AutoDesign Agent Skills

August 19, 2026 · View on GitHub

English · 简体中文

AutoDesign Agent Skills

Turn one research paper into an editable poster, slide deck, project webpage, or narrated conference video directly inside your Coding Agent. These are four independently installable, self-contained Agent Skills:

SkillCreatesPrimary delivery
autodesign-posterSource-grounded academic conference posterEditable HTML, one-page PDF, preview, provenance
autodesign-pptResearch talk and slide deckEditable HTML, native editable PPTX, PDF, notes
autodesign-webpageResponsive research project webpageEditable local HTML, assets, desktop/mobile QA
autodesign-videoNarrated and subtitled conference videoEditable HyperFrames project, 1080p MP4, audio, SRT/VTT

Each Skill carries a lightweight local harness for evidence, attempts, deterministic checks, rendering, and delivery. No AutoDesign application server is required. Run state and generated artifacts stay in an output directory selected by the user. Managed runtimes stay in a versioned user cache; the workflow never writes generated output into the installed Skill.

Important

The Skill edition is the portable, convenient way to use part of AutoDesign from an existing Coding Agent. It does not replace the full AutoDesign Harness. Read Skills vs. the full AutoDesign Harness before evaluating output quality.

Agent Skills v0.2.0 · 2026-08-19

v0.2.0 is the first feature release after the initial standalone Skills bundle. It packages the current Poster, PPT, Webpage, and Video Skills as four independently installable, checksum-verified archives.

Highlights:

  • Poster Agent-first PDF ingestion. The host Agent now works from complete PDF page renders, creates reviewed source crops, and can return to the paper when a visual is incomplete or unsuitable.
  • Revision-bound Poster work. Source catalogs, plans, attempts, review routes, and final delivery are hash-bound and recoverable. Earlier attempts remain immutable, so a repair does not silently rewrite prior evidence.
  • Read-only Poster DOM QA. Screen and print probes report layout defects without modifying the Agent-authored poster.
  • White primary canvas across all four Skills. Poster, PPT, Webpage, and Video now require an opaque white primary canvas with no background image or paint effect that changes the rendered white. This keeps the paper's usual white-background figures visually coherent. Restrained light cards, panels, controls, captions, and overlays remain allowed as local surfaces.
  • Portable release integrity. Deterministic packages, SHA-256 sidecars, install receipts, read-only installed-tree checks, and standalone validators keep the release independently installable and auditable.

Download the agent-skills-v0.2.0 release. The original agent-skills-v0.1.0 release remains available as a historical snapshot.

Poster Agent-first v2 · 2026-08-18

Poster is the first portable Skill to receive the Agent-first source workflow. The paper PDF remains the source of truth: the host Coding Agent decides what matters, while local tools provide deterministic inspection, cropping, provenance, validation, and recovery. No paper-specific figure list or acceptance label is built into the Skill.

What improved:

  • Direct PDF curation. The Agent can inspect immutable page rasters and request exact pixel crops instead of being limited to one automatically extracted visual catalog. Fragmented diagrams can be re-ingested as a clean, single source visual.
  • Reviewed source catalogs. Canonical selection and review records bind every approved visual to the source page, crop, hash, evidence role, and catalog revision before an authoring attempt can use it.
  • Revision-bound plans and attempts. Each attempt snapshots its source, catalog, plan, and authorized assets. Failed reviews route to the right level: retry layout, create a new content plan, or return to the PDF for replacement crops. Earlier attempts remain immutable and inspectable.
  • Read-only screen and print QA. One browser probe checks both media modes for 13 stable failure types, including clipping, overflow, undersized text, table escape, broken source-flow structure, and underfilled canvases. It reports findings without rewriting the authored HTML, CSS, text, or layout.
  • Stronger portable provenance. Atomic recovery, append-only revision records, archive and release receipts, read-only installation checks, and zero-bytecode operation make runs easier to resume and audit across hosts.

These changes address the main quality bottlenecks observed in the first Skill release: weak important-figure selection, fragmented PDF extraction, repairs at the wrong workflow level, and layout checks that could not distinguish screen from print. The merged code passed a 481-test integration matrix, all 53 product smoke checks, the four standalone Skill validators, and real Chromium Poster screen/print regression suites.

This is a Poster-first update. PPT, Webpage, and Video remain installable and retain their existing generation workflows. The 70–80% Harness-equivalence goal remains a roadmap target, not a measured claim. Poster Agent-first v2 is included in the downloadable agent-skills-v0.2.0 release.

Quick install

Download the checksum-verified release bundle with GitHub CLI:

mkdir -p autodesign-skills-v0.2.0
gh release download agent-skills-v0.2.0 \
  --repo Yaxin9Luo/AutoDesign \
  --dir autodesign-skills-v0.2.0
cd autodesign-skills-v0.2.0

Then install all four Skills into the shared Codex and DeepSeek Harness user directory:

DESTINATION="${DESTINATION:-$HOME/.agents/skills}"

for skill in autodesign-poster autodesign-ppt autodesign-webpage autodesign-video; do
  python3 -I ./package_agent_skills.py install \
    --archive "./${skill}-0.2.0.zip" \
    --checksum "./${skill}-0.2.0.zip.sha256" \
    --destination "$DESTINATION"
done

Install only the artifact types you need by running the command for one Skill. The installer checks SHA-256, validates the extracted package, promotes it atomically, and refuses to overwrite an existing installation.

Install by Coding Agent

Agent Skills are directories containing a required SKILL.md plus optional scripts, references, and assets. AutoDesign follows the open Agent Skills specification.

Codex

Codex discovers user Skills under ~/.agents/skills. Older local setups may also use ~/.codex/skills.

DESTINATION="$HOME/.agents/skills"

Use this value for DESTINATION in the Quick install shell. After installation, start a new Codex task if the Skills do not appear immediately. See the official Codex Skills documentation.

Claude Code

Claude Code discovers personal Skills under ~/.claude/skills:

DESTINATION="$HOME/.claude/skills"

Set this in the same shell before running the Quick install loop; the loop preserves a destination you already selected. Claude Code normally detects changes in an existing Skills directory. Restart it if this is the first Skill directory on the machine. See the official Claude Code Skills documentation.

DeepSeek Harness

DeepSeek Harness supports project roots .dsh/skills and .agents/skills, plus the user roots ~/.dsh/skills and ~/.agents/skills. For a DSH-only user installation, use:

DESTINATION="$HOME/.dsh/skills"

Set this in the same shell before running the Quick install loop; the loop preserves a destination you already selected. Use DESTINATION="$HOME/.agents/skills" instead when you want one shared installation for Codex and DeepSeek Harness. See the official DeepSeek Harness Skills subsystem.

Installer details and Windows

The agent-skills-v0.2.0 release contains four ZIPs, four SHA-256 sidecars, a manifest, and a standalone installer. Download it with GitHub CLI:

mkdir -p autodesign-skills-v0.2.0
gh release download agent-skills-v0.2.0 \
  --repo Yaxin9Luo/AutoDesign \
  --dir autodesign-skills-v0.2.0
cd autodesign-skills-v0.2.0

Choose a Python 3 launcher: python3 on most macOS/Linux systems, python as the fallback, or py -3 on Windows. Then install one archive and its matching checksum:

python3 -I ./package_agent_skills.py install \
  --archive ./autodesign-poster-0.2.0.zip \
  --checksum ./autodesign-poster-0.2.0.zip.sha256 \
  --destination "$HOME/.agents/skills"

Replace autodesign-poster with another Skill name and choose the destination for your Coding Agent. The installer verifies SHA-256, validates extracted contents, promotes atomically, and refuses to overwrite an existing Skill.

Windows example for all four Skills (change $Destination for Codex or DeepSeek Harness):

New-Item -ItemType Directory -Force autodesign-skills-v0.2.0 | Out-Null
gh release download agent-skills-v0.2.0 `
  --repo Yaxin9Luo/AutoDesign `
  --dir autodesign-skills-v0.2.0
Set-Location autodesign-skills-v0.2.0

$Destination = "$HOME\.claude\skills"
foreach ($Skill in @("autodesign-poster", "autodesign-ppt", "autodesign-webpage", "autodesign-video")) {
  py -3 -I .\package_agent_skills.py install `
    --archive ".\$Skill-0.2.0.zip" `
    --checksum ".\$Skill-0.2.0.zip.sha256" `
    --destination $Destination
}

First run

Give the Coding Agent the paper PDF, name the Skill, and state the intended audience and output directory. For example:

HostExample
Codex$autodesign-poster Turn /path/paper.pdf into an editable CVPR landscape poster. Save the complete run under /path/output.
Claude Code/autodesign-ppt Turn /path/paper.pdf into an 18-slide conference talk for a technical audience.
DeepSeek HarnessUse autodesign-webpage to turn /path/paper.pdf into a responsive research project page under /path/output.
Any supported hostUse autodesign-video to create a six-minute narrated conference video from /path/paper.pdf with optional English subtitles.

The Coding Agent reads the Skill workflow, runs its local doctor, prepares the source evidence, authors bounded attempts, executes deterministic QA, and requests a fresh visual review before final delivery. The first browser, PPT, or Video run may download pinned runtime dependencies into a user cache.

Generated deliverables live under the run directory you requested. The workflow does not write run state, caches, or artifacts into the installed Skill.

Requirements

All four harnesses require Python 3.10 or newer. Video currently requires Python 3.10–3.12 for its locked Kokoro and HyperFrames runtime.

SkillAdditional local requirements
PosterPoppler: pdftotext, pdfinfo, pdftoppm, pdfimages; first setup installs a pinned browser in the user cache
PPTPoppler, LibreOffice; first setup installs pinned browser and native-PPT runtimes in the user cache
WebpagePoppler; first setup installs a pinned browser in the user cache
VideoPoppler, Node.js 22+, npm, ffmpeg, ffprobe, Python 3.10–3.12; setup installs exact hyperframes@0.7.86 and locked Kokoro assets

The setup tools select supported macOS, Linux, or Windows packages and fail closed on unsupported platforms or architectures. Initial runtime setup requires network access. Run the Skill through your Coding Agent so it can act on doctor diagnostics rather than skipping a required gate.

Skills vs. the full AutoDesign Harness

The four Skills preserve many of AutoDesign's artifact contracts and local QA checks, but they run inside a host Coding Agent rather than the complete AutoDesign system.

AreaStandalone SkillsFull AutoDesign Harness
Entry pointExisting Coding Agent conversationAutoDesign Workbench and pipeline
OrchestrationHost agent follows one portable workflowCoordinated artifact pipeline, shared context, routing, and lifecycle management
Source understandingHost agent plus local evidence and provenance toolsIntegrated paper memory, planners, artifact-specific harnesses, and shared source state
Review and iterationDeterministic gates plus fresh host-agent or subagent reviewIntegrated critics, retry routing, cross-stage diagnostics, and workbench feedback
Editing experienceFiles and the host Coding AgentWorkbench, previews, attempt history, Canvas editing, and artifact export
Output consistencyMore dependent on the selected agent, model, and local environmentMore controlled by the full optimized pipeline

The standalone Skill path is intentionally easier to install, but it does not replace AutoDesign's full effect or guarantee the same artifact quality. In particular, PDF interpretation, important-figure selection, visual hierarchy, and multi-round repair still depend more heavily on the host Coding Agent.

Our future target is to bring these portable Skills to roughly 70–80% of the full Harness experience across supported artifact workflows. That is a roadmap target, not a claim about current measured performance.

For the complete system, use the AutoDesign Workbench and the full DesignHarness pipeline.

Roadmap and contributing

The Skills are under active development. Contributions are welcome in:

  • PDF understanding, figure ranking, crop selection, and source grounding;
  • artifact planning, context management, and repair routing;
  • layouts and editable export for Poster, PPT, Webpage, and Video;
  • deterministic evaluators, visual review rubrics, and regression cases;
  • real-paper examples that expose gaps between the Skills and full Harness.

Open an issue with a reproducible paper/run case, or submit a pull request with focused tests. Contributing to the portable Skills is also a direct way to help move them toward the 70–80% future target without weakening their source-grounding and delivery contracts.

Maintainer verification

Validate all four packages:

python3 -B scripts/validate_agent_skills.py --root agent_skills

Build deterministic, non-overwriting release archives:

python3 -B scripts/package_agent_skills.py build \
  --source-root agent_skills \
  --output-dir dist/agent-skills-v0.2.0 \
  --version 0.2.0

The output directory contains one versioned ZIP and SHA-256 sidecar per Skill, manifest.json, and the release-local package_agent_skills.py and validate_agent_skills.py tools. Build twice and compare every byte before publishing a release.