nj-agents¶

A project-agnostic toolkit of Claude Code skills and agents covering the software development lifecycle. Install once; invoke with /name in any git repo, any stack, any language.

40 skills · 30 agents · 6 classes

The classes¶

Class	Skills	What it promises
Review	11	Advise only.
Authoring	10	Writes exactly one artifact into the repo, then proposes the commit — never runs git itself.
Workflow	4	Reads a diff and drafts a change artifact — a PR or a commit message — then proposes it.
PM-authoring	4	Writes a work item into whatever tracker is connected, then proposes the create.
Social	1	Produces paste-ready copy.
Testing	10	Writes test source into the repo and executes it — the only class that does both.

How it fits together¶

40 skills · 30 agents · 6 classes · 3 agents that can write files. Counted from the source on every build — the diagram deliberately carries no tallies.

What an agent is allowed to do¶

A skill orchestrates; an agent does one job and hands the result back. 27 of the 30 agents cannot write files at all — they return their output and the skill writes it. That is not an accident of configuration: each agent's body says so, and its declared tools match.

Only 3 hold a write tool, because each genuinely produces a file itself:

Every agent declares an explicit tools: allowlist. That is required rather than optional, and for portability rather than taste: Claude Code and Cursor read a missing key as inherit every tool, but Gemini CLI reads it as no tools, so the agent loads unable to act. An explicit list is the only spelling that means the same thing on every runner.

check.sh enforces both halves — the key must be present, and an agent whose body claims to be read-only may not declare a write tool. The two statements of the same contract cannot drift apart.

Always current, by construction¶

Every skill and agent here is written and maintained by hand — the procedures, the safety rails, the conventions each one follows. This site is the presentation of those files, rebuilt from them on every commit.

That distinction is what makes it trustworthy:

A page cannot disagree with the thing it documents. It is rendered from skills/*/SKILL.md and agents/*.md at build time, so a change to a skill is a change to its page.
The wiring is derived, not retyped. Which agents a skill spawns — and which skills spawn a given agent — is recovered from the definitions themselves.
A broken reference fails the build. A skill pointing at an agent that does not exist stops the site from publishing rather than shipping a dead link.

Built with the help of Claude, and verified the same way everything else here is: a validator that runs on every push, and a behavioural suite that asserts what these skills must never do.