Repository navigation
Releases
unaverage design 3.4.4
Fixes for unaverage design, with the skill at 3.4.4. The design rules are unchanged; the effect of the fixes on builds is unmeasured.
What changed
npm auditis clean. The lockfile now pinspuppeteer-core25.17.0, which bringszod4.6.7 instead of 3.25.76 and closes the three "high" reports onzod. The scripts render, export and measure exactly as before.- Quick asks before it builds when a tool is missing. At Quick the skill now names the missing tool with the exact command in its first message, counts that as one of its two questions and waits for your answer.
npm installruns only after your OK. - Findings from the Windows test of 16 November 2026: the generic check warns per surface when a surface has no text to measure and when a selector matches nothing, and the skill now tells the agent to render with its own script at Quick, to keep the direction log in the template's structure and to build posts and ads as HTML or SVG instead of drawing them with Python (probed in Claude Code and in one Codex run).
- The NOTICE says where the skill runs, and the README names the ZIP as
unaverage-design-<version>.zip.
Install or update
npx skills add jakob-meyer/unaverage --skill unaverage-design -g -a claude-code -a codexAs a Claude Code plugin: /plugin marketplace add jakob-meyer/unaverage, then /plugin install unaverage-design@unaverage. To update an existing install, see "Update and remove" in the README.
By hand or in the Claude app: download unaverage-design-3.4.4.zip below, as described in the README.
Tests
Checked in pair runs of the scripts against 3.4.3, in fixtures, in a probe of five Quick runs in Claude Code and in fresh installs on macOS. Not measured in a blind test. Method, results and limits: docs/testing.md, section 2.15.
Files
unaverage-design-3.4.4.zip: the skill folder of this release. SHA-256d71c4de5f467490795420ec114247bee5635563b6198b2e9220752b4ab22e0d1.
Apache License 2.0. All changes: CHANGELOG.md.
Assets 3
unaverage design 3.4.3
First public release of unaverage design, a design skill for Claude, Claude Code and Codex. It aims for studio-level results instead of the generic AI look, across websites, product UI, decks, social posts, ads and brand.
What it does
- When the direction is open, it first asks about things from your world: what your customers hold or look at, what you love the look of, what you can no longer stand to see in your field.
- You agree on the effort first (Quick, Standard or Studio), then it shows three ranked directions as mockups.
- It builds in HTML, CSS and SVG at the real size of the medium, measures contrast, overflow and font sizes, and runs a generic check that counts signatures of AI design.
- AI reviewers in a fresh context judge the result against a one-page design contract.
- One skill for English and German, with modules for German copy and for EU and DACH law.
Install
npx skills add jakob-meyer/unaverage --skill unaverage-design -g -a claude-code -a codexAs a Claude Code plugin: /plugin marketplace add jakob-meyer/unaverage, then /plugin install unaverage-design@unaverage.
By hand: download unaverage-design-3.4.3.zip below and unzip it into ~/.claude/skills/ (Claude Code) or ~/.agents/skills/ (Codex).
In the Claude app: download unaverage-design-3.4.3.zip below, turn on code execution (Settings, Capabilities, Create and edit files) and upload the ZIP as a skill, as described in Anthropic's help center. In a chat without a project folder, name the skill in your request, for example "Use the unaverage-design skill for this: ...".
In the ChatGPT app it runs in Work mode once it is installed for Codex (-a codex above).
Where it runs
Built and tested in Claude Code. It also runs in the Claude app as an uploaded skill, and in Codex, including Work mode in the ChatGPT app. In a Claude chat without a project folder, name it in your request; in ChatGPT's Chat mode it does not work. In Codex and ChatGPT it often builds without its measurement scripts. In Codex's default sandbox Chrome does not start, so rendering and the measured generic check need your permission to run outside the sandbox (not yet tested with 3.4.3). It also ran on Windows 11 on 16 November 2026, in Claude Code and in the ChatGPT app.
For rendering and measuring it uses Node 22.12 or newer, a local Chrome, Chromium, Edge or Brave, and Python 3 with Pillow and NumPy. Without them it is written to say what is missing and to list every check that did not run. No vendor server, no telemetry; it runs locally, or in the Claude app in that app's cloud workspace.
Tests
unaverage design was developed against blind tests: AI agents built, played the users and judged, with three jury models (two Claude models and GPT-6.1-Sol); all measured skill runs used Claude Opus 5.5. No human judge and no real user stands behind these numbers yet. The measured round ran on version 3.4.0; 3.4.1 to 3.4.3 are corrections checked in probes, and their effect on builds is unmeasured. Method, results and limits: docs/testing.md.
Files
unaverage-design-3.4.3.zip: the skill folder of this release, the same files as in the install tests and byte for byte the file used in the Codex runs and the Windows and Claude app tests. SHA-25664df5ee74a1fd9da6e4888e05947110a69f1e10376bbdce26b52c6ea34ff03d7.
Apache License 2.0. All changes: CHANGELOG.md.