claude-resources

Production-grade Claude Code agents and skills for spec-driven development. Turns a generalist AI coding assistant into a team of specialist agents that collaborate through structured specs and reports.

What this is

Most AI coding assistants generate code without engineering discipline. This framework changes that by providing:

Structured workflows that enforce spec-first, test-proven, review-gated development
Specialist agents that collaborate like a real engineering team — architect, builder, tester, reviewer
Skills that encode battle-tested patterns, not just documentation
Project constitutions — enforce non-negotiable invariants across every spec and every review
Clarification round-trips — user answers blocker questions before the spec is synthesized, not after
Explicit contracts — optional contracts.md for API/data specs pins endpoints, schemas, and error codes
Parallelization markers — [P] on every subtask makes parallel-safety auditable
Operational learning that captures project-specific knowledge across sessions
Token-efficient output that reduces cost at scale without sacrificing reasoning or spec quality

The result: AI agents that work the way senior engineers do — clarify before building, spec before coding, test before shipping.

Prerequisites

Claude Code CLI or IDE extension
Git for plugin installation and project detection
Bash 4+ — hooks are pure bash, no other runtime required

Language support. Go-specific skills ship by default (15 under skills/go/). Other languages are supported via the same extension mechanism; session-start auto-detects pyproject.toml, Cargo.toml, package.json, and angular.json and emits the language context, but the skills/<lang>/ directories for Python, Rust, Node, etc. are not yet populated. See docs/extending.md for the steps to add a language tier.

Quick start

1. Install:

claude plugin add https://github.com/deandum/claude-resources

(Alternatives — clone, submodule — are in docs/getting-started.md.)

2. Configure your project:

cp claude-resources/EXAMPLE_CLAUDE.md your-project/CLAUDE.md

Edit CLAUDE.md with your tech stack, build commands, and project boundaries.

3. Run a workflow:

For complex tasks, use the single-entry-point flow:

/orchestrate  build a rate limiter for our API gateway with per-client 100 req/min limit

Main Claude — running the core/orchestration skill — walks the full workflow (Phases 1–4, Gates 1–3) and pauses for your approval at every phase boundary and between execution groups. The audit trail lives in docs/specs/<slug>/group-log.md.

For ad-hoc steps (no spec, no gates):

/ideate  a vague idea you want to sharpen           # optional — produces a clear task
/define  a clear task                                # runs Phase 1 + Phase 2, stops after Gate 2
/build   an already-defined task                     # resumes a spec's Phase 3, or spawns builder directly
/test    a package or function to test               # standalone tester spawn
/review  a diff or branch                            # standalone reviewer spawn
/ship    add containerization + observability        # in-orchestration or ad-hoc shipper

Main Claude IS the lead — there is no separate lead subagent. Specialist agents (critic, scout, architect, builder, cli-builder, tester, reviewer, shipper) execute bounded work; main Claude gates every phase boundary via AskUserQuestion.

Philosophy

Specs prevent rework. Clarifying requirements before writing code catches misunderstandings early — when they're cheap to fix.
Specialist agents produce better results. An architect thinking about structure and a builder thinking about implementation produce better work than a generalist doing both.
Common Rationalizations matter. The most dangerous shortcuts are the ones that sound reasonable. Explicitly cataloging them prevents clever circumvention.
Learning compounds. Capturing project-specific knowledge across sessions means agents get better at your project over time, not just better at coding.
Token efficiency matters at scale. Verbose output burns tokens without adding value. Compress human-facing prose; never compress specs or agent-to-agent artifacts.

Documentation

Getting started — install + first workflow walkthrough
Workflow — spec-driven workflow deep dive
Reference — every command, agent, and skill
Extending — add your own skills, agents, commands, or languages
Operations — hooks, learning system, and the opt-in ops plugin

License

MIT