Guide · Agent skill
How to use the ArtDirection agent skill: a quality loop for visuals
Updated · written with AI from the project's own documentation and checked against our catalogue data
ArtDirection is an agent skill that adds a quality-control loop to visual work such as images, films, slides and site captures. Instead of accepting the first draft, the agent writes a brief, has vision judges score the work against a written rubric, lists ranked defects to fix, and repeats up to four rounds. This guide explains what the skill does, how to install it, how to trigger it with plain requests and what it needs to run.
Art Direction Quality Loopdanielmiessler/LifeOS · MIT Download the skill Quick start
- 1Install the ZIP. Download the ZIP from this directory and unzip it into ~/.claude/skills/ for personal use, into .claude/skills/ in a project, or upload it under Settings → Capabilities → Skills in the Claude apps.
- 2Prepare the judge config. Create Config.md in your ArtDirection customizations folder and name the vision vendors in art_judges. Sign in to the codex CLI with your ChatGPT account and make sure the claude CLI has subscription access.
- 3Install the framework files and ffmpeg. The skill needs Bun and the LifeOS framework files. Install ffmpeg if you critique films. Follow Setup.md for the first-time checks.
- 4Write a brief. Describe the audience, register, references and bar. The brief schema is in References/Rubric.md, and References/BriefExample.json shows a filled example.
- 5Ask for a critique. Say something like rate the visuals or iterate until it's great. The agent runs Critique and reads the ranked defects it returns.
- 6Fix, re-check and deliver. The agent iterates up to four rounds. When the result passes, ask it to deliver the render. Deliver only hands over a passing artifact.
- 7Record your own score. Tell the agent your score and note, such as that one's a 6, the type is cramped. This appends a label you can use for later judge calibration.
What the ArtDirection skill makes the agent do
ArtDirection runs a quality loop for artistic deliverables. A brief defines the audience, register, references and quality bar. A direction council turns the visual intent into a Direction.md file. A separate domain skill then produces the artifact, and ArtDirection judges it.
The judging step uses two vendors' vision models. Each scores the work against anchored craft criteria and returns ranked, located defects that a producer can fix. The result is a Critique.json file with both vendors' scores, the defects and a pass, fail or unavailable verdict.
The skill does not create the artwork itself. The source is explicit that a domain skill must produce it, and that text writing audits and rating an ISA are outside its scope.
Install the ArtDirection skill
Download the ZIP from the skill's page on this directory and unzip it into one of these locations:
Option one is your personal skills folder at ~/.claude/skills/. Option two is a project folder at .claude/skills/ inside that project. Option three is to upload the ZIP in the Claude apps under Settings → Capabilities → Skills.
- Personal install: ~/.claude/skills/
- Project install: .claude/skills/ in your project
- Claude apps: Settings → Capabilities → Skills, then upload the ZIP
Set up the judges and the config file
The judge panel is set in a Config.md file inside your ArtDirection customizations folder. That folder is LIFEOS/USER/CUSTOMIZATIONS/SKILLS/ArtDirection/ under your config root, which is $CLAUDE_CONFIG_DIR when set, otherwise ~/.claude. Config.md names the vision vendors in art_judges. OpenAI, Anthropic and Google have seats. If the file or a required key is missing, the critique stops with an error that names both vendors.
OpenAI vision runs through the codex CLI, signed in to a ChatGPT account. Anthropic vision runs through the claude CLI with subscription access. Setup.md covers first-time account access, ffmpeg for films and a no-cost proof that the setup works.
- Config.md at the customizations folder names the judge vendors in art_judges
- codex CLI signed in to ChatGPT for OpenAI judging
- claude CLI with subscription access for Anthropic judging
- Never put credentials in briefs or logs
Trigger the skill with natural requests
You do not need special syntax. The skill's trigger phrases include critique this, art direction, quality loop, rate the visuals, is this pro, iterate until it's great, deliver the render and art council. Say what you want in plain words and the agent picks the matching intent.
Loop runs the full cycle until the work reaches the bar. Direction builds the visual direction. Critique scores one artifact. Deliver hands over a passing artifact. Calibrate records your own score.
- "Iterate until it's great" on a render starts the Loop intent
- "Rate the visuals" starts Critique and returns Critique.json
- "That one's a 6, the type is cramped" starts Calibrate and records a label with your note
- "Deliver the render" starts Deliver, which only hands over a passing result
bun skills/ArtDirection/Tools/Critique.ts --video <mp4> | --image <png> [--images <glob>] --brief <Brief.json> [--bar 8.5] [--samples 12] [--concurrency N] [--vendors openai,anthropic] [--sample-seed <n>] [--out <dir>] [--json] [--allow-identity]Read the verdict and the defects
A pass needs both vendor medians to reach the bar minus 0.5, the pooled median to reach the bar, and no UNKNOWN score or critical defect. If the loop reaches its four-round cap without passing, the report states the shortfall. A vendor that is unreachable is reported as unavailable, not as a quality failure.
A one-vendor panel can still pass, but its report carries the vendor_missing flag as weaker evidence. Figure anatomy problems such as extra limbs or a grip that does not match the object held are treated as critical defects.
- pass: both vendor medians and the pooled median meet the bar, with no critical defect
- fail: a score or defect blocks the pass; the report lists ranked, located fixes
- unavailable: a vendor could not be reached, so there is no quality verdict
Deliver and calibrate
Deliver copies the artifact to the directory you name, but only with a current passing critique bound to the exact bytes and a passing Proof receipt. For MP4 files, the adjacent review/ folder comes along. Deliver will not write anywhere except the intended hand-over location.
Calibrate appends your own verdict, an integer from 1 to 10, and your note to labels.jsonl. Judges are never changed automatically. Labels only inform a deliberate calibration later.
- Deliver: Tools/Deliver.ts --artifact <file> [--critique <Critique.json>] --to <dir>
- Calibrate: Tools/Calibrate.ts --artifact <file> --verdict <1-10> --note "<text>"
bun skills/ArtDirection/Tools/Deliver.ts --artifact <file> [--critique <Critique.json>] --to <dir>What people use it for
- Checking a batch of generated images against a fixed rubric before publishing them to a portfolio or blog.
- Getting a repeatable pass or fail check on slide decks or site captures for a team that reviews visual work.
- Adding a quality gate after another skill that generates images, films or slides, so the first draft is not shipped as-is.
Frequently asked questions
Does ArtDirection create images, films or slides?
No. The skill only judges and gates the work. The source says a separate domain skill for images, motion, story films, video essays, web design or slides must produce the artifact.
What do I need to run ArtDirection?
You need the Bun runtime, the codex CLI signed in to a ChatGPT account for OpenAI judging, the claude CLI with subscription access for Anthropic judging, and a Config.md file that names your judge vendors. ffmpeg is needed for film critiques, and the LifeOS framework files are required.
Is ArtDirection free and what is its licence?
The source lists the licence as MIT. It does not state a price. The judge vendors require their own accounts or subscriptions, as described in the requirements.
Can I use ArtDirection without an OpenAI or Anthropic account?
The panel is built from the vendors named in your Config.md. OpenAI vision needs the codex CLI signed in to ChatGPT, and Anthropic vision needs the claude CLI with subscription access. The source says a missing file or key stops the critique with an error.
Does ArtDirection run scripts and use the network?
Yes. The source says it runs scripts and uses the network, because the vision judges are called through those CLIs. Review the scripts before installing if that matters for your setup, and never put credentials in briefs or logs.
How is ArtDirection different from just asking an AI to critique my image?
It runs two vendors against a fixed, written rubric, returns ranked and located defects, repeats up to four rounds, and only hands over work that passes. A single free-form opinion has none of these checks. The source also notes that scores are not absolute truth, so your own eye remains the final check.
What if my judges disagree or the result looks wrong?
Record your own score with Calibrate. The label is kept verbatim and judges only change through a deliberate calibration, not automatically.
New agent skills, repos and AI tools — every week
The best new skills and open-source projects, the fastest-growing repos, and one free prompt. Five minutes to read. Unsubscribe anytime.
Our verdict
ArtDirection suits creators and teams who want a repeatable second opinion and a pass or fail gate on images, films, slides and site captures, and who already have a generation skill. It is not a generator and does not replace your own judgement. If you need a tool that creates the artwork, or a text writing audit, pick a different skill.
Download the skill