# system-atlas

Adopt this skill: https://proskills.md/skills/other/inkboard-system-atlas-skills-system-atlas

Page: https://proskills.md/skills/other/inkboard-system-atlas-skills-system-atlas
This document is the public skill bundle for that page. It contains SKILL.md and every helper file the catalog carries. No login is required.

## File: skills/system-atlas/SKILL.md
size: 8031
upstream: https://github.com/inkboard/system-atlas/blob/main/skills/system-atlas/SKILL.md

```text
---
name: system-atlas
description: Build and maintain an explorable, progressively-disclosed isometric "atlas" of a system's architecture — an interactive page (hover to read, click to pin, go inside for steps, moving data packets you can inspect, chapters that reveal the system a few structures at a time) plus a generated text twin (SYSTEM.md) and question tracking by ID, all from one data file in the repo. Use this whenever someone wants to discuss, design, review, or explain an architecture visually — "make an atlas", "map the system", "make the architecture explorable", "visualize the codebase/agent/pipeline so we can talk about it", "a diagram I can click around", "walk me through how it fits together", or when an architecture discussion is producing a pile of open questions that need tracking across feedback rounds. Also use it to update an existing atlas after decisions change.
---

# System Atlas

An atlas is one data file that renders two views: an **interactive isometric map** (a single self-contained HTML file), and a **generated text twin** (`SYSTEM.md`) with the decisions table, every structure, the flows, and the open questions by ID. The data file is the only thing anyone edits; both views rebuild from it. It sits beside a hand-written glossary (`CONTEXT.md`) and ADRs.

The reason for this shape: an architecture discussion produces decisions, questions, and vocabulary faster than any one document can hold, and the person you are discussing with wants to *see* the system, not read it. The map is for them; the text twin is for the repo and for you next session; the single source is what keeps the two honest.

This skill was distilled from building a real agent-architecture atlas across one long design session and several rounds of feedback. The user's corrections from that session are the rules below; `references/process-and-lessons.md` has the story.

## When to reach for it — and when not

Use it when the system is new enough that vocabulary, decisions, and questions are still moving, and there will be more than one feedback round. Don't use it for a finished system that only needs a README, or for one diagram in a PR.

## Process

Follow the order — each step was earned by a correction the first time round.

1. **Read the inputs before drawing.** The vision doc, the repo's existing surfaces, and whatever prior art the user allows (ask — they may forbid a branch or a source). If you will build on a framework, read its docs first; hand long docs to a subagent with your specific design questions and have it return a primer with gotchas and a "what it does not give us" list. Drawing before this produces boxes that don't map to anything real.
2. **Discuss before drawing.** Propose the structure in chat, mapped to the runtime's real primitives, and ask only the questions you cannot derive from the repo. Take defaults for the rest and say which. Ask as plain chat text.
3. **First atlas — the whole system.** Copy `assets/` into the atlas home (`template.html`, `build.mjs`, and `data.example.mjs` renamed to `data.mjs`), fill the data, build, publish. **Where the atlas home is depends on the repo's docs policy.** Some repos commit design docs freely — then `docs/<system>/atlas/` in-tree is right. Other repos deliberately commit only ADRs and `CONTEXT.md`, with specs and evidence going to the issue tracker instead; in that case put the atlas, `SYSTEM.md` and `research/` in a git-ignored scratch directory and attach `SYSTEM.md` plus the research to the spec issue as comments when the spec is published, keeping only `docs/<system>/adr/` and `docs/<system>/CONTEXT.md` in-tree. Ask which policy applies before committing anything. Learned the hard way: committing the whole set produced a 3,900-line docs PR and four review rounds reconciling three restatements of one design — with ADRs plus a glossary only, there is one place to be consistent. If your agent has a design-guidance skill for HTML artifacts, load it before touching the template; read `references/design-language.md` for the visual rules either way.
4. **Progressive disclosure.** A whole system at once reads as noise ("hard to parse" was the first correction). Ten-ish chapters; each adds at most three structures and runs one small flow that only touches revealed structures; the last chapter shows everything with a flow picker. Unrevealed structures stay in the index, dimmed, with their chapter number. Panels are summary-first: one sentence, then *Read more* and *Steps* folded.
5. **Shapes and labels.** Letters on boxes are not enough ("better box shapes/labelling" was the second correction). Give each role a shape and put a readable name label on the canvas under every structure — see design-language.
6. **Text twin.** `CONTEXT.md` is a glossary and nothing else (domain-model format: the nouns, one line each); ADRs only for decisions that are hard to reverse, surprising without context, and the result of a real trade-off — these two are the in-tree pieces. `SYSTEM.md` is generated and `research/` holds evidence; both live with the atlas (scratch dir or `docs/`, per step 3). Don't open issues unless asked.
7. **Feedback by question ID.** Every question is `Q-<code><n>` with a state: open (a string), resolved `{q, r}` (answer + date), or routed `{q, to}` (handed to a named next step such as a deep dive). Record the user's words. If they call something "not a question", drop it; if they say "I don't get this", explain with a concrete example *before* resolving. After each round: rebuild, republish, update memory.
8. **Deep dives feed back.** Research with subagents against one shared brief (the interface we own, the requirements that separate candidates, a usage model for cost, a fixed deliverable shape). Write a synthesis with a normalized cost/fit grid. Fold resolutions into the data as `{q, r: '… (from the deep dive, date)'}`. If the user rejects a proposal, sweep *every* file and rewrite — a banner on top of a stale section is not enough; they will find it.
9. **Keep it current.** "The atlas is great for me — but not if it's not up to date." One source, rebuild and republish after every change, never hand-edit generated files, and leave a `README.md` in the docs folder explaining the set (table in process-and-lessons).

## Publishing the map

`atlas.html` is one self-contained file — no build step, no external assets beyond a Google Fonts stylesheet. Publish it whichever way the person can actually open:

- If your agent can publish a hosted HTML artifact, publish it there and keep the URL stable across rebuilds; put it in `META.artifactUrl` so the generated `SYSTEM.md` links to it.
- Otherwise serve the folder with any static server (`npx serve`, `python3 -m http.server`) and hand over the local URL, or commit the file and let the repo's pages host serve it.

Either way the rule is the same: one URL, republished after every data change, never a second copy.

## What done looks like

- `<atlas home>/data.mjs` exists and is the only edited source; `bun <atlas home>/build.mjs` writes `SYSTEM.md` and `atlas.html` without error.
- The atlas is published at a stable URL and republished there after every data change.
- Every structure has `one`, `what`, `how`, a `short` label, a role `kind`, and its questions; ghosts are marked; chapters exist with per-chapter flows; the last chapter is the whole system.
- `SYSTEM.md` carries the decisions table, the question index with IDs and states, and the "how this file is maintained" footer.
- Project memory records the atlas URL, docs paths, locked decisions with dates, what the user rejected and why, and the next step.

## Verify before publishing

Syntax-check the built script (`new Function(js)`), then look at it: serve the folder with a static server and open it in a real browser — `file://` renders as a static snapshot in some in-app browsers and the fonts may not load. Resize to ~1280×800 and screenshot a first chapter, a middle chapter, the last
```

## File: README.md
upstream: https://github.com/inkboard/system-atlas/blob/main/README.md

````text
# System Atlas

**An agent skill that turns an architecture discussion into something you can click around.**

![A walkthrough of a real system atlas](./demo.gif)

*Above: a real atlas of a production agent system, built with this skill — chapters reveal the system a few structures at a time, hovering reads a structure, and going inside one shows its steps in execution.*

[![skills.sh](https://skills.sh/b/inkboard/system-atlas)](https://skills.sh/inkboard/system-atlas)

```bash
npx skills add inkboard/system-atlas
```

## Why

Architecture discussions produce decisions, questions, and vocabulary faster than any one document can hold, and the person you are discussing with wants to *see* the system, not read it. So you draw a diagram, and a week later the diagram is a lie. An atlas fixes the shape of that problem rather than the diagram: one data file in the repo is the single source, and it renders two views that always agree — an **interactive isometric map** you explore with a mouse, and a **generated text twin** (`SYSTEM.md`) that carries the decisions table, every structure, the flows, and every open question by stable ID. The map is what you put in front of a person; the text is what the repo and your next session read; because both are generated from the same file, updating the design is one edit and a rebuild, and there is exactly one place where the system can be wrong.

## What you get

- **An interactive isometric map** — hover a structure to read it, click to pin it, arrow-right to go *inside* and see its steps in execution, pan and zoom the whole thing. One self-contained HTML file, no build step, no runtime dependencies.
- **Progressive-disclosure chapters** — a whole system at once reads as noise. Each chapter reveals at most three new structures and runs one small data flow that only touches what you have already seen; the last chapter shows everything with a flow picker.
- **Moving data packets you can inspect** — the dots travelling the hops are real requests. Click one and the panel shows its route and a representative JSON payload, so "how does a turn actually work" has a visual answer.
- **A generated text twin** — `SYSTEM.md`, rebuilt from the same data: the decisions table with ADR links, every structure's what/how/steps, the flow tables, and the question index.
- **Question tracking by ID** — every open question gets a stable `Q-<code><n>` and a state: open, resolved (with the answer and the date), or routed to a named next step. You can run several rounds of feedback against it without losing track of what was asked or who answered it.
- **Role shapes and readable labels** — the brain, the stores, the tool decks, the gates and the scheduled jobs each get their own isometric form and a name tag on the canvas, so the map is legible without a key.
- **The process, not just the renderer** — the skill carries the design language and the session lessons that produced it, so the agent knows the order to work in and the corrections to skip.

## Using it

Ask for one in plain language. The skill triggers on things like:

> "make me an atlas of this pipeline"
> "map the system so we can talk about it"
> "I want a diagram I can click around and poke holes in"
> "walk me through how this fits together"
> "update the atlas — we dropped the nightly job and picked a different vendor"

Then the agent works in this order:

1. **Read the inputs before drawing** — the vision doc, the repo's real surfaces, the framework's docs. Boxes that don't map to something real are worse than no boxes.
2. **Discuss before drawing** — propose the structure in chat, mapped to the runtime's actual primitives, and ask only the questions it cannot derive from the repo.
3. **Build the first atlas**, then **cut it into chapters** — the first version is always "hard to parse"; progressive disclosure is the fix, and it is the default here.
4. **Generate the text twin** and keep the glossary (`CONTEXT.md`) and ADRs beside it.
5. **Run feedback rounds by question ID** — record the person's actual words, resolve with a date, and sweep every file after a rejection. A banner on a stale section is not enough.
6. **Rebuild and republish after every change.** An atlas that is out of date is worse than no atlas.

## How it works

```
<atlas home>/data.mjs          ← the only file you edit
<atlas home>/build.mjs         ← bun build.mjs
<atlas home>/template.html     ← the renderer
        ↓
   atlas.html   +   SYSTEM.md  ← both generated, never hand-edited
```

`data.mjs` holds the structures (position, footprint, role shape, prose, steps, questions), the flows as hop lists with payloads, the chapters with what each reveals, and the decisions. `build.mjs` writes both views from it. `assets/data.example.mjs` is a working starter with every field documented — copy it, fill it in, run the build, and you have an atlas.

Where the atlas lives depends on your repo's docs policy. Repos that commit design docs can keep it at `docs/<system>/atlas/`; repos that deliberately commit only ADRs and a glossary should keep it in a git-ignored scratch directory and attach the generated `SYSTEM.md` to the spec issue instead. The skill asks before committing anything.

The map is a single self-contained HTML file, so publishing it is whatever the person can open: a hosted HTML artifact if your agent can publish one, a static server (`npx serve`, `python3 -m http.server`), or your repo's pages host. The rule is one URL, republished in place.

## What's in the repo

| Path | What it is |
|---|---|
| `skills/system-atlas/SKILL.md` | The skill: when to use it, the process, what done looks like |
| `skills/system-atlas/assets/template.html` | The atlas renderer |
| `skills/system-atlas/assets/build.mjs` | `data.mjs` → `atlas.html` + `SYSTEM.md` |
| `skills/system-atlas/assets/data.example.mjs` | Documented starter data — copy to `data.mjs` |
| `skills/system-atlas/references/design-language.md` | Layout, palette, isometric grammar, role shapes, copy rules, the chapter recipe |
| `skills/system-atlas/references/process-and-lessons.md` | How the first session went, the subagent deep-dive pattern, and the things that bit |
| `skills/system-atlas/evals/evals.json` | Three eval prompts: a new atlas, an update, and a small greenfield one |

## Installing

The [`skills` CLI](https://github.com/vercel-labs/skills) installs into Claude Code, Cursor, Copilot, Codex, Gemini, Cline, Amp and others:

```bash
npx skills add inkboard/system-atlas
```

Or drop `skills/system-atlas/` into your agent's skills directory by hand — for Claude Code that is `~/.claude/skills/system-atlas/` (personal) or `.claude/skills/system-atlas/` (per project).

## Credits

The visual grammar takes after the "codebase as interactive isometric diagram" style — khaki paper, black hatched structures, an index on the left, a reading panel on the right, and data packets you can catch. The process rules are the corrections from a real design session, kept in the order they were earned.

## License

MIT — see [LICENSE](./LICENSE).
````
