pm

pm — Rally-style PM tree plugin

> 📊 Live example dashboard → — a worked PM tree, rendered by the skill's own tooling, showing how pm plans and executes project development. (The example tree under example/ documents how this very skill was built.)

A Claude Code plugin with a single pm skill for working a Rally hierarchy (Epic → Feature → User Story → Task) stored as a tree of CLAUDE.md files under pm/E<NNN>/. The skill has three modes, selected by the first token of the invocation (or inferred from intent).

The workflow was originally authored for the landfinder repo's pm/E100/ tree but applies to any repo laid out the same way.

Three modes of /pm

| Invocation | Mode | What it does | Triggers on | |------------|------|--------------|-------------| | /pm init … | Scaffold | Builds a tree from a design brief: writes pm/PLAN.md, the E<NNN>/CLAUDE.md charter, the F/US/TASK folder skeleton, installs the two helper scripts, and auto-assigns each story a Model:. | "scaffold the PM tree", "bootstrap the Rally hierarchy", "set up pm/E100" | | /pm plan … | Refine | Reshapes an existing tree in a planning session: split stories, repair dependency graphs, re-balance phases, sharpen acceptance criteria, re-evaluate model assignments. Additive, never destructive. | "refine the plan", "groom the backlog", "re-decompose this feature", "re-prioritize" | | /pm … | Execute (default) | Builds the tree out: a long-running PM agent owns a Feature and spawns short-lived Worker agents — one per story, in a dedicated git worktree, on that story's assigned model — then reviews and merges their work. | "act as pm for F1xxx", "claim a story", "worktree for a story", "I'm the pm" |

The init and plan workflows are reference prompts (references/init.md, references/plan.md) that SKILL.md routes to based on the first word of the request; execute is the default in SKILL.md itself. One skill, one set of bundled scripts/ and references/, no duplication.

The Model: contract

Each story's CLAUDE.md carries a Model: line (claude-sonnet-4-6 by default, claude-opus-4-8 for architectural / ambiguous / algorithmic / cross-cutting / security-sensitive work). init mode assigns it at scaffold time, plan mode re-evaluates it, and execute mode reads it to decide which model to spawn each Worker on. Picking the model is a planning decision, not a spawn-time guess.

Status log & timestamps

Every epic/feature/story file keeps a ## Status log — one UTC-timestamped line per transition (Created → In progress → Done, with optional notes), giving each node a full audit trail rather than just a current state. Status is changed through the installed set-status.ps1, which rewrites the Status: line, stamps the log entry, bumps Last updated:, and regenerates the HTML tracker in one atomic call, so the file, its history, and the tracker never drift apart. The tracker surfaces created/started/done timestamps as chips on each page.

Self-contained pm/ folder

Everything the workflow needs lives under the target repo's pm/: PLAN.md (the roadmap, rendered into the dashboard), the two helper scripts, the generated index.html dashboard, and the epic tree. init mode lays this down; if you run execute mode on a tree missing the scripts, it installs them from the skill's own bundled copies (${CLAUDE_SKILL_DIR}/scripts/).

Repository layout

skills/
└── pm/                                 ← the single skill
    ├── SKILL.md                        ← mode dispatch + the execute-mode workflow
    ├── scripts/
    │   ├── build-pm-html.ps1           ← tracker generator (installed into target repos)
    │   └── set-status.ps1              ← status change → timestamped log → HTML regen
    └── references/
        ├── init.md                     ← `/pm init` reference prompt (scaffold)
        ├── plan.md                     ← `/pm plan` reference prompt (refine)
        ├── tree-structure.md           ← master templates, numbering, model heuristic, status-log standard, HTML structure
        └── worktree.md                 ← git worktree commands for the per-story workflow
.claude-plugin/
├── plugin.json
└── marketplace.json

SKILL.md reads the first token of the request and routes to references/init.md or references/plan.md, or stays in execute mode. references/tree-structure.md is the single source of truth for the templates; init mode seeds them into a target repo's charter, and plan/execute read the seeded copy from there.

Install

claude plugin marketplace add chrishuffman5/pm
claude plugin install pm@pm --scope user

Restart Claude Code; the skills then activate automatically on their trigger phrases.

License

MIT — see LICENSE.