Projects Code
Full-Chain Development Skills
Twenty-one open-source agent skills that walk a project from idea to release — research, PRD, spec, design, test-driven build, and retrospective — with real stop points.
- When
- 2026
- My role
- Author and maintainer
- Built with
- Agent Skills, Claude Code, Python, MCP, Markdown
- 21 skills covering 11 stages of development
- 11 structural checks, each with its own tests
- MIT licensed, usable chained or one at a time
Project introduction
Why
Coding agents are fast, but the expensive mistakes happen before any code is written: the wrong scope, a spec nobody agreed on, an assumption nobody tested. I wanted a way of working where an agent moves quickly inside each stage and stops where changing your mind later is costly.
What it is
A library of reusable agent skills, one per stage, from a project ledger and kickoff research through adversarial architecture selection, MVP convergence, the PRD, a four-step spec pipeline, interface design, project context, test-driven implementation, testing, retrospective, and release packaging. A thin orchestrator skill works out the next stage and keeps a record of where the project stands.
Design choices
- Every skill also works alone. Each upstream artifact has a “provide it directly” path, so you can use just the PRD writer or just the test router.
- Gates are real stops. Where a stage is gated, the skill says what it’s waiting for and waits.
- Honest labelling. When a skill falls back, its output says what wasn’t covered. A report with “unverified” in it beats one that looks green and isn’t.
- Safe installs. The installer is idempotent, reports customized skills instead of overwriting them, and keeps English and Chinese editions strictly separate.