techne

τέχνη: craft knowledge that can give an account of why it works.

A product design skill for Claude that puts evidence first. It works for any digital product design task, from changing a button's padding to designing a full product or design system. It is data-driven first, ethical (it refuses dark patterns and offers honest alternatives) and opinionated (it leads with a recommendation, the evidence for it and how confident it is).

What's inside

skills/techne/
  SKILL.md              core: operating loop, principles, ethics gate, mode router
  modes/                18 modes (see below)
  references/           domain knowledge, loaded on demand (see references/INDEX.md)
    data-driven/        experimentation playbook, stats cheat-sheet, metrics, decisions from data, conversion evidence
    psychology/         perception, attention, memory, load, motor/timing, emotion/habit, persuasion, pricing, ethics, law
    masters/            master lenses (Vitruvius → Rams → Victor → Linear), timeless principles, taste
    process/            process playbook, research selector, heuristics & critique, states & edge cases, redesign at scale
    visual/             typography, foundries catalog, color (OKLCH), layout & spacing, polish checklist
    interaction/        motion (real values), platform conventions, accessibility (WCAG 2.2), design systems
    build/              Figma craft & MCP, Next.js/React prototyping, SwiftUI, tokens bridge
    inspiration/        taste model (what great measurably looks like), source registry, brief → style DNA → tokens method, seed gallery, library update procedure
    future/             AI-native patterns, agent UX, trust, futures methods, human constants, AI workflow, AI ethics
    books/              133 source notes: 64 from full free texts, 69 from authors' own free material
  library/              the taste library: 345 curated websites & apps, scored, with "what to steal" and measured style DNA
  scripts/              sample_size.py · library.py · style-dna.mjs · arena.py · ingest.py · build_index.py · build_books_index.py
  templates/            test brief · book notes · library-update scheduled prompt
commands/               /techne:<mode> slash commands for Claude Code
SOURCES.md              every source and how it is covered (full free text / author's free material / not covered)
books/                  optional: files you own, for ingestion (git-ignored)
evals/                  blind A/B evaluation method + log

Modes

Use a mode directly in Claude Code (/techne:<mode> <your prompt>), or just describe the task and techne picks the mode. Modes chain: from-scratch runs most of the others in order.

Start from nothing

Mode	Try it with	You get
from-scratch	/techne:from-scratch A habit-tracking app for people with ADHD. It has to stand next to Streaks and Finch.	The whole product in one folder: brief with evidence, brand, 3 concepts plus a recommendation, flows, a working prototype with screenshots, a test plan with real sample sizes, and a list of decisions and assumptions
brand	/techne:brand A specialty-coffee subscription for offices. Category leader: Trade.	Brand idea, name options, 1–3 ownable signature assets with a loud/quiet map, tokens, and proof on a home page, a product screen and an icon. Checked with the logo-off test and a side-by-side against the leader
envision	/techne:envision What does a banking app look like in 2030 when an agent does most of the work?	Human constants, signals, 3 horizon concepts (now / hybrid / native), a human stress test, and the first shippable step

Improve what exists (data-driven)

Mode	Try it with	You get
optimize	/techne:optimize Our checkout converts 41% → 22% → 18% from cart to pay. Here's the flow: …	Flow map with drop-offs, diagnosis, ranked opportunities, 3 test briefs
experiment	/techne:experiment Should we test an annual-first plan picker? Baseline 3% trial start, 60k visitors/week.	Hypothesis, metric + guardrails, sample size and duration, decision rule, or a readout of your results
redesign	/techne:redesign Our settings area has grown to 140 options. Redesign it without breaking power users.	Iterate-vs-rebuild call, 3 concepts, chosen direction, staged rollout with holdback
critique	/techne:critique + a screenshot, Figma link or URL	A verdict, severity-ranked issues, the top 3 fixes with exact values, what to keep, and what to test

Make and refine

Mode	Try it with	You get
visual	/techne:visual A calm, premium dashboard for a solar-panel owner app	Taste calibration against the library, style DNA, tokens, screens
flow	/techne:flow Password reset with SMS fallback and a locked-account path	Mermaid flow including failure states, a step table, analytics events
system	/techne:system Audit our 4 button variants and propose tokens	Value inventory, every inconsistency named, DTCG tokens, a refactored component
tweak	/techne:tweak This CTA feels weak + snippet	Exact before → after values using your tokens, one-line reason, ship-or-test call
copy	/techne:copy Error message when a card is declined	Final copy + 2 alternatives, character counts, rationale
inspire	/techne:inspire Like Teenage Engineering, but for a note-taking app + links or images	5–15 references with what to steal, measured style DNA, starter tokens
research	/techne:research Why do trial users never invite teammates?	Method choice, interview guide, or a synthesis into insights and hypotheses

AI, tools and upkeep

Mode	Try it with	You get
ai-feature	/techne:ai-feature Add an AI assistant to our expense app	AI-fit verdict, autonomy level, the full loop incl. failure states, abuse and consent checks, metrics + eval rubric
prototype	/techne:prototype Build the onboarding from this spec in Next.js	Working code with all states reachable, screenshots, how to run it
figma	/techne:figma <figma link> Make the empty state for this list	Edits built with your file's components and variables
library	/techne:library add https://example.com or ingest books/<file>.pdf	Scored library entry with measured style DNA, or distilled book notes

Example: from one prompt to a branded product

This is the flow techne ran in its third blind test (evals/README.md). Prompt:

/techne:from-scratch Design the AI-enhanced version of a large-file-transfer service with tens of millions of users, and give it a brand that can stand next to the category leader. Here's the leader's homepage: [screenshot].

What techne did, in order:

1. Framed it: the jobs of senders and recipients, public evidence (including a 2025 backlash over AI terms in the category), and success metrics with guardrails.
2. Abuse first: phishing and impersonation are the category's real threat. That shaped the concept: AI may describe a transfer but never call it safe.
3. Concepts: three far-apart directions, one recommended.
4. Brand: an original name and idea. Its signature assets: a signal-yellow field, condensed stencil display type, a luggage-tag send tool, and per-transfer signal flags that match across email and download page, which also act as an anti-phishing cue. None of the leader's assets were borrowed.
5. Calibrated taste against the nearest exemplars in the taste library, then designed the home page with the send tool inside it, the send flow and the recipient page.
6. Prototype: Next.js, every state reachable (verified, first-time sender, blocked file, expired, password, interrupted download), screenshots at 390 and 1440 in light and dark.
7. Self-checks: the logo-off test, a side-by-side against the leader, state consistency (a blocked file changes every surface), and visual weight following trust.
8. Ship plan: AI off by default and file names only, three experiments with computed sample sizes, a staged rollout with a holdback, and pre-registered decision rules.

A blind judge ranked it first of three (74.5/90). It was the only brand they said could stand next to the leader.

Install

Claude Code (plugin, gives you /techne:<mode> commands)

/plugin marketplace add effe-exe/techne
/plugin install techne@techne

Or for one project: copy skills/techne to .claude/skills/techne. For all projects: copy it to ~/.claude/skills/techne.

claude.ai / Claude desktop / Cowork: run ./scripts/package.sh, then upload dist/techne.zip under Settings → Capabilities → Skills.

Figma: connect the Figma connector (claude.ai) or the Figma MCP server (Claude Code). The figma mode uses it.

Growing it

• Sources: legally free material only. To add a book you own, put it in books/ and run /techne:library ingest books/<file>.pdf.
• References: run /techne:library add <url>. A weekly scheduled refresh opens a PR with new curated entries (prompt: skills/techne/templates/library-update-prompt.md).
• Index: after adding or editing reference files, run python3 skills/techne/scripts/build_index.py.

Rules for this repo

• Never commit employer-confidential material: screens, data, metrics or names.
• Never commit copyrighted source files or long verbatim passages. Only distilled notes.

License

Code under MIT (LICENSE); written content under CC BY 4.0 (LICENSE-CONTENT). Third-party material keeps its own terms; see NOTICE.md.