Subagent Instructions

Use this file only when the main agent has explicitly assigned you a bounded
slide or asset task. The main agent owns deck coherence. Your job is to rebuild
the approved reference image into one clean editable artifact-tool presentation JSX slide.

Do the following loop exactly:

1. Read the assigned deck brief, source-style audit or style-direction file,
   design-system lock, approved reference image, reference QA notes, source
   notes, icon manifest, and data files before editing.
2. Create one ESM slide module that exports a single artifact-tool presentation JSX
   slide function for your assigned slide.
3. Build a temporary one-slide artifact-tool deck by calling only your
   exported function.
4. Render that slide to PNG and layout JSON with the provided absolute path to
   `render_artifact_slide.mjs`. Always pass `--layout`.
5. Run `node <check_layout_quality.mjs> --layout <your-layout-json>` after every
   render.
6. Compare the rendered PNG to the approved reference image at readable size and
   note a prioritized list of issues, including visual mismatches and any
   layout-quality failures.
7. Repeat from step 2 to fix issues. Repeat this process at least 3 times until
   the slide matches the reference image as well as practical and
   `check_layout_quality.mjs` has no errors. If exact matching would be worse
   because of factual, legibility, or editability constraints, get as close as
   possible, make it look clean, and document the deviation.
8. If `check_layout_quality.mjs` reports an error, fix the slide. Do not return
   a slide with checker errors unless the main agent explicitly provided an
   allowlist and written reason.
9. Warnings from `check_layout_quality.mjs` must be inspected visually. Fix
   warnings involving title/hero gutters, KPI text, split inline text, tight
   text boxes, or text/image collisions unless there is a clear reason not to.
10. Return only your slide module and any cropped/reused assets. Do not return a
   standalone deck as the final artifact.

Reference fidelity rules:

1. Use the reference image as a template/layout comp, not loose inspiration.
   Preserve composition, hierarchy, visual rhythm, spacing, color balance, scale,
   and mood where practical.
2. Do not just use the reference image as a background image for the whole slide.
3. Do not replace a strong reference composition with a generic shared layout.
   Shared components are for repeated deck chrome, not for flattening slide
   bodies.
4. Fix imagegen artifacts: garbled text, fake UI, distorted icons, rough labels,
   bad contrast, decorative clutter, unsupported metrics, and any
   do-not-rebuild elements called out by the main agent.
5. Keep slide copy concise. Do not add paragraphs or extra explanatory text just
   because space exists.

Asset and icon rules:

1. Use vetted brand assets provided by the main agent whenever possible.
2. For generic icons in the reference image, first look for the closest matching
   Lucide icon and use it, preferably with
   `ctx.addLucideIcon(slide, { name: "IconName", left, top, width, height,
   color, strokeWidth })`. You may also use pre-rendered Lucide assets supplied
   by the main agent.
3. Do not manually draw icons with lots of shapes, lines, or SVG paths when a
   Lucide icon is close enough.
4. For photos, complex decorative visuals, device mockups, abstract hero art,
   non-Lucideable diagrams, shadows/glows, and complex visual effects from the
   reference comp, crop/reuse the visual asset when native reconstruction would
   be slower or visibly worse.
5. For logos and complex brand marks, use high-quality official assets or clean
   source-provided crops. Do not use rough, fuzzy, distorted, or off-brand
   crops. Do not use imagegen-cropped logos as final assets.
6. Do not use imagegen crops for factual text, chart labels, source notes,
   metrics, or data-bearing UI.
7. Record final icons, logos, and repeated asset marks in the icon manifest with
   source type, icon/library name or asset path, color, size, slide usage,
   approval status, and notes.

Hard layout rules:

1. Declare layout zones before drawing. Use constants such as `SAFE`, `HEADER`,
   `FOOTER`, `TEXT_ZONE`, `VISUAL_ZONE`, `CHART_ZONE`, and `KPI_ZONE`. Keep all
   major text, KPI cards, charts, and hero visuals inside their assigned zones.
2. No major text box may overlap a hero image, diagram, chart, table, or visual
   zone unless the prompt explicitly asks for text-over-image.
3. Leave at least 24px of gutter between major text blocks and hero
   images/diagrams/charts. Do not let bounding boxes merely look close enough;
   the layout JSON must show separation.
4. Do not split a single sentence across multiple text boxes just to style one
   phrase. Use one text box, separate lines, or a standalone emphasized phrase.
5. KPI cards must reserve separate vertical slots for label, value, suffix, and
   note. Leave enough room for font fallback.
6. Treat font fallback as normal. Every text box needs slack: at least 10-20%
   extra width and height beyond the expected text.
7. Cropped images must be placed in aspect-ratio-compatible frames unless
   intentionally masked.
8. Do not invent quantitative claims. Every number must be sourced from the
   prompt/source notes, derived transparently, or omitted.

Your module must look like this: