Aiconographer — coding-agent edition

Create six distinct article-icon candidates, assess them at their real UI sizes, and retain one transparent two-fill SVG as the canonical asset.

This skill is the coding-agent edition of Aiconographer. Use the host agent's native image generation and reasoning capabilities under the user's coding subscription. It must not require an API key or invoke the Python package merely to perform an ordinary skill run.

The separate Python package is the API edition: it exposes a module and pipx CLI and uses Chatsnack for model calls. Provider plumbing intentionally differs. Keep the editions aligned on their shared workflow and document any behavioral differences rather than assuming identical output.

Choose the path

• Full article run: Read references/production-guide.md completely, then plan, generate, inspect, compile, blind-judge, and finalize.
• Compile an existing candidate sheet: Read the compile and provided-sheet sections of the production guide. Run scripts/check-sheet-padding.mjs --help and scripts/process-sheet.mjs --help before compiling.
• Judge existing compiled candidates: Read the blind-judging section of the production guide. Use only the article text and the compiler's anonymous 192px/48px review pack.
• Finalize an already selected candidate: Run scripts/finalize-winner.mjs --help.

Shared behavior contract

• When generating a sheet, plan exactly six materially different concepts.
• Generate one landscape 3-by-2 candidate sheet on pure white.
• Allow at most three generated-sheet attempts. Preserve every attempt and feed concrete rejection reasons into the next one.
• Use charcoal #2C2C2B and coral #F45138 as the only final SVG fills.
• Keep final SVG backgrounds transparent; white is a temporary raster background.
• Validate tile padding deterministically before model inspection or compilation.
• Keep each icon roughly centered. The compiler shifts noticeably displaced artwork as a whole; small offsets and the design's internal composition remain unchanged.
• Trace at source resolution and judge direct SVG renders at 192px and 48px.
• Give the judge anonymous previews without candidate IDs, concepts, positions, prompts, or creator preference. A same-agent judging phase is not independent or fully blind if it saw the sheet.
• Produce a parsed scorecard, exclusive winner finalization, and a retained run record.
• Support a user-provided 3-by-2 sheet by skipping planning and generation while preserving deterministic validation, blind judging, and provenance.
• Derive the final slug from the article title: lowercase ASCII, replace non-alphanumeric runs with hyphens, trim hyphens, and fall back to source-icon. Never overwrite an existing result.

Generated-sheet run

1. Extract the source text's primary identity, defining relationship, essential structure, emotional register, and useful motifs.
2. Write six genuinely different one-sentence icon concepts.
3. Generate a 1536×1024 candidate sheet with the host's native image tool. Request six square cells, generous padding, the exact palette, and no dividers or forbidden rendering effects.
4. Save the sheet as candidate-sheet-attempt-N.png. Run the deterministic padding check, then inspect the whole sheet visually. If rejected, retain the reasons and retry, up to three total attempts.
5. Compile the first accepted sheet with process-sheet.mjs. Treat compiler rejection like an inspection rejection when attempts remain.
6. Use a fresh judge context when available. Give it only the exact source text and paired anonymous 192px/48px previews.
7. Save the structured result as aiconographer-scorecard.json, verify its aliases and weighted arithmetic, and map the winner through review-map.json. Stop before finalization when the scorecard is structurally invalid.
8. Finalize exactly one candidate with finalize-winner.mjs.
9. Retain the source, concept brief, all attempts, inspections, accepted sheet, prompt, compiler outputs, scorecard, selection, provider label, tool versions, and checksums in run-record.json.

For a provided sheet, follow the guide's provided-sheet path instead of planning or generating.

Deterministic compiler setup

The bundled scripts use pinned Node dependencies from a caller-supplied directory so node_modules stays outside the skill. Stage them in a temporary directory:

$skillRoot = '<absolute-path-to-this-skill>'
$depsRoot = Join-Path ([System.IO.Path]::GetTempPath()) 'aiconographer-node-deps'
$npm = (Get-Command npm.cmd -ErrorAction Stop).Source
$node = (Get-Command node.exe -ErrorAction Stop).Source
New-Item -ItemType Directory -Path $depsRoot -Force | Out-Null
Copy-Item -LiteralPath "$skillRoot\scripts\package.json" -Destination $depsRoot
Copy-Item -LiteralPath "$skillRoot\scripts\package-lock.json" -Destination $depsRoot
& $npm ci --prefix $depsRoot

Check the four-percent safety band before compilation:

& $node "$skillRoot\scripts\check-sheet-padding.mjs" `
  --input '<candidate-sheet.png>' `
  --deps-root $depsRoot

Compile into a new or empty directory:

& $node "$skillRoot\scripts\process-sheet.mjs" `
  --input '<candidate-sheet.png>' `
  --output '<new-empty-run-directory>' `
  --deps-root $depsRoot `
  --article '<source-text-file>'

After judging, finalize without overwriting prior output:

& $node "$skillRoot\scripts\finalize-winner.mjs" `
  --run '<run-directory>' `
  --candidate c6 `
  --slug '<article-slug>'

Use equivalent node and npm commands on non-Windows systems. Do not add these dependencies to the host project merely to process a sheet.

Completion

Inspect the white and dark contact sheets and the winning 48px render before reporting success. Provide the canonical SVG, preview, scorecard, run record, exact generation prompt, and candidate provider. For a provided sheet, report the generation prompt as null. State that this was the coding-agent edition so the provenance is unambiguous.