AutoDesign Poster

Create the Poster with the host coding Agent. The Agent reads the paper, chooses evidence, authors HTML/CSS, and makes every semantic repair. Bundled scripts render, crop, hash, version, measure, validate, and package the result.

Resolve SKILL_ROOT to this installed Skill folder. Put run state, browser cache, and generated files in a user-selected output directory. Never write generated artifacts, caches, or run state inside this Skill directory.

Resolve <PY> to the first available launcher in this order: python3, python, then Windows py -3. Substitute the selected launcher verbatim; do not assume it is one shell token. Use <PY> "$SKILL_ROOT/scripts/poster_harness.py" --help for exact flags.

Read at the point of use

• Before source selection, read agent-first-source.md. It is the visual evidence authority for Poster v2.
• Before authoring or validation, read output-contract.md.
• Before either fresh review, read review-rubric.md.

Invariants

• PDF and complete page renders are the primary semantic surface. pdfimages is discovery-only and never authorizes evidence.
• The host Agent or a fresh vision-capable subagent owns semantic selection. There is no mandatory image-count quota.
• Scripts never edit the Poster. In particular, DOM audit reports defects but never changes authored HTML, CSS, layout, or content.
• Render the primary .paper-poster canvas as opaque pure white in both screen and print media so white-background paper crops remain visually integrated. Do not use a canvas gradient, image, tint, dark theme, transparent root, or root/ancestor paint effect that changes white; restrained light fills remain available inside local sections and panels.
• Fresh source review must pass before plan. The artifact reviewer may escalate a repair route but never downgrade it.

Workflow

Run every command as <PY> "$SKILL_ROOT/scripts/poster_harness.py" COMMAND .... Each successful command prints JSON; retain its IDs, hashes, relative paths, active revisions, and next_action.

1. Run doctor with an external --cache-root. If Chromium is absent and the user permits the download, rerun it with --install-browser.
2. Run init --run-dir "$RUN" --source "$PAPER". Add only optional style references with repeated --reference; they can never become paper evidence.
3. Run inspect-source --run-dir "$RUN". Read the returned immutable PDF directly and inspect every complete returned page when direct PDF vision is unavailable. Query grounded text with evidence as needed.
4. For each important complete region, write the canonical request from the source guide and run crop-source. Inspect every crop at useful zoom, then run list-source-assets; extraction hints remain untrusted.
5. Write the exact selection object, run source-review-context, give its bound previews and source pages to a fresh reviewer, then run record-source-review. Repair and repeat until a passing review commits an immutable catalog revision.
6. Write the exact revision-bound plan from the source guide and run plan. Use only reviewed assets, permitted roles, and catalog reuse limits.
7. Run begin-attempt. Read the returned attempt ID, revision IDs, staged assets, and authoring-context.json. Author its requested poster.html and claim/source map yourself. Preserve each source visual and its evidence-bound native readout as direct siblings in one .source-flow-unit. Never guess an attempt number.
8. Run the strictly read-only dom-audit --attempt "$ATTEMPT"; repair the authored Poster yourself if it reports findings.
9. Run validate --attempt "$ATTEMPT" --source-map "$SOURCE_MAP". Continue only after all HTML, DOM, local-dependency, preview, and one-page PDF gates pass.
10. Run review-context, then give every bound screen/PDF frame, source map, evidence, and rubric to a fresh Agent that did not author the attempt. Run record-review with its complete canonical review.
11. On failure, follow the recorded route. A layout repair starts a new begin-attempt; content or source repair first uses reopen-curation, then commits the required new plan or source-review/catalog and plan revisions.
12. On pass, run finalize --attempt "$ATTEMPT" and deliver the complete verified final/ closure, not selected files from it.

After interruption, run resume --run-dir "$RUN", take the active attempt and revision IDs from its JSON, and perform only its named safe next action:

• author means run begin-attempt and author the attempt it returns. This starts the first attempt, starts the next layout-repair attempt, or returns the active authoring attempt as appropriate.
• retry_current_attempt means run begin-attempt to recover the same active attempt without consuming a new semantic attempt.
• reopen_curation means submit the bound content/source reopen request before any new plan, source review, or attempt.
• Other actions name the command/stage to run using the active IDs in JSON.

A blocked, stale, failed, or unreviewed run is not a finished Poster.

Compatibility and quality boundary

diagnose-v1 is read-only. It reports legacy metadata but does not load, execute, migrate, or continue the old snapshot; continue v1 only with its exact prior package, or start a new v2 run.

This portable Skill does not replace the full AutoDesign Harness. The Harness has stronger orchestration, assistance, and quality-control loops. Use this Skill for convenient standalone Agent workflows while continuing to improve it toward Harness quality.