Phased Agents — AI skill for Claude Code

phased-agents

**A methodology for AI-driven long-running software projects, plus a copy-and-go template for Claude Code.**

If you've started a multi-month project with an AI coding agent and watched it drift into chaos around month two — losing track of what's decided, why, and what actually works vs. what merely compiles — this is for you.

What it adds

Standard Claude Code is great at single tasks and weak on long projects: no memory across sessions, no decision audit trail, no enforced verification, no guard against bugs that pass tests but don't work. `phased-agents` adds:

1. A project bible (CLAUDE.md) loaded every session — goal, current phase, conventions, how to run things.
2. Five specialized subagents — researcher, implementer, reviewer, eval-runner, doc-keeper — each with one job, so they catch each other's mistakes.
3. Phases with hard acceptance gates — measurable contracts per chunk of work. No "directionally good enough."
4. A pipeline — research → implement → review → eval → doc-keep — that each sub-phase runs and lands as a PR.
5. PR discipline — only the project-lead merges to main, so it stays a clean audit trail. Independent sub-phases run in parallel, each on its own branch/worktree.

Come back after a week and you know exactly where things are, why each decision was made, and what's next.

When to use it

Worth the overhead when a project has at least one of: a multi-month timeline, risk of silent failures (algorithms, security, performance, correctness), multi-domain expertise (papers, RFCs, design systems), or decisions with lasting consequences.

Overkill for one-week builds, throwaway prototypes, or marketing sites — use a plain Claude session for those.

Quick start

Open Claude Code in your project directory (or an empty one for a new project) and paste:

Use the phased-agents methodology in this directory. Clone
https://github.com/ameri110/phased-agents into a temp dir, read its SETUP.md,
and fo