CCPM (Claude Code Project Manager) — Summary

CCPM is a spec-driven project management skill for AI coding agents that implements a five-phase PRD→Epic→GitHub Issues→Parallel Agents→Shipped Code pipeline. It ships as a single SKILL.md file installable via the agentskills.io format, making it compatible with Claude Code, Codex, OpenCode, Factory, Amp, and Cursor. The skill routes its work through five named phase documents (plan.md, structure.md, sync.md, execute.md, track.md) plus a set of bash scripts for deterministic operations, combining LLM reasoning for planning work with script execution for reporting. GitHub Issues is used as the source-of-truth for cross-agent coordination, with git worktrees providing workspace isolation for parallel agent execution. CCPM is notable for its benchmark-driven validation: the README claims 100% eval score on a structured 3-scenario test vs 27.7% for baseline agents.

Compared to seeds: most similar to taskmaster-ai (multi-phase workflow, GitHub/file-based task storage, parallel agent execution) but differs by using GitHub Issues as canonical task store (not local JSON), requiring gh CLI integration, and packaging everything as a single agentskills.io SKILL.md rather than an MCP server with 37 tools.

CCPM — Overview

Origin

Built by Automaze (https://automaze.io); developed and maintained by @aroussi. 8,146 GitHub stars as of analysis. MIT licensed. Active, last commit 2026-03-18.

Philosophy

"Stop losing context. Stop blocking on tasks. Stop shipping bugs. CCPM gives your AI agent a structured PM brain: turn ideas into PRDs, PRDs into epics, epics into GitHub issues, and issues into production code — with full traceability at every step."

"Core Principle: No Vibe Coding. Requirements live in files, not heads."

The framework positions itself against "vibe coding" — AI-generated code without structured requirements. Its argument: agents operating on documented, traceable issues produce better outputs than agents operating on verbal memory.

Manifesto-style positions

• GitHub Issues as source of truth: not a separate database, not a local file, but the team's existing issue tracker — visible to humans and agents simultaneously
• Script-first for deterministic operations: whenever the answer can be computed by a bash script (status, standup, epic list), run the script; save LLM calls for work requiring reasoning
• Parallel execution by design: every epic decomposition identifies independent vs sequential tasks upfront; agents launch in separate worktrees
• Full traceability: PRD → Epic → Issue → PR → Code — every artifact links to its parent

Target Audience

Small engineering teams and solo developers using AI agents with GitHub repos.

CCPM — Architecture

Distribution

Single SKILL.md file in the agentskills.io format. Install by copying skill/ccpm/SKILL.md into the agent's skill directory, or via the agentskills.io registry.

Install Complexity

Clone + copy one directory:

cp -r skill/ccpm ~/.claude/skills/ccpm

Or via agentskills.io compatible harness.

Required Runtime

• gh CLI (GitHub CLI) — required for GitHub Issues sync
• gh extension install yahsan2/gh-sub-issue — for parent-child issue relationships (optional fallback available)
• Git with worktree support

Directory Tree (after CCPM setup)

<project-root>/
├── .claude/
│   ├── prds/
│   │   └── <feature-name>.md          # PRD files
│   └── epics/
│       └── <feature-name>/
│           ├── epic.md                 # Epic definition
│           ├── 001.md                  # Task file (pre-sync)
│           └── 1234.md                 # Task file (post-sync = issue number)
└── skill/
    └── ccpm/
        ├── SKILL.md                    # Main skill dispatch file
        └── references/
            ├── conventions.md          # Path standards, frontmatter schemas
            ├── plan.md                 # Phase 1: PRD writing + parsing
            ├── structure.md            # Phase 2: Epic decomposition
            ├── sync.md                 # Phase 3: GitHub sync
            ├── execute.md              # Phase 4: Parallel agent launch
            ├── track.md                # Phase 5: Status + standup
            └── scripts/
                ├── status.sh
                ├── standup.sh
                ├── epic-list.sh
                ├── epic-show.sh
                ├── epic-status.sh
                ├── prd-list.sh
                ├── prd-status.sh
                ├── search.sh
                ├── in-progress.sh
                ├── next.sh
                ├── blocked.sh
                └── validate.sh

Config Files

None — CCPM is configuration-free. All state lives in .claude/prds/, .claude/epics/, and GitHub Issues.

Target AI Tools

Claude Code (primary), Codex, OpenCode, Factory, Amp, Cursor — any agent with skill support.

CCPM — Components

Skill File

skill/ccpm/SKILL.md — single dispatch skill that triggers on PM-related phrases. Contains the description: frontmatter for semantic matching and phase routing logic.

Phase Reference Documents

Bash Scripts (deterministic operations)

Artifact Schemas

PRD frontmatter:

---
name: <feature-name>
description: <one-line summary>
status: backlog
created: <ISO-8601>
---

Epic frontmatter:

---
name: <feature-name>
status: backlog
created: <ISO-8601>
progress: 0%
prd: .claude/prds/<name>.md
github: (set on sync)
---

CCPM — Uniqueness & Positioning

Differs from Seeds

CCPM is closest to taskmaster-ai in the seed set — both implement multi-phase PRD-to-task workflows with parallel agent execution. The key architectural delta: taskmaster-ai stores all tasks in a local tasks.json file accessed via an MCP server with 37 tools, while CCPM uses GitHub Issues as the authoritative task store (synced from local .claude/epics/ files) and packages its intelligence as a single SKILL.md rather than an MCP server. This means CCPM requires a GitHub repo and gh CLI but gains cross-machine visibility, human-readable issue tracker integration, and no local database dependency. Unlike bmad-method (role-based persona agents), CCPM is a single-skill dispatch system. Unlike superpowers (behavioral skills for Claude Code only), CCPM works across any agentskills.io-compatible harness. The "script-first for deterministic operations" pattern (routing status/standup to bash scripts rather than LLM calls) is unusual and not found in any seed framework.

Distinctive Positioning

• Only framework in this batch with an explicit benchmark claim (100% vs 27.7% baseline on 3-scenario eval)
• GitHub Issues as canonical task store — unique in the batch; most use local files
• Script-first routing (bash for reporting, LLM for reasoning) — explicit LLM token optimization
• Lazy-loading phase reference documents (each phase reads only its relevant references/*.md) — reduces initial context size
• Git worktree isolation for parallel agents — specific execution isolation strategy named explicitly

Observable Failure Modes

• Requires GitHub repo — doesn't work offline or with non-GitHub hosts (GitLab, Bitbucket)
• Requires gh CLI and gh-sub-issue extension — setup friction for new users
• Local .claude/ files can drift from GitHub Issues if another agent works from a different machine
• No conflict resolution strategy if two agents close the same issue simultaneously
• PRD/epic quality depends on LLM's brainstorming quality — no structural validator for requirements completeness beyond checklist prompts