meta-dev

One authority for every Claude-config artifact. meta-dev gives skills, agents, hooks, plugins, and project instruction files (CLAUDE.md/AGENTS.md) a single modular standard, a right-sized lifecycle for each, and enforcement that runs itself — so you stop re-litigating "what goes where, in what format, at what quality" on every project.

It is opinionated on purpose: loaded text informs, only hooks prevent; each rule lives in exactly one place; verification scales to the change. Everything else follows from those.

Install

claude plugin marketplace add farzanmrz/meta-dev
claude plugin install meta-dev@farzan-dev

Then start a session (or /reload-plugins). One entry point routes everything:

/meta-dev:skill-creation      # "I want to do <anything> to a skill / agent / hook / plugin / instruction file"

You rarely type it — sharp descriptions auto-route to the right operation, and a bundled guard nudges you onto the lifecycle when you edit an artifact directly.

Skills

The lifecycle operations, each right-sized to how often that artifact actually changes.

Router

• skill-creation — the entry point; diagnoses which artifact and which operation, then dispatches the matching skill.

Skill lifecycle

• create-skill — build a new Agent Skill from scratch: inventory for overlap → capture intent → draft to standard → static review → proportional empirical verification.
• update-skill — change an existing skill's behavior for a new requirement without regressing what it already did.
• improve-skill — raise quality with zero behavior change (debloat, formatting, description/triggering); runs hands-free.
• restructure-skill — split, merge, or move content between SKILL.md, references, and scripts when the shape is wrong.
• review-skill — read-only scored audit against the standards; routes each fix to the operation that applies it.
• retire-skill — safely delete or deprecate a skill: cross-reference scan → archive → remove.

Agent lifecycle

• create-agent — author a Claude Code subagent: least-privilege tools, model-by-task, the battle-tested system-prompt skeleton, and the skill↔agent seam checks.
• update-agent — change an agent's behavior, tools, model, or preloaded skills without regression.
• improve-agent — quality pass on an agent (prompt bloat, description, tool least-privilege); hands-free.
• review-agent — read-only audit of an agent, including why it isn't being delegated to.
• retire-agent — safely remove an agent, scanning for fork targets and hook agent_type bindings first.

Hook lifecycle

• create-hook — design and author a hook: narrowest event, right type, correct decision contract, robust script.
• review-hook — read-only audit of hook config and scripts (why one didn't fire, or isn't safe).

Plugin lifecycle

• scaffold-plugin — scaffold a plugin, wire its components, and set up skills-dir or marketplace distribution.
• validate-plugin — read-only validation of a plugin or marketplace, wrapping claude plugin validate plus a standards audit.

Instruction-file lifecycle

• setup-instructions — bootstrap a repo's AGENTS.md/CLAUDE.md + path-scoped .claude/rules/, skills and info derived per area.
• update-instructions — change instruction-file content and re-scope rules as the project evolves, under approval.
• review-instructions — read-only audit that verifies every documented claim against the actual codebase (stale versions, moved routes, dead commands), plus standards and cross-file consistency; routes each fix.
• improve-instructions — declutter and reformat instruction files without changing what they commit to; hands-free.
• revise-agents-md — user-invoked only (/meta-dev:revise-agents-md): a full-surface tune-up that audits every claim against the codebase, sorts each fact to its right altitude, reorganizes onto the canonical skeleton, and strips code-recoverable bloat — one autonomous pass, read-only toward your own notes.

Currency

• sync-docs — on-demand only (/meta-dev:sync-docs): diff every standard against current official Claude Code docs and route the drift. This is what keeps meta-dev from aging into fiction.

Agents

Shared judges and workers the lifecycles dispatch — read-mostly, single-purpose.

• grounder — read-only per-area grounding worker; revise-agents-md and setup-instructions dispatch one per path-area in parallel to ground the code and draft each rule cheaply, instead of reading serially on the caller's model.
• reviewer — read-only scored audit of a skill, agent, or instruction-file artifact against the standards, with routed fixes.
• grader — grades one eval run's assertions against the transcript and outputs, with cited evidence.
• comparator — blind A/B judge of two outputs without knowing which produced which.
• improvement-analyst — unblinds a comparison, explains why the winner won, and proposes prioritized fixes.
• benchmark-analyst — read-only cross-run pattern finder over benchmark data (non-discriminating assertions, flaky evals, cost tradeoffs).

How it works

• standards/ — the law, modular: core.md governs every .md artifact (formatting, per-line quality, model policy, rigor ladder, autonomy split, enforcement doctrine); per-artifact modules (skill, agent, hook, plugin, instruction-files) add only what's unique. Every skill and agent cites these, so a creator and a reviewer can never disagree.
• hooks/ — enforcement across five surfaces. A guard hard-denies create/delete/large edits of skills and agents (routing them through the lifecycle) while letting a one-line edit through with a nudge, and hard-gates two prevention-critical surfaces with no small-edit escape: hook-config changes and edits to upstream/vendored skills (any folder carrying an .upstream marker — blocked even inside the lifecycle, so vendored content is never locally forked by accident). Direct edits to project instruction files (CLAUDE.md/AGENTS.md) get a non-blocking nudge instead — a constitution is living documentation, so the guard just points you at /meta-dev:revise-agents-md (the full-surface tune-up) rather than blocking. Bash mutations are matched by operand-adjacency, so a filename merely named in a commit message never trips the guard. A md-format hook nudges (never blocks) on any .md write and warns when an AGENTS.md/CLAUDE.md surface grows past ~200 lines (a signal to move path-specific detail into .claude/rules/).
• skills/skill-creation/ — the shared harness: a vendored eval pipeline (parallel with-skill-vs-baseline runs, grading, benchmarking, blind comparison, an HTML review viewer) plus tools/ validators for skills and agents.

Licensing

meta-dev's own code and content are MIT (see LICENSE). The eval harness under skills/skill-creation/ (scripts/, eval-viewer/, assets/, references/schemas.md) is vendored byte-identical from anthropics/skills under Apache-2.0, with its LICENSE.txt retained; provenance and the re-sync procedure are in docs/vendoring.md.

Developing

meta-dev develops itself — editing an artifact routes through its own lifecycle. Dev loop when installed from a local marketplace:

# edit the repo, then:
claude plugin marketplace update farzan-dev
claude plugin update meta-dev
/reload-plugins            # hook/agent changes don't hot-reload