Skill Loop

skill-loop

A Claude Code plugin that learns a codebase's conventions and applies them on later sessions. It generates a small set of coding skills from an existing repository, then keeps them current as you work, learning from corrections, new patterns, clean merges, and failures.

Learned skills are personal to you and scoped per project. Nothing is committed to or pushed into your repositories, and most sessions cost no additional tokens.

Installation

/plugin marketplace add joeljohn159/skill-loop
/plugin install skill-loop@skill-loop

Then run once per repository:

/skill-loop:bootstrap

Bootstrap reads the codebase and writes a small set of convention skills. After that the plugin works in the background. You interact with it only to promote recurring lessons, and occasionally to change models or view its log.

How it works

The plugin is organized as six layers. The expensive, high-capability work is rare and explicit. The recurring per-session work is cheap, and is skipped entirely when there is nothing to learn.

Layer	Trigger	Cost	Responsibility
Bootstrap	/skill-loop:bootstrap	Opus, once	Read the repo. Judgment rules become skills; mechanically enforceable rules become formatter and linter configuration. Each rule ships with a command that verifies it.
Enforce	PostToolUse hook	none	Run the project formatter on each edited file. Deterministic rules are fixed automatically and never become skills.
Capture	hooks	none	Record raw signals: corrections, new patterns, approvals, failures.
Reflect	Stop hook	gated	A deterministic pre-scan runs first; with no signal it exits at no cost. Otherwise one model pass extracts and de-duplicates candidate rules.
Promote	/skill-loop:promote	Sonnet, rare	Generalize recurring candidates, write them into your skills, and prune stale entries. Manual, never automatic.
CI feedback	/skill-loop:learn-from-ci	model_ci	Turn a failing build and the fix that resolved it into a candidate rule, processed locally.

flowchart LR
  bootstrap --> work[work normally]
  work -->|edit a file| enforce[enforce: format]
  work -->|signals| capture
  capture --> reflect{Stop: signal?}
  reflect -->|no| done[exit, no cost]
  reflect -->|yes| stage[stage candidate]
  stage -->|recurs| promote
  promote --> skill[(personal skill)]
  skill -.->|applies next time| work

Reflection is the only automatic per-session model call, and it is gated: a correction, a new dependency, a failed test, or a clean merge will trigger it, while an uneventful session triggers nothing.

Commands

Command	Purpose
/skill-loop:bootstrap	Read the repo and generate its convention skills. Run once per project; prompts for your model profile on first use.
/skill-loop:promote	Move recurring candidates into your skills. Manual, never automatic.
/skill-loop:learn	Reflect on the current session immediately, instead of waiting for it to end. Accepts an optional lesson to record.
/skill-loop:learn-from-ci	Paste a failing CI log and let the fix become a candidate rule. Works with any CI; runs locally.
/skill-loop:configure	Choose which model each stage uses.
/skill-loop:logs	Open a readable activity log in a separate terminal tab.

Models

Each stage's model is configurable through /skill-loop:configure (bootstrap also prompts on first run). Reflection is the only automatic call, so its model is the main cost lever.

Profile	bootstrap	promote	reflect	ci
Maximum	opus	opus	opus	opus
Balanced (default)	opus	sonnet	haiku	haiku
Economy	sonnet	haiku	haiku	haiku
Custom	chosen per stage

Storage and privacy

Everything skill-loop learns is personal, per-project, and kept in your home directory. Nothing is written into a repository.

• Skill: ~/.claude/skills/sl-<project>/ — one folder per project: a SKILL.md router plus one file per concern (error-handling.md, layering.md, …) that loads on demand. Scoped with a paths: rule so it activates only inside that project.
• Per-project state: ~/.skill-loop/projects/<project>/ (signals, candidates, snapshots, activity log).
• Global preferences: ~/.skill-loop/config (model choices, thresholds, scope).

Projects stay isolated, so one repository's rules never apply in another. To share a single set across all projects, set scope=global in the config. The only file bootstrap may add to a repository is a standard formatter config such as .prettierrc, which is shared team infrastructure rather than a skill.

Safety