This directory contains a portable Agent workflow kernel. It is meant to be copied into a project that wants Agent-assisted work to be planned, auditable, and able to produce reusable project experience.
The mainline has two layers. The per-target loop runs automatically on every target:
plan
-> implementation
-> validation
-> log
-> commit
The experience layer runs across targets and is human-triggered, not part of close-out:
human reviews plans, logs, or failures
-> human decides a reusable lesson is worth extracting
-> Agent drafts a candidate lesson note with evidence
-> human accepts, edits, or rejects
AGENTS.md: repository-level Agent working rulesplan/README.md: how target plans and logs workplan/target-plan.template.md: a starter plan formatplan/log.md: factual maintenance log starterdocs/experience/README.md: how to extract reusable lessonsdocs/experience/lesson.template.md: a starter experience note format
This kernel is not a project-management system, a note-taking app, or a complete development methodology. It gives Agents and humans a small shared protocol for bounded work:
- plan before editing
- declare file ownership
- validate changes in proportion to their behavior, risk, and affected surface
- record factual outcomes
- support human-triggered extraction of reusable lessons, with the Agent assisting rather than deciding
Project-specific repositories should add their own domain rules, validation commands, branch policy, release policy, and review expectations.
- Copy this directory's contents into a new or existing repository.
- Keep
AGENTS.mdat the repository root, or merge its generic rules into the project's existingAGENTS.md. - Keep
plan/README.md,plan/log.md, and the target-plan template. - Keep
docs/experience/README.mdand the lesson template if the project wants reusable experience notes. - Add project-specific validation commands, including any conditions that require a full suite, and branch policy.
A plan is intent before work.
A log is a factual record after work.
An experience note is a reusable judgment extracted from evidence.
Do not call raw logs "experience" until they explain a transferable pattern, failure mode, rule change, or decision principle.