🚀 Rocketman

Idea to production, managed in the repo.

A project management hub that lives in your codebase as one beautiful, offline HTML file — and a Claude skill stack that walks a product from raw idea to shipped, iterated reality.

Built by Winsen Labs · MIT licensed

spec · kanban · roadmap · decisions · debugging · docs — all in PM/index.html

Responsible token-maxing

Token-maxing has become a thing — throw enough tokens at a model and watch it go. We've always been fans of the responsible kind: tokens spent with a plan, a memory, and a paper trail, so the output is something you'd actually ship.

That's what Rocketman is for. Rocketman + Claude Code turns any idea you have into a production-grade application on autopilot — the whole nine yards. Not a demo, not a prototype that falls over on the second click. The idea gets pressure-tested, specced, planned, built by a fleet of agents, verified, human-tested, launched, and iterated — and every token spent leaves a trace in a hub you can read in ten seconds.

The track is the discipline; the hub is the memory; the agents are the throughput. Point them at an idea and let them run.

Why a local HTML hub beats online PM tools in the age of AI coding agents

Project management tools were built for humans clicking around a browser. But the primary worker on a modern codebase is increasingly an AI coding agent running inside the repo — and for that worker, a SaaS board behind an API, an OAuth wall, and a rate limiter is exactly the wrong shape. Rocketman is built on a different bet:

1. The context should live where the work happens — in the repo. An agent shouldn't need network access, an API token, or a third-party MCP round-trip to learn what the project is, what's blocked, and why a decision was made. It should read a file. Rocketman puts the spec, the board, every decision, and every debugging trail in PM/ as structured data that any agent loads instantly — no auth, no latency, no rate limits, no "the integration is down."

2. HTML is the format AI is unreasonably good at. As Thariq Shihipar argues in The Unreasonable Effectiveness of HTML, agents produce and reason over HTML extraordinarily well, and a self-contained .html file is immediately usable — no build step, no export, no viewer. HTML makes spatial, structured information (a kanban, a diff, a timeline, a decision tree) visible at a glance instead of flattened into prose. His line captures the whole idea: it turns "a document you'd skim into one you'd actually read." That is precisely the gap between a markdown spec nobody opens and a hub you actually look at.

3. It version-controls with the code. Because PM/ is committed, your project state branches, diffs, reviews, and merges with the code itself. A feature branch carries its own spec and board changes; they show up in the PR; they can never silently drift from the work they describe. Try that with a Linear board.

4. It's offline, portable, and yours. One double-clickable file. Works on a plane. Outlives any vendor. No seat pricing, no data export, no lock-in.

5. Structured data keeps it honest. Raw HTML is miserable to diff and unreliable for an agent to hand-edit. So Rocketman splits data from presentation: small, diffable JSON is the source of truth; the gorgeous HTML is generated and read-only. Agents edit data; the build renders the hub.

Local-first, sync-optional. None of this means abandoning the tools your team already loves. When a stakeholder wants it in Linear, FanDesk, Cheetah, or Winsen, the /rm-mirror skill mirrors the whole hub into them on demand. The repo stays the source of truth; the SaaS board becomes a read-out.

Hand it to Claude

New here? Don't read the docs — let your agent evaluate it for you. Copy this into Claude Code (or any coding agent) inside a project you care about:

Take a look at Rocketman: https://github.com/winsenlabs/rocketman
(npm: @winsendotai/rocketman). It's an in-repo HTML project hub plus a Claude
skill track that walks an idea from product thinking → spec → plan → build →
verify → launch, all managed inside the repo.

Evaluate whether it would actually be useful for THIS project. If it is, run
`npx @winsendotai/rocketman init` to scaffold it, open PM/index.html, and walk
me through what it set up and how I'd use it day to day. If you don't think it
fits, tell me why instead of installing it.

The agent reads the repo, makes the call, and sets Rocketman up for you if it's a fit.

Quickstart

# in any repo
npx @winsendotai/rocketman init      # scaffold PM/ + the Claude skill stack
npx @winsendotai/rocketman build     # render PM/data/*.json → PM/index.html
open PM/index.html      # zero deps, opens offline

That's it. Commit PM/ alongside your code and you have a living project hub that travels with the repo.

Or clone this repo and open PM/index.html to see Rocketman managing its own development.

What you get

| | | |---|---| | 🗂 One offline file | The whole project renders to a single self-contained PM/index.html. No server, no deps, no network. | | 🎛 Eleven views | Dashboard · Board · List · Spec · Roadmap · Decisions · Debug · Docs · Activity · Fleet · Dependencies — a "calm command center" UI. | | 🧠 AI summary lines | Every entity carries a one-line summary, so the whole board is skimmable at a glance. | | 🔗 Backlinks | [[id]] wiki-links between tasks, ADRs, debug logs, and spec — with automatic "referenced by". | | 👥 Provenance | Every item and event is attributed human vs agent (opus / sonnet / haiku). | | 🛠 The Rocketman Track | A Claude skill stack that forces product thinking and drives idea → production. | | 🔁 Sync-optional | Mirror the hub into Linear / FanDesk / Cheetah / Winsen whenever you want. | | 🤖 Fleet-ready | The board is a work queue; a parent agent allocates ready tasks to sub-agents on demand. | | 📡 Agent relay + Fleet view | Agents in separate terminals message each other and hand off tasks over a conflict-free file bus — and the Fleet view shows every conversation, handoff, and who's online. No more copy-paste between tabs. |

The Rocketman Track

The skill stack walks a product from a raw idea all the way to a shipped, iterated product. Every stage reads from and writes back to the same PM/ data, so the hub is always current.

  Ideate ─▶ PRD ─▶ Plan ─▶ Build ─▶ Verify ─▶ Test ─▶ Launch ─▶ Iterate
   force    spec    track   fleet     gate    human    ship    changelog
  thinking         + board                    test
            └──────────── /rm-research ────────────┘
            verify the current, correct stack before you build on it

The skill stack

| Skill | Stage | What it does | |---|---|---| | /rm-ideate | Ideate | Forces product thinking — six questions on demand, status quo, wedge, and why-now before any code. | | /rm-research | Any | Verifies the current, correct tech stack / library / API / version before you build on it — the antidote to agents reaching for a stale stack from memory. | | /rm-prd | PRD | Turns the idea brief into a structured spec (vision, personas, requirements, non-goals, risks). | | /rm-plan | Plan | Decomposes the PRD into milestones, epics, and a board of ready tasks. | | /rm-build | Build | Turns the board into a work queue; allocates ready tasks to sub-agents on demand. | | /rm-verify | Verify | Runs the automated gate — build, lint, tests, and the hub doctor. | | /rm-test | Test | Generates a crisp human test script and records pass/fail back onto the task. | | /rm-launch | Launch | Ships: tags, updates the changelog, cuts the release. | | /rm-iterate | Iterate | Captures post-launch changes as new tasks + changelog entries. | | /rm-relay | Multi-agent | Lets agents in separate terminals message each other and hand off tasks over a polled PM/comms/ bus. Pair with /loop. | | /rm-mirror | — | Mirrors the hub into Linear / FanDesk / Cheetah / Winsen (sync-optional). | | /rm-init · /rm-sync · /rm-decision · /rm-debug · /rm-brief | Infra | Scaffold, rebuild, record an ADR, log an investigation, brief project status. |

Skills live in .claude/skills/ and load automatically in Claude Code. The conventions that make any agent fluent live in PM/CLAUDE.md.

How it works

        you / an agent edit                      the build                    you open
   ┌───────────────────────────┐         ┌──────────────────┐         ┌──────────────────┐
   │  PM/data/core.json        │         │  engine/build.mjs │         │  PM/index.html   │
   │  PM/data/tasks.json       │ ───────▶│  flat-merge +     │ ───────▶│  one offline,    │
   │  PM/data/spec.json        │         │  inline css/js    │         │  read-only hub   │
   │  PM/data/content.json     │         │  → single file    │         │  (8 views)       │
   └───────────────────────────┘         └──────────────────┘         └──────────────────┘
      source of truth (diffable)            deterministic, 0 deps          generated artifact

• Data is four small JSON files — easy to diff, easy for an agent to edit reliably.
• build.mjs flat-merges them into one #pm-data island and inlines the engine CSS + JS into a single HTML file. Deterministic: same data in, byte-identical hub out.
• The HTML is generated and read-only. Never hand-edit it — edit the data and rebuild. (ADR-2)

See PM/docs/ for the full Diátaxis documentation (tutorial · how-to · reference · explanation) — the same Markdown also renders inside the hub's Docs view. The how-tos cover adding tasks, decisions, debug logs, spec sections, Markdown docs, and PDF/image attachments.

Project structure

rocketman/
├── PM/
│   ├── index.html          # the generated hub — open this  (do not hand-edit)
│   ├── CLAUDE.md           # conventions that make agents fluent
│   ├── data/               # the source of truth (core, tasks, spec, content)
│   ├── docs/               # Markdown docs → rendered in the hub's Docs view
│   ├── files/              # attachments (PDF, images) embedded into the hub
│   └── comms/              # the agent relay bus (presence, messages, handoffs)
├── engine/                 # the hub engine (reusable, zero-dep)
│   ├── build.mjs           # the content pipeline: data + docs → index.html
│   ├── md.mjs              # the Markdown → HTML converter
│   ├── app.js              # the hub renderer
│   └── hub.css             # the "calm command center" theme
├── templates/              # what `rocketman init` scaffolds into a repo
├── bin/rocketman.mjs       # the CLI (init · build · new · doctor · serve · relay)
└── .claude/                # the Rocketman Track skills + guard hook + settings

Requirements

Node 18+. That's the whole list — Rocketman has zero runtime dependencies.

Browser note: the theme uses oklch() colors; open the hub in a 2023+ browser (Chrome/Edge/Safari/Firefox). See BUG-3 in the hub for the supported baseline.

Contributing

Rocketman dogfoods itself — its own roadmap lives in PM/index.html. See CONTRIBUTING.md and CHANGELOG.md.

License

MIT © Winsen Labs.