self-forge

self-forge

Turn how you actually work into reusable Claude Code artifacts.

self-forge mines your local Claude Code usage — prompt history, session transcripts, settings, and the skills/commands/agents/hooks you already have — finds the things you keep doing by hand, and forges each into the right shape: a slash command, a skill, a subagent, a hook, a workflow, or a whole publishable plugin. It proposes from the evidence of how you actually work — not from a blank page, and not from a codebase scan. All analysis stays on your machine.

Not the same as claude-automation-recommender. That official skill reads your codebase's tech stack and recommends. self-forge reads your own behaviour — what you repeat and correct — and ranks, scaffolds, and publishes. Different axis, complementary tools.

It is the meta-tool of this marketplace: where a specialist already exists, self-forge routes to it — CLI-wrapping to mcp-forge, corrections to behavior-hooks, durable knowledge to matts-second-brain, permission noise to fewer-permission-prompts, codebase-based suggestions to claude-automation-recommender.

Why

Every repeated prompt, every "no, always do it this way", every five-step deploy you re-explain is a latent artifact. You rarely stop to capture them. self-forge does the noticing for you: it quantifies the repetition, ranks what is worth automating, and builds it in the exact format Claude Code expects.

Install

/plugin marketplace add wrxck/claude-plugins
/plugin install self-forge@wrxck-claude-plugins

Requirements

• python3 (standard library only — no pip installs).
• A populated ~/.claude/history.jsonl and some session history. self-forge works best after you have used Claude Code for a while; it reads roughly your last 30 days of usage by default. On a fresh profile it will honestly tell you there is not enough signal yet, rather than invent work.

Use

Command	What it does
/self-forge	Mine your usage, rank candidates, present a proposal report (no files written)
/self-forge:scaffold <n>	Build one approved proposal into a real artifact
/self-forge:publish <name>	Bundle artifacts into a plugin and publish to the marketplace + docs

Or just ask: "what should I turn into a skill?", "analyze my Claude usage", "make a hook out of the thing I keep correcting" — the skill loads on those triggers.

What you'll see

/self-forge returns a ranked, evidence-anchored report — and stops there until you approve a build:

## self-forge — 3 proposals from 30d of usage

| # | Forge   | Name                       | Evidence     | Priority | Confidence |
|---|---------|----------------------------|--------------|----------|------------|
| 1 | hook    | block-compose-down-volume  | 7x corrected | High     | 0.9        |
| 2 | command | /review-security           | 12x typed    | High     | 0.85       |
| 3 | skill   | deploy-nutrition-saas      | 5x walked    | Medium   | 0.7        |

Scaffold any with /self-forge:scaffold <number>.

If your history is thin, it says so plainly and tells you when to come back — it never pads the list with invented suggestions. That honesty is the point: every proposal is backed by a real count.

How it works

MINE -> CLUSTER -> CLASSIFY -> SCORE -> PROPOSE -> (approve) -> SCAFFOLD -> (opt) PUBLISH

1. Mine — a dependency-free Python miner (scripts/mine.py, stdlib only) reads ~/.claude/history.jsonl, samples project transcripts, and inventories what already exists. It emits aggregate counts and short snippets of your own prompts — never assistant output, file contents, or anything network-bound.
2. Classify — each repeated pattern maps to the smallest artifact that captures it, via an explicit signal→artifact taxonomy with tie-breakers and anti-patterns.
3. Score — priority = (frequency x friction x automatability) / build_cost, gated by confidence, with a hard floor so weak ideas are dropped, not padded.
4. Scaffold — on approval, forge the files in exact format from bundled templates, then validate that they load.
5. Publish — optionally bundle into a plugin and ship it to the marketplace and docs site.

What is in the box

• skills/self-forge/ — the core skill, with references (signal-taxonomy, mining-playbook, artifact-formats, scoring-model, publishing), templates for every artifact type, and the miner script.
• agents/ — usage-miner, artifact-architect, artifact-smith for fanned-out runs.
• commands/ — /self-forge, /self-forge:scaffold, /self-forge:publish.

Privacy

All analysis is local. The miner reads files under ~/.claude and prints aggregates plus short snippets of your own prompts to justify proposals. Nothing is uploaded.