craftsman makes engineering discipline the default in Claude Code: the smallest correct change, no fix without a root cause, and an honest report of what was actually verified. It bundles thirteen agents, seven skills, five slash commands, and a cross-platform hook system — installable in two commands.
claude plugin marketplace add bufferBrew/craftsman
claude plugin install craftsman@craftsman
Or from inside a Claude Code session:
/plugin marketplace add bufferBrew/craftsman
/plugin install craftsman@craftsman
Restart Claude Code (or start a new session) after installing, then run /craftsman:init in any
project to get started.
| Problem | Before | With craftsman |
|---|---|---|
| Choosing the right agent pipeline | Trial-and-error chaining of agents | @orchestrator classifies the task, picks the minimal pipeline, and gates each stage on fresh evidence |
| Diffs grow beyond the request | "Fixed it" — plus an uninvited refactor of three other files | Minimal-diff coding — the smallest change that solves the problem; anything extra needs your OK first |
| The same bug returns under a new name | Symptom patched, root cause untouched | Root-cause debugging — no fix without an established root cause; a graphify knowledge graph catches recurring bugs before they're filed as new ones |
| Finding relevant code takes forever | Raw grep through files | Query the graphify graph first — hooks nudge every grep/find toward it in graphed projects |
| Environment quirks rediscovered every session | Same OS/shell/tool failure re-derived by trial and error | Environment-quirk memory — discovered once, recorded, never re-derived |
| Scaffolding writes files you didn't ask for | Surprise CLAUDE.md and config files |
Ask-before-writing — /craftsman:init proposes exact file content and waits for confirmation |
| "Done!" but nothing was actually run | Claims without evidence | Honest completion — every nontrivial task ends with a Verified / Assumed / Not covered status block |
If "fix the off-by-one" comes back as a fix plus renamed variables and a reformatted file, that's
what the smallest-change-first skill and the coder agent's minimal-diff discipline are for:
anything beyond the literal request needs your OK first. See
use case #1.
A one-line ask shouldn't turn into a drive-by refactor of three unrelated files. craftsman's
coder agent runs a seven-step ladder before writing anything new (does it need to exist? already
in the codebase? stdlib? platform feature? existing dependency? one line?) and stops there unless
you say otherwise.
"Done!" followed by a crash on the first real run usually means the build was never actually
executed. The caveats-and-status skill ends every nontrivial task with a fixed
Verified / Assumed / Not covered block, and orchestrator pipeline gates require fresh build/test
output at each stage — not a claim that it passed. See
use case #4.
When a null-check bug fixed in OrderService resurfaces in InvoiceService two months later,
nobody usually connects the two. With a graphify graph in
the project, the debugger agent queries it before grepping and checks new bugs against
KNOWN_ISSUES.md for a resolved match — surfacing the original fix to reuse instead of
re-deriving it. See
use case #2.
- Full plugin reference — agents, skills, commands, hooks, installation options, troubleshooting.
- Use cases — worked scenarios showing when each piece earns its keep.
- Results — real, verifiable before/after evidence, honestly scoped to what exists so far.
- Contributing — how to add agents/skills, validate, and test hooks.
- Changelog
If craftsman saves you a debugging session, a ⭐ helps others find it.