Prezit

prezit — presentations from a subtree

One anchor, one story. The body of the deck is the anchor and everything under it; the intro is built from the anchor's ancestors and the wiki pages they link. A hypothesis anchor is legal and yields a shorter deck. No time-window, no multi-anchor: "what happened this week" is answered by choosing the anchor whose subtree moved.

The three-way split (who does what)

You never retype the chrome (copy the template), and you never let the engine author a sentence — selection and phrasing are your whole job.

The mapping — what each crux object becomes

Shorthand: intro = wiki, methods = design (not implementation legwork), results = tree.

Workflow

1. Harvest. crux deck <anchor> --json from inside the vault. On a large vault you may delegate the initial sweep to a read-only search subagent — it returns paths and structure only.
2. Read end to end. Every linked report under results/<hid>/, every finding, the approved synthesis. Inventory the figures and the metrics.json addresses.
3. Agree the arc with the PI before drafting. Structured questions, each with a recommended default: single arc vs mini-arc per sub-question; redesign figures from metrics.json vs reuse; which 1–3 findings get a slide. Trivia (output path, framework) is stated, not asked.
4. Draft. Copy assets/deck.html to <vault>/presentations/<anchor-id>/index.html and fill the stubs. Study examples/q1_scaling_deck.html — it is the worked example, built on skills/crux/examples/scaling_vault q1, verify-green and lint-clean.
5. Check. crux deck --verify <deck> --strict and crux deck --lint <deck> until both are green. Then the manual pass: open from file://, walk every slide at 16:9, confirm nothing collides with the fixed footer and in-SVG text is room-legible.
6. Refine with the PI. Deck-building is a refinement loop by nature — "split that slide", "drop that model everywhere", "enlarge the figures" each must be a small local edit. That is why this is a skill, not a fire-and-forget agent.
7. Re-presenting later: crux deck --refresh <deck>, then re-read the prose around every changed number — refresh fixes values; only a human fixes the sentence that interprets them ("clears the floor by +0.05" can be silently falsified by a correct refresh). validate --check=decks reports stale decks from the vault side.

The addressing contract (non-negotiable)

Every number the vault owns carries both the cached value it renders and the address it came from:

{ src:'h1#task_a.delta', v: 3.3, lo: 1.9, hi: 4.7 }      // chart data

<span data-src="h1#task_a.delta">+3.3</span>              <!-- a numeral in prose -->
<span data-derived="h1#a.x,h1#b.y">7.9</span>             <!-- computed; inputs checked -->
<span data-src="literal">95%</span>                       <!-- definitional constant -->

• Addresses resolve as <hid>#<dotted.key.path> → results/<hid>/metrics.json. The PI or the run harness writes metrics.json; crux never computes it, and neither do you — if a number you need has no address, ask for it to be added, don't invent it.
• A chart annotation comparing two cached values (a bar gap, a delta) must be computed at render from those values, never hand-typed — a cached display string is invisible to --refresh and rots.
• Figures are inline SVG re-derived from metrics.json, not embedded matplotlib PNGs. A PNG that must appear is base64-embedded.
• The deck lives at <vault>/presentations/<anchor-id>/index.html. It is derived, not evidence: never link it under any ## Artifacts block.

When the only artifact is a 10-panel PNG (one-panel extraction)

One figure = one claim: contribute the single panel that tells the story, rebuilt on its own. Fallback ladder, in order:

1. First, re-run the plotting code when the report links it, extracting the one panel as SVG.
2. Rebuild from the underlying table or metrics.json with the template's chart scaffolds.
3. Last resort: embed the cropped panel base64 and mark it visibly as an unverified figure — never present an unaddressed chart as if it were checkable.

Content rules

• One claim per slide, stated in one sentence — a slide whose claim cannot be written in a sentence does not exist. Deck length is elastic; only the results section grows.
• Keep the negatives. An honest "this did not work, and here is the number" is often the most valuable slide. Never spun, never dropped.
• Dual register: a plain-language lead a non-specialist grasps in one read, backed by exact numbers a specialist can check.
• Quantify → interpret → caveat. Not "more data helps" but "+3.3 points, CI clear of zero — because the extra data lands where the model was thinnest — but at 1.4× the compute."
• Show uncertainty; a gap smaller than the SE is not a gap.
• Bullets are conclusions, not axis descriptions.
• Numbers by role (D5): motivation slides carry none; method slides carry addressed definitional constants only; result slides carry at least two addressed numbers; the verdict slide introduces none of its own.
• "Rules locked before results": surface the anchor's ## Protocol when it has one.
• No background, related-work, or future-work slide unless asked.

The slide contract

Every <section class="slide"> carries a comment header: job / source / numbers / cut (plus claim/shape where relevant). It is the override surface — the user's request wins over every default, and when you deviate you update the slide's contract header to match, so the deviation is a deliberate call rather than drift. crux deck --lint enforces the headers and the 7-content-unit budget; footer overlap and projector fit stay on your manual checklist.

Technical spec (already encoded in the template — do not regress it)

One self-contained HTML file: no CDN, no framework, no build step, opens by double-click from file:// and survives being emailed. Fade-only transitions with prefers-reduced-motion honored; keyboard nav (→ ← space Home End); thin progress bar; slide counter derived from the DOM; print stylesheet gives page-per-slide PDF via the browser. Color is semantic — one color per arm, defined once as CSS variables, reused identically across chart, legend and bullets. In-SVG fonts are sized for a projector (~1.5× laptop taste); chart height is capped in vh so a figure never crowds out its caption bullets.