live-panel

Turn a description of a system into an always-running architecture diagram: one fixed page that looks like a monitoring panel of a running system. Every frame is a complete, readable diagram; what moves is system state, not decoration. Two built-in themes: terminal-dark (the original look) and light-pastel (white background, rounded pastel boxes - for infographics, e.g. Xiaohongshu). Canvas presets: 4:5 (X, 1200x1500), 3:4 (Xiaohongshu, 1080x1440), 1:1 (1080x1080). Intended as the base for posts on X and Xiaohongshu.

Credit. The look and the motion idea come from an architecture-diagram clip by @thedelost (a Codex agent-tree panel, generated with GPT and screen-recorded as a web page): https://x.com/thedelost/status/2105398038026195279 . It reached many people through a quote-post by @slashui: https://x.com/slashui/status/2105850132365443528 . This skill is an independent re-implementation of the method (the motion grammar was learned by analysing that clip). examples/codex-agents/ is a recreation of the original picture and stays the original author's design. Keep this credit when you publish output made with the skill's Codex example.

When to use

• The user asks for an architecture / agent-tree / org-chart / data-flow diagram that should feel alive (video, GIF-like loop, or live page).
• The user has a description of a system with a few boxes, a flow between them, a few "on call" side triggers, a log and some numbers.
• Not for: diagrams that need interaction, real telemetry, or step-by-step explanations (use a slide or a normal diagram tool).

Input

A single JSON config (schema: references/config-schema.md). It holds everything: canvas size, colours, boxes with lines of text, wires, packet paths, side-rail triggers, log wording, numbers. The template never needs editing. Start from the closest example: examples/codex-agents/config.json (terminal-dark, 4:5, recreation of @thedelost's picture), examples/airbnb/config.json (terminal-dark, Chinese content, figures from an interview) or examples/agent-architecture/config.json (light-pastel, 3:4, a static Xiaohongshu infographic made to move, original by 小红书 @林纾). Theme: theme.preset; canvas: canvas.preset. Positions are absolute canvas pixels, so another aspect ratio needs its own config.

Choosing a view (视图)

If the content has one clear shape, use a {"type":"view"} element instead of placing boxes by hand (contract: references/views.md; gate: scripts/verify_view.py must PASS all 7 combos):

• repeating loop → ring; stage-by-stage drop-off → funnel (both ready, see examples/ring/, examples/funnel/).
• nested workflow with decisions and loop-backs (retry on "no", outer/inner loop, human gate) → loop (ready, see examples/loop-poteto/).
• relationships/clusters → nebula; work over time/parallel tracks → timeline; hierarchy → tree; a few watched numbers → kpi (planned, not in this version yet).
• an architecture with arbitrary boxes and wires → no view, free-form elements as in the original examples.

Fixed procedure

1. Collect content and where each number comes from. List boxes, who talks to whom, which things are "on call" triggers, what the log would say. Write the source (link or "simulated") next to every number. If a number has no real source, it is simulated: say so on screen (a footer line, or "(illustrative)" next to the counter). Never invent a figure and present it as real.
2. Write the config - scaffold first, then iterate. For boxes in rows (an agent tree, a pipeline, a fan-out/merge, a monitoring panel, with or without a side rail of on-call triggers), write a short spec - what each box says, rows top to bottom, optional lane (running/done), gauge (bar that flips past a threshold), rail, counters, log, note, credit; python3 scripts/scaffold.py --example prints one - and run python3 scripts/scaffold.py --spec spec.json --out out/config.json. It does the pixel layout, wires, packets, rail arrows, log columns and text wrapping, and its output usually passes check_frames.py on the first run; then edit the output (or the spec) in small steps. For a layout it cannot express (a centre box with satellites, curved arrows, a recreated picture with its own geometry), copy the closest example to your output path with a shell command (cp examples/<closest>/config.json out/config.json), or write a short Python script that builds the dict and dumps the JSON; then change it in small edits. Never type a whole config into a single message or tool call: a 20-40 KB JSON in one write is the most common way an agent run dies before anything is saved. Have a runnable draft on disk by your fifth tool call at the latest, then work in a loop: change one to three boxes, run check_frames.py, read its PROBLEM lines, fix, repeat. Do not plan coordinates or count text widths in your reasoning: if you catch yourself doing pixel arithmetic for more than a few lines, stop and write the edit instead - long in-head layout planning is the other way runs die before saving anything. Place boxes on a pixel grid (x, y, w, h in canvas px); text lines flow inside boxes at one fixed line height. Put the credit/source line in the top-level credit key ("credit": {"text": "...", "y": ...}); a plain text element at the bottom is not read as the credit. Skim references/motion-grammar.md before choosing machines and periods; references/config-schema.md lists every key, so you rarely need to read assets/template.html.
3. Render. python3 scripts/render.py --config my.json --out my.mp4 (needs Chrome/Chromium and ffmpeg on PATH, or --chrome / --ffmpeg; on macOS the browser is also found under /Applications/*.app). Linux and macOS only - on Windows run the scripts inside WSL. --html-out page.html also keeps the self-contained page, which runs live in any browser.
4. Check by frames, not by trust. python3 scripts/check_frames.py --config my.json --out-dir frames --repeat samples ~120 time points, measures text overflow and overlaps from the DOM, exports PNGs and proves replay determinism. Open the PNGs and look: the checker catches geometry, not taste. Fix the config and repeat until it exits 0. For text overflows its box, take the cheapest fix - shorten the text, drop a size step, or widen that one box - rather than redesigning the layout; a config that passes beats a better one that is still being planned when time runs out.

Motion rules in one screen

(Full text and rationale: references/motion-grammar.md.)

• Layout never moves. No camera, no step-by-step reveal. Frame 0 is already a complete diagram.
• Three tempos at once: fast (packets with trails on every wire, spinners, fast counters), medium (log scrolls with newest line bright and older grey; bars re-roll and flip colour and label past a threshold), slow (side-rail triggers light up one at a time in order, their arrow changes colour and carries a packet, advice is typed out, totals accumulate).
• One truth everywhere. The number in the log is the number on the bar; the trigger lit in the rail is the one the log mentions; counters in the status bar equal counters in boxes. In this engine that holds by construction: log lines are generated from the same state machines as the boxes.
• No real data? Say "illustrative". Fixed facts stay fixed; only the animation's own counters move, and they are labelled.

Determinism (why it renders the same every time)

The page exposes window.seek(t). Every visual is a pure function of t (and a fixed integer seed for the few "random" picks): no wall clock, no Math.random, no CSS animation. The renderer calls seek(i/fps) and screenshots; check_frames.py --repeat seeks away and back and compares bytes. Opened in a normal browser (no ?manual), the same function is driven by requestAnimationFrame and loops.

Files

• assets/template.html - the generic page. scripts/scaffold.py - short spec -> complete config (layout done for you). scripts/render.py - mp4. scripts/check_frames.py - checks + PNGs. scripts/livepanel.py - shared helpers (Chrome over the DevTools pipe, stdlib only).
• references/motion-grammar.md, references/config-schema.md.
• examples/codex-agents/, examples/airbnb/, examples/agent-architecture/ - config, mp4 and an animated preview.webp. examples/ring/, examples/funnel/, examples/funnel-karpathy/ (log scale, mono palette), examples/loop-poteto/ (grouped loop with decisions) - view examples.
• assets/views/<kind>.js - one file per view; references/views.md, references/view-rules.md, references/verify-view.md; scripts/verify_view.py.

Light theme notes

For infographics keep the original colours and structure; motion is gentler: small packets drift along every arrow, state-bound arrows turn the accent colour and carry two packets, active boxes get a glow, lists (steps, tools) light one row at a time from one shared cycle machine so that everything on screen refers to the same step. Use container:true on boxes that hold other boxes so the checker does not call nested boxes overlapping, and inside:true on icon text placed on a box.

Credit rule

Always put the original author and link of any picture you re-animate or recreate in the video footer (credit) and in the post: the author's handle and the full URL exactly as the user gave it (e.g. 原图：小红书 @作者《标题》 https://... · 动态复刻，非原作者作品). A handle without the link, or a link without the handle, is not enough. When the user gives a source for a real figure (a report, an interview), name that source in credit too. State when a figure is simulated.