Skip to content

About

A small Claude Code plugin: plain explanations, confidence % on options, verify before done, a cross-check by a different model, and a read-only audit.

Topics

Resources

Code of conduct

Contributing

Security policy

Stars

12 stars

Watchers

0 watching

Forks

Repository files navigation

Blacksmith

A small plugin for Claude Code. Claude Code's defaults already do most of the work, so Blacksmith adds only what they leave out: a few always-on rules, two commands you run yourself, and four agents.

Day to day there is nothing to learn. You work in Claude Code as usual, and the rules apply on their own.

What you get

Rules in every session

One short file, plugin/rules.md. A hook loads it when a session starts, on resume, after /clear and after a compact.

Rule What you notice
Plain explanations Claude leads with the result, in short sentences and in your language. Jargon is explained and technical details come last.
Confidence on options When you must decide, you pick from options or reply with a number. Every option shows a %, and the recommended one says what its figure rests on.
Verify before done Claude does not say "done", "fixed" or "passing" until it has run the command that shows it.
The right agent for the job Planning goes to a stronger model, routine coding and review to a cheaper one.
Compact early Near 60% of the context window, Claude asks you to compact and says what to keep. One setting makes it automatic.

Two commands you run yourself

Claude cannot start these: /audit is costly and /cross-check sends code to other vendors. After a non-trivial change Claude offers /cross-check, and the choice is yours.

  • /cross-check [base] [criteria] asks a different model to review your change: codex, read-only, plus DeepSeek if you opt in. For example, /cross-check main "rejects empty input". A finding counts only once Claude reproduces it. Without codex it reports "skipped" and nothing breaks (setup).
  • /audit [axis ...] [path] reads the project for performance, code-quality and security problems, one agent per axis in parallel. For example, /audit or /audit security src/. It changes nothing, and your CLAUDE.md counts: a design it documents as deliberate is not a finding (details).

Four agents

planner splits bigger work into parts. coder builds one part, test-first, in its own git worktree. reviewer reads a diff with fresh eyes. auditor covers one axis of /audit. Each has its own model and effort (routing, and how to change it).

Install

Inside Claude Code:

/plugin marketplace add juzser/blacksmith
/plugin install blacksmith@blacksmith

Then start a new session. claude plugin details blacksmith should list 2 skills and 4 agents.

From a shell: claude plugin marketplace add juzser/blacksmith, then claude plugin install blacksmith@blacksmith. To try a checkout without installing: claude --plugin-dir ./plugin. An agent can do all of this from INSTALL.md: ask Claude to "install Blacksmith".

What leaves your machine

Only /cross-check sends your code to another vendor, and only when you run it. By default that is codex, which can read any file in the repo. DeepSeek gets the diff, the changed tests and your criteria, and only if you opt in. Exactly what goes where.

More

Phrasing ideas drawn from ayghri/i-have-adhd (MIT). License: MIT.

About

A small Claude Code plugin: plain explanations, confidence % on options, verify before done, a cross-check by a different model, and a read-only audit.

Topics

Resources

Code of conduct

Contributing

Security policy

Stars

12 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages