pick Y', or before changing code whose purpose is unclear.

★ 0
Ventorius
Ventorius/ven-stack

verify

Prove a change works with the check that matches its type, do the read-back, and log the pass against the exact tree. Use when a change is ready to call done, before review, commit or stopping, and whenever ven-decisions end refuses an unverified tree.

★ 0
Ventorius
Ventorius/ven-stack

unslop

Cut AI writing tells (filler, model vocabulary, metaphor nouns, hedging, formatting tics) from text while keeping every fact. Use when editing or checking any prose before it ships: docs, PR bodies, reports, skill text.

★ 0
Ventorius
Ventorius/ven-stack

type-discipline

Apply when designing types or data shapes, reviewing a signature, or handling outside data, in any language. Make wrong states fail early: in the compiler for typed languages, in validated runtime contracts for untyped ones.

★ 0