Figure

One animated SVG, no scripts, no upload: it renders and plays anywhere markdown does — GitHub README, PR and issue comments, blog posts, Notion. ~15 kB for a figure that would be a 2 MB video, sharp at any size, and it follows the reader's light or dark theme.

You write a JSON spec — never SVG, never a .ts figure file (those are only for the docs site's interactive figures, see below). One command renders it. No install, no build, no browser.

1. Write the spec

A spec is { "props": { "layout": …, "edges": […], "steps": […] } }. Do not save it as a file in the repo. It is an input, not a deliverable, and a stray spec.json is the one mess this skill must not leave behind — pipe it in (step 2), and the rendered SVG keeps a copy of it for you.

{
  "props": {
    "speed": 2000,
    "layout": {
      "gap": 48,
      "children": [
        { "id": "agent", "label": "Your AI Agent" },
        {
          "id": "api",
          "label": "Hindsight API",
          "direction": "column",
          "gap": 24,
          "children": [{ "id": "retain", "label": "Retain", "sub": "LLM extraction" }]
        },
        {
          "id": "bank",
          "label": "Memory Bank",
          "direction": "column",
          "gap": 28,
          "children": [
            { "id": "facts", "label": "Facts", "sub": "world · experience", "shape": "store" },
            { "id": "obs", "label": "Observations", "shape": "store" }
          ]
        }
      ]
    },
    "edges": [
      { "id": "call", "from": "agent", "to": "retain", "label": "retain()" },
      { "id": "store", "from": "retain", "to": "facts", "label": "extract" },
      { "id": "consolidate", "from": "facts", "to": "obs", "label": "consolidate", "quiet": true }
    ],
    "steps": [
      {
        "label": "retain()",
        "flow": [
          {
            "edges": { "edge": "call", "data": "“Alice joined Google in March”" },
            "say": "Your agent sends what happened."
          },
          {
            "edges": "store",
            "show": {
              "facts": [
                {
                  "tag": "world",
                  "tone": "blue",
                  "text": "Alice joined Google",
                  "meta": "Mar 2026",
                  "mark": "new"
                }
              ]
            },
            "say": "An LLM pulls out the facts."
          },
          {
            "edges": "consolidate",
            "ms": 2600,
            "show": { "obs": [{ "text": "Alice works at Google", "meta": "2 sources" }] },
            "say": "The worker merges them into one belief."
          }
        ]
      }
    ]
  }
}

Layout — a tree. A group has children, and label (which draws a frame around it), direction: "row" | "column", gap, align. Anything else is a box: { id, label, sub?, shape? }, where shape is "store" for a database cylinder (data at rest) or "decision" for a diamond. Plain boxes are the things that do something. Give every box a stable id.

Edges — { from, to, label?, id?, around?, quiet? }; from/to name a box or a group. around: "above" | "below" arcs over the boxes in between; quiet: true draws the edge only while a step uses it (for long edges that would cut across the picture).

Steps and beats — each step is one story the figure tells; the SVG plays them in a loop. A beat is one moment: edges (a hop id, { edge, back, data } for a reverse hop or a data chip, or an array to run several at once), say (the caption; it stays until the next say), show (fills the content card inside a box and persists to the end of the step), light (highlight boxes for that beat), ms (how long the beat lasts).

Card rows — { tag?, tone?, text, meta?, mark?, mono? }. tone is blue | purple | green | orange | gray. Use tag for the kind of thing (world, user, page), meta for a detail, and mark for what happened to it (new, ✓, cited, ↻).

Keep it honest and specific: real example data beats placeholders, and every claim in a label, card or caption must match what the code actually does — check the code, don't assume.

2. Render it — from stdin, so nothing is left on disk

cd hindsight-interfig
npm run svg -- - ../path/to/out.svg <<'SPEC'
{ "props": { "layout": …, "edges": […], "steps": […] } }
SPEC

The quoted <<'SPEC' heredoc passes the JSON through untouched, and the only file produced is the SVG. Zero dependencies, no browser, no build.

Other forms:

npm run svg -- what-hindsight-does out.svg   # a figure from figures/, by name
npm run svg -- --spec out.svg                # print the spec an SVG carries, to edit and re-render

Every SVG embeds its own spec in <metadata>, so a figure stays editable without anyone keeping the JSON: read it back with --spec, change what you need, render again. That is why a spec file is never worth committing.

Use npm run svg, not node scripts/... directly: the script name is the interface, the path is not.

3. Look at it before you ship it

Always. Text that overflows its box, an arrow crossing a box, a caption that does not match what is moving — obvious on sight, invisible in the source.

cd $(dirname out.svg) && python3 -m http.server 8777 &   # the browser tool blocks file:// URLs

Then open http://localhost:8777/out.svg, screenshot it, wait a few seconds and screenshot again to catch a later beat. open out.svg works too when a human is watching.

4. Put it where it belongs

• PR or issue comment — commit the SVG on the branch, then reference its raw URL: ![figure](https://raw.githubusercontent.com/<owner>/<repo>/<branch>/<path>.svg). GitHub renders and animates it immediately. (Dragging a file into a comment also works, but only a human can do that, and GitHub keeps a fresh upload private for several minutes.)
• README or repo docs — commit it and link it with a relative path.
• Blog post — hindsight-docs/static/img/blog/, referenced as /img/blog/<name>.svg. An SVG under hindsight-docs/static/ is also read by the docs-skill generator: it pulls the spec out of the image and writes the narration into the skill, so an agent reading the docs gets the content and not a dead image link.
• The docs site's own pages — don't use an SVG. Those pages embed the interactive React figure, which has tabs, pause, speed and hover. Add a hindsight-interfig/figures/<name>.ts instead and <Flow {...figure.props} /> on the page (see hindsight-interfig/README.md).

A worked example

The spec in step 1 is a complete, working figure — copy it and edit. Any SVG you find carries its own spec too: npm run svg -- --spec <file>.svg prints it back, so an existing figure is the fastest starting point for a new one.

What the SVG cannot do

It loops through every step with no controls: no tabs, no pause, no hover. If the figure needs those, it belongs on the docs site as a React figure. Keep an SVG to one or two steps so the loop comes back round quickly.