Exotic skills for design taste. Each skill is a contract the agent loads BEFORE picking colors, layouts, prose, or pixel-level choices, and it refuses the consensus aesthetic the model defaults to.
Most agent harnesses choose what is statistically average. These
skills are the opposite: they exist because the operator did not want
average. Drop them next to your other skills and reference them by
name (couch-design, ulus-baker) at the start of a session where
the agent's defaults would not survive the operator's review.
The design-taste contract. Load this before any web, frontend, or visual-design work where the result has to feel like one person made it, on purpose, while caring about who would see it.
The skill carries the operator's verbatim 02:00 am brief on what taste means to them: a brief that refuses generic, refuses corporate, refuses "usually-picked" options, and asks the agent to design with the weight of someone who knows the visitor will remember the page after they close it. The skill then translates that feeling into eight operational rules the agent has to defend every design choice against:
- The childhood-memory test. Does this element have the shape of something the visitor could remember from their own life?
- The single-goal rule. Every element earns its place. Filler is banned.
- The no-repetition rule. Nothing repeats in shape, type, color, layout, motion, or voice.
- The pattern-detection rule. The operator detects AI defaults instantly. Refuse the popular pick unless the popular pick is the only thing that fits a real constraint.
- The respect-the-reader rule. Real WCAG contrast, real data, real punctuation, no scroll hijack, no custom cursor, no smooth-scroll override on the whole page.
- The bond-don't-perform rule. First-person voice where it fits, copy admits the gaps, no marketing-department register.
- The search-before-committing rule. Look at unusual references (Amiga workbench, risograph zines, sci-fi paperback covers, concrete poetry) before locking an aesthetic direction.
- The 02:00 am clause. The brief is unconditional. There is no fast path that bypasses it.
Trigger phrases the agent should recognise: ultimate taste,
couch design, the brief, not generic, no usually-picks,
no AI default.
Full skill text: couch-design/SKILL.md.
The prose-side companion to couch-design. Named after Ulus Baker
(1960-2007), Turkish sociologist whose central move was that opinion
is the enemy of thought. Load this for essays, manifestos,
philosophical writeups, grounding sections, "why we exist" pages,
and longform that has to read like a human wrote it from their
heart and not like an LLM smoothing itself into mush.
Seven operational rules the prose has to pass:
- The friend-at-the-bar rule. The claim opens the paragraph. The attribution follows. The reader meets the idea before they meet the source.
- Internalize, then say. No quoting philosophers you have not read. Paraphrase in your own English.
- The Baker move. Refuse opinion. Opinion is the comfort of "yes, I agree before reading." Thought is the public struggle against that comfort.
- Admit what you do not know, mid-argument. The strongest move available to a writer is "I don't know how to fix this part" said in the middle of the essay, not in a footer.
- The ending has to land somewhere else. The close cannot be the open in different words.
- Real punctuation, no ornament. Zero em-dashes. Zero ornament semicolons. Zero triple-adjective lists.
- Vary sentence length, repeat words. Mix four-word sentences with full-paragraph sentences. Say the same noun twice instead of synonym-hunting; the reader is already tracking it.
Plus a banned-words list inherited from ~/.claude/CLAUDE.md and
security-report-prose: the standard ChatGPT prose lexicon
(delve, essentially, tapestry, seamless, holistic,
cutting-edge, testament to, dive into, ...) is hard-banned;
filler intensifiers (actually, literally, simply, merely)
are soft-banned.
Trigger phrases: ulus baker, couch prose, writing with taste,
real philosophical voice, no parroting, no slop, filozofik,
from the heart.
Full skill text: ulus-baker/SKILL.md.
Sites built with these skills loaded into the agent at the start of the session:
- aila.sh -- the AILA project landing page.
Visual surround:
couch-design. Grounding / why-page prose:ulus-baker. - lambda-zero.com -- the
project-lambda-zeroorg page. Visual surround:couch-design. Manifesto and section prose:ulus-baker.
If you ship something with these skills loaded, open a PR adding it here. The bar is the same one the skills set on themselves: nothing generic, nothing repeated, everything earns its place.
Either copy the skill directories into your harness's skills root,
git clone https://github.com/project-lambda-zero/touch-skills.git
cp -r touch-skills/couch-design ~/.claude/skills/
cp -r touch-skills/ulus-baker ~/.claude/skills/or symlink them so updates to this repo flow through immediately,
git clone https://github.com/project-lambda-zero/touch-skills.git
ln -s "$PWD/touch-skills/couch-design" ~/.claude/skills/couch-design
ln -s "$PWD/touch-skills/ulus-baker" ~/.claude/skills/ulus-bakerPath may differ if you use a non-Claude-Code harness; the skills
themselves are plain Markdown + YAML frontmatter, so any harness that
reads a SKILL.md will work.
You are welcome to clone, fork, or quote from these. What you should
NOT do is adopt them as a default style without understanding what
they refuse. The no-default rule IS the point. Reading each
SKILL.md in full before loading is part of using it correctly.
If you write new skills in this register and want to contribute them, open a PR; the bar is "would the operator load this at 02:00 am next to the existing two?"
AGPL-3.0. See LICENSE.