design-harness

Turn scattered reading and half-formed ideas into a system design you can defend. The human adjudicates, the agent runs the errands.

An Agent Skill (Claude Code plugin, works with any SKILL.md-compatible agent): feed it sources and rough ideas as you go — it files and links them, records how your thinking evolves, and assembles the evidence into a system design.

Markdown keeps the record; the canvas makes it readable — every source and idea you add lands on the board, one glance away from any decision's why.

▶ Click the board to open the live canvas — this repo's own design workspace, dogfooded, rebuilt from the markdown on every merge.

What This Does

• Sources file themselves — drop papers, repos, blog posts; the agent files one card per source, grades it, and anchors every claim to where it came from.
• Ideas are yours alone — only a human creates an idea card; the agent links it to evidence, clusters neighbors, and surfaces conflicts instead of picking a side.
• Output assembles on your word — one command turns surviving ideas into the deliverable, and every later sync runs on your word too; edit output directly and it flows back into ideas.
• Everything is traceable — every design element links back through ideas to the sources that earned it, with an append-only log on every card.
• Plain Markdown, no lock-in — three folders in your repo; versionable, greppable, renders on GitHub. The agent is just the runtime.
• One visual canvas — a self-contained HTML projection of the whole board: no server, no dependencies, rebuilt on demand and thrown away.

Installation

Claude Code:

/plugin marketplace add tigerless-labs/design-harness
/plugin install design-harness@design-harness

Codex (CLI and the ChatGPT desktop app share one plugin system):

codex plugin marketplace add tigerless-labs/design-harness

then install from /plugins in the CLI or the desktop app's plugin directory.

For other SKILL.md-compatible agents, copy the skill folder into your agent's skill directory:

git clone https://github.com/tigerless-labs/design-harness
cp -r design-harness/plugins/design-harness/skills/design-harness \
  ~/.claude/skills/

Usage

Talk to your agent in plain language:

> file these papers onto the board

> assemble the design

1. Drop material — the agent files and grades each source as a card.
2. Think out loud — your judgments land as idea cards, linked to their evidence.
3. Say go — the ideas assemble into the output your target.md declares.
4. Ask for the canvas — the agent builds the board and hands you a clickable link.

Render your own workspace in ten seconds — the demo board is the same projection run on a real workspace:

python3 plugins/design-harness/skills/design-harness/scripts/build_canvas.py \
  path/to/your-workspace -o /tmp/canvas
open /tmp/canvas/canvas.html

Canvas Styles

Five skins ship with the canvas; switch live from the toolbar, Pin & Paper is the default. Screenshots below are this repo's own workspace, unretouched — the board view and a close-up for each style.

Pin & Paper (default)

A handmade pinboard: yellow paper ground, ink-blue hand script, pinned notes, taped slips, and dashed hand-drawn threads between them.

Notebook Tabs

Warm ivory stationery: serif italic headings, index cards with colored tab tongues, sticky-note ideas, stitched dashed connections.

Swiss Modern

White, ink, and one red accent: hairline grid, giant grotesk world titles with red ordinals, ruler-straight edges.

BlockFrame

Neo-brutalist pastel decks: thick borders, hard offset shadows, cyan / pink / yellow card blocks that tilt just enough to feel alive.

8-Bit Orbit

A CRT arcade in deep navy — scanlines, neon cyan / yellow / magenta frames, and orthogonal circuit-trace edges. Dark in both color schemes; a CRT has no daylight mode.

Philosophy

1. The human adjudicates, the agent runs the errands — agents execute well and judge poorly, and a wrong design costs the whole branch of code built on it.
2. Markdown is the only source of truth — every other surface, the canvas included, is a disposable projection.
3. Rejected ideas are archived with their reasons, never deleted — six months later, "why didn't we do X?" has an answer.
4. Every element of the output must answer "why?" with a link.

This repo is dogfooded on itself: its own design was produced by running the method on ~80 sources about decision-making methods.

License

MIT — vendored renderers (marked, DOMPurify) keep their original headers.