reflect · git:20261108.ac1df7f · 2026-11-08 · sha256 ed0084f8db5391d7
reflect git:20261108.ac1df7fA
Immutable. This exact content never changes and is served at /api/v1/blob/ed0084f8db5391d7.
No tool to install first
claude plugin marketplace add https://markdownregistry.com/mp/a/art_v236bmqtgxkm3eci/ed0084f8db5391d7925d046204a3669378c1da0aa5b2b8f5c4d301c06e10e7d1.json
Then /plugin install reflect@markdownregistry-v236bmqtgxkm3eci in Claude Code. It installs this exact version, the one graded A here, pinned to the commit it was graded at, and leaves it there.
---
name: reflect
description: "After a long task, a three-reviewer panel audits the session transcript — tooling, judgment, missed alternatives — and proposes concrete skill edits. Findings feed skillify. For banking marathon-session lessons."
aliases: []
metadata:
credits:
transcript-review-panel: pstack
author: Lauren Tan
url: "https://github.com/cursor/plugins/tree/main/pstack"
port: pstack-claude
port-author: Michael Denyer
port-url: "https://github.com/michael-denyer/pstack-claude"
---
# Reflect
Long sessions teach things nobody writes down. This skill audits the
session transcript with independent reviewers and converts the findings
into proposed edits to OMC's own skills — before the lessons evaporate.
## When to use
- a marathon session just ended (or is ending) and went well or badly in
ways worth banking
- the user asks "what did we learn", "bank this session", or says reflect
## When not to use
- routine sessions with nothing anomalous — reflection on nothing is noise
- the user wants a summary of what happened — `ketchup` does that
## Workflow
1. **Locate the transcript** for this session and hand each reviewer the
path, not a paste — reviewers read it themselves.
2. **Seat the panel**: three read-only reviewers in one message, in the
background, each with one lens:
- **tooling reviewer** — which commands, skills, and agents helped or
wasted time; what a script or skill should have existed
- **judgment reviewer** — where calls were wrong, late, or unverifiable;
which principle would have caught it
- **divergent reviewer** — the alternative approaches never tried; what
a fresh agent would have done differently
Each returns at most three findings, each with transcript citations.
**Wait for every reviewer to reach a terminal state and collect its
output before Step 3** — background dispatch returns before the
reviewers finish, and synthesizing on a partial panel proposes
misleading skill edits. A reviewer that times out is named as missing,
not silently dropped.
3. **Synthesize**: merge duplicates, drop unevidenced claims, rank by how
often the lesson would recur. A lesson that applies once is a note; one
that applies every session is a skill edit.
4. **Propose skill edits**: for each surviving recurring lesson, the
concrete edit — which skill file, which section, what wording — or a
proposal for a new skill via `skillify`. Human-ruling wording stays
verbatim (`why` skill epistemics).
5. **Report** the findings and the proposed edits; apply an edit only with
the user's go — reflection proposes, the user disposes.
## Rules
- Reviewers cite the transcript; a finding without a citation is dropped.
- The panel never edits skills itself — proposals go to the user and then
to `skillify` for execution.
- Never edit a skill inside the installed plugin cache. Before the user
approves an edit, show its destination — the repo's `skills/<name>/` for
a shipped skill, or the user's `.claude/skills/<name>/` copy — and after
approval edit that destination in place. Reflection never opens a PR or
pushes.
- Findings about the user's own choices are reported as observations, not
corrections.