← Back to work

/design-flow — the meta-artifact

AI-assisted design workflow that structures discovery, research, prototyping, review, and iteration into a repeatable product-design loop.

The documented design workflow lists its phases from discovery through review and retrospective.

Problem

Asking an agent to “design and build a launchpad” produces something. Asking an agent to “design and build your launchpad” produces nothing useful — the model has no opinion about your taste, your users, your constraints, and the resulting work feels generic because it is.

I kept rebuilding the same structure by hand: a discovery interview, a taste calibration, three visual directions, a chosen blend, the build, a UX review, fixes, a retrospective. After the fourth time it became obvious this was the artifact.

Approach

/design-flow is a Claude Code skill — a SKILL.md that orchestrates other skills through seven phases. Each phase has explicit inputs, outputs, and an approval gate before the next phase starts.

  1. Phase 1 — Discovery. /design-discovery produces brief.md. Or, if a PRD exists, fast-path imports it and runs a 2–4k verification pass.
  2. Phase 2 — Research. Parallel Haiku agents scout patterns, references, and prior art.
  3. Phase 3 — Taste. /design-taste calibrates aesthetic preferences. Skipped if a profile exists for the chosen preset.
  4. Phase 4A — Shape. Three visual directions, named, with concrete reasoning.
  5. Phase 4B — Build. /frontend-design craft receives only brief.md + chosen visual-direction.md. Lean context = better output.
  6. Phase 5 — Enhancements. Up to three of /adapt, /animate, /delight, /polish.
  7. Phase 6 — Review. /ux-review against the project. Fixes loop until score ≥ 95% or 5 passes.
  8. Phase 7 — Retrospective. What worked, what didn’t, what to carry forward.

State lives in a single checkpoint file, so any phase can resume without re-reading every artifact.

The numbers that matter

After Session 5’s optimization pass, the token budget dropped from ~144–216k per run to ~98–162k — a 30–35% cut. Wins came from:

  • Selective artifact loading on resume — checkpoint + only what the next phase needs (~5–15k per resume)
  • Leaner Phase 4B build context — only visual-direction.md + brief.md reach the build agent (~5–10k saved)
  • Capped enhancement passes — three max per run; dropped /refine as redundant with /polish
  • Sonnet routing — Phases 3 (taste) and 7 (retrospective) run on Sonnet since neither needs code generation

Why this is the most-reused thing I’ve built

Because every prototype since the parent-portal proving ground has run through it, including the launchpad case study. Skills compose: /design-flow calls /ux-review which calls /critique which references /design-memory. Each one is small and replaceable; the orchestration is what makes the whole thing feel like a process instead of a series of prompts.

The first time you build the orchestrator, it feels like overhead. The fourth time, it feels like the only sane way to design.

What I’d do differently

  • Stack mismatch detection at startup. Right now you can ask /design-flow to build an Angular app inside a React project and it’ll cheerfully comply for a while before things break.
  • Preset-scoped taste profiles. taste-profile-launchpad.md, taste-profile-insights.md — one file per product surface so the calibration phase has somewhere to land.
  • /design-flow estimate — a dry-run that reports the projected token spend before committing to a full run.
  • Post-completion /linear-prd handoff — automatically file the produced artifacts as Linear issues for follow-up.