mcpField
R
reflect

reflect

github.com/hachejAgent workflows161 installsUpdated 3d ago
Repository

About reflect

reflect is an agent skill for turning useful session learnings into durable changes in a repository. It helps route discoveries to code rules, tests, skills, documentation, or upstream projects instead of keeping them in a separate memory vault.

Key features

  1. 01Review session mistakes and discoveries
  2. 02Skip lessons already documented
  3. 03Route learnings to durable locations
  4. 04Encode repeated mistakes as rules
  5. 05Summarize resulting changes

Use cases

  1. 01Capture a recurring agent mistake as a check
  2. 02Document a codebase or tool quirk
  3. 03Improve a skill procedure
  4. 04Share a lesson with a template or upstream project

Install

Copy the command and run it in a terminal.

$npx skills add https://github.com/hachej/boring-stack --skill reflect

After installing, type/reflectin Claude Code to confirm it loaded.

Details

From the project README

Reflect

Adapted from Poteto's brainmaxxing reflect (MIT, see THIRD_PARTY.md). This repository has no separate memory vault: what the next agent must know lives in the repository itself, where it is reviewed like code.

Process

  1. Scan the session for:
    • mistakes made and corrections received;
    • the product owner's preferences (how they want the app to behave or read);
    • codebase knowledge gained (gotchas, why something is the way it is);
    • library or tool quirks (vendor/boring/, Flue, the model provider);
    • friction in a skill (boring-app-template, build-app) or in the app's checks;
    • repeated manual steps.
  2. Skip anything trivial or already written down (grep AGENTS.md, docs/, the skills).
  3. Route each learning (below), smallest durable form first.
  4. Put the changes in the current branch, or a new one, and say what you changed.

Routing, in this order

  1. Structure (principle-encode-lessons-in-structure): can it be a lint rule (.oxlintrc.json), an architecture rule (scripts/verify/ARCHITECTURE.json), a unit test, a verify check step, a validator in an agent's index.mjs, a law with its evidence (docs/INVARIANTS.md + scripts/verify/VERIFY.json) or a model (formal/)? Then encode it there and write no prose. A mistake an agent repeats becomes a rule in scripts/verify/ARCHITECTURE.json, with the law or decision it enforces, a fix hint, and a fixture in test/unit/architecture.test.mjs proving it fires; better still, make the mistake impossible (a type, a single helper, a constraint) so no rule is needed.
  2. A skill (a pull request on hachej/boring-stack, skills/<skill>/SKILL.md): the learning is about how a skill's procedure works: fix the procedure.
  3. The feature map (docs/verify/<area>.md): a gotcha about driving a feature.
  4. docs/DECISIONS.md: a decision and its reason ("we do X because Y; not Z because W").
  5. AGENTS.md: a rule every agent needs on every task. Keep it short: move detail to a skill.
  6. Upstream: a gap in @boring/agent, @boring/files or @boring/chat goes to hachej/boring-ui-v3 as an issue or a pull request, never as an edit of vendor/.
  7. The template (hachej/boring-app): a lesson every new app needs. Say so, so it lands in the template; existing apps take it by a normal pull request.

Summary

## Reflect summary
- Structural: [rules, tests, checks added]
- Skills: [skill files changed, one line each]
- Docs: [DECISIONS / AGENTS / feature map entries]
- Upstream or template: [issues or PRs to open]

Comments

Markdown supported