Infographic (alias /visual-proof)

Transforme un extrait d'étude en un visuel mémorable — poster, slide, thread hook. Aucune recherche, aucun WebSearch.

Also invoked as /visual-proof — same skill, same outputs.

When to use (vs other skills)

Optional pipeline (never auto-chained):

/bear-hours → notes/dense-read → /infographic (optional shareable visual)
/deep-research → /infographic (hook) OU /layout-html (rapport complet)

Hard rules

1. No research — never WebSearch / WebFetch. Work only from provided text/files.
2. One visual — single PNG or single prompt export per run (iterations → visual-proof.v2.png).
3. Faithful to source — never invent sources, numbers, quotes, or claims absent from input.
4. Spec always — write visual-proof.spec.md before generate or prompt export.
5. Privacy strip — apply references/privacy-strip.md before spec finalization.
6. Language — on-image text matches source (lang in spec).
7. ≤7 labels on canvas — image models garble text; if more needed → suggest /layout-html.
8. Style first — ask style via AskUserQuestion before reading content (unless smart-skip).
9. No silent brief edits — mention output paths in chat; do not append to brief.md without consent.
10. Metaphor + hero — every spec names one concrete visual metaphor (a drawable scene, never "boxes and lines") and one hero element that dominates the canvas. See Step 2.

Step 0: Style — AskUserQuestion (always first)

Use AskUserQuestion — one call, wait for answer before continuing.

Title: Infographic — Style

Hard rules:

• Never list style options as plain chat bullets — always AskUserQuestion
• STOP after the call; wait for answer
• Smart-skip — if user already named a preset or described custom style (≥5 words), skip (see references/palettes.md)

Question: "Quel style visuel pour cette infographie ?"

Options (exactly 4 — AskUserQuestion max; custom style goes through the automatic "Other" option):

If user answers via "Other": treat the free text as a custom style. Map to palette per references/palettes.md. One short follow-up only if too vague.

If preset selected: load palette from layout-html/references/presets/{{PRESET_ID}}.md via references/palettes.md.

Default when skipped: warm.

Step 1: Intake — detect input mode

Resolve input in this priority order:

Slug derivation (mode D): lowercase-hyphen from title or first heading; ask once if ambiguous.

Guards:

• No resolvable text → ask for slug, path, or paste
• Multiple substantive files in study → AskUserQuestion "Which file is the source of truth?"
• Source < ~150 words → warn; still allow card type
• If brief.md exists → read ## Narrow wedge; reject visual elements outside wedge (one question if excerpt spans topics)

Load references/visual-types.md, references/spec-schema.md, references/palettes.md, references/privacy-strip.md before extracting.

Step 2: Extract visual kernel (no invention)

From source, produce fields for visual-proof.spec.md:

Metaphor quality bar: if you can't picture the scene in one sentence without the words "box", "diagram", "chart" or "section", it's not a metaphor — pick a composition variant from references/visual-types.md and make it concrete with the source's own imagery.

Apply references/privacy-strip.md before spec is finalized.

For long reports: use executive summary / key findings section only — never the whole report.

Step 3: Select visual type

Auto-score per references/visual-types.md. Types: schema, timeline, comparison, flow, card (fallback).

If top two scores within tie threshold → AskUserQuestion:

Title: Infographic — Layout

Options (max 4): Schema | Timeline | Comparison | Flow — put the auto-scored winner first, labeled (recommandé); picking it = agent's choice

User override always wins. Record type_rationale in spec.

Escape: If source needs >7 precise labels or exact numeric data on canvas → warn user; suggest /layout-html for SVG reliability. Proceed with simplified visual or PROMPT only if user confirms.

Step 4: Build spec — always write

Write {output_dir}/visual-proof.spec.md per references/spec-schema.md.

This file is the audit trail and source for both IMAGE and PROMPT modes.

Step 5: Generate or export

Load references/tool-detection.md. Determine MODE before acting.

Step 5a: IMAGE mode

When: image_gen or GenerateImage available in this session (unless user asked prompt-only).

Before generate: Apply the labeled-infographic verification loop from references/tool-detection.md.

1. Compose prompt from spec + references/prompt-template.md § IMAGE mode
2. If using a prompt-style image tool: pass the composed prompt and aspect ratio (16:9 default; 1:1 if square/Instagram; 9:16 if story)
3. If using a description + filename image tool: write visual-proof.png in the output dir
4. Verify (fidelity loop, see references/tool-detection.md): read generated image back — check text legible and label count/palette/hero match the spec; one targeted retry, then fallback PROMPT
5. Write {output_dir}/visual-proof.png
6. Write {output_dir}/visual-proof.alt.md — 2–3 sentence alt text + caption for posts

Iterate: image_edit on existing PNG → visual-proof.v2.png

open "{output_dir}/visual-proof.png"   # macOS, IMAGE mode

Step 5b: PROMPT mode

When: No image tool, user requested prompt-only, or IMAGE failed.

1. Fill references/prompt-template.md → write {output_dir}/infographic-prompt.md
2. Include blocks: Universal, Midjourney, DALL·E, Flux/SD
3. Print Universal block in chat for immediate copy-paste

Step 6: Deliver

Tell the user:

1. Mode used (IMAGE | PROMPT)
2. Visual type + preset
3. Output paths (spec always; png or prompt)
4. Source file + wedge (if any)
5. Reminder: spec = source of truth; image = illustrative proof, not new research
6. Suggest pairing: /layout-html for full article — do not create proof.md unless user asks

Do not silently edit brief.md. Optional on consent: add ## Visual proof section with paths and date.

Escape hatches

Self-check before delivery

• No WebSearch / WebFetch
• All labels/numbers traceable to input
• visual-proof.spec.md written with type_rationale
• Spec has a concrete metaphor (drawable scene) + hero element
• Privacy strip applied
• ≤7 on-image labels, 1–2 words each (or user warned + confirmed)
• IMAGE: PNG + alt.md OR PROMPT: infographic-prompt.md with 4 blocks
• Universal block printed in chat (PROMPT mode)
• Clear differentiation from /layout-html in delivery message
• IMAGE mode: browser opened (open on macOS)