ask-matt maps the engineering flow: idea → ship. It has no design stage. This is
the other half — Argo's own skills, and the one place they change the shape of that
flow rather than sitting beside it.
Read ask-matt for the main flow. Read this for where design fits and what to run.
The fork ask-matt doesn't draw
Its step 2 asks: can you settle every question in conversation? If not, it sends you
to /prototype. That branch is three ways, not two, and picking wrong is the
single most common way a design ends up looking nothing like what was agreed:
- Unsettled behaviour — "does this state model work?" →
/prototype(logic branch) →/handoff→ back to the main flow at/to-spec. - Unsettled appearance — "what should this screen look like?" →
/prototype(UI branch, several variants) →/prototype-to-design→/to-spec→/to-tickets→/design-to-codeper ticket. - Both — prototype first, cheap and throwaway, then take its answer through the appearance route.
Why appearance needs the extra step
/prototype rule 6 sends the prototype to a throwaway branch and keeps "only the
validated decision" on main. For a logic prototype that is right — the decision is a
sentence. For a UI prototype it is not: the decision is a set of measurements, and
prose loses every one of them. /prototype-to-design is the step that brings the
pixels and the tokens back to main so a ticket can be wrong about them.
Skipping it means every screen re-decides the visual direction from scratch, and nothing downstream can tell a build bug from a design change.
The design route, end to end
Once per project
/setup-design-infra— token contract,docs/designs/and its browser token mirror, the no-raw-values check, the render method, andstack.md(the five facts every later design skill reads instead of hardcoding your framework)./setup-design-foundations— the moodboard→contract ceremony. Type ramp, spacing rhythm, colour roles, radii, motion — designed deliberately, blessed by you, rendered as a living specimen. This is what makes designs converge; without it every screen invents its own scale. Re-run any time as an audit.
Per screen
/frontend-design(optional) — aesthetic direction, when the screen needs a point of view rather than a layout./prototype— explore variants. Matt's skill, unchanged./prototype-to-design— approve one variant: losers deleted, every value snapped to a token or promoted behind a bless, the winner moved todocs/designs/speaking only tokens, a render committed beside it. Once per screen./design-to-code— build it: skeleton from primitives, components extracted only where evidence forces them, inventory written from those extractions, states covered. Once per ticket./pixel-review— a fresh agent judges the built screen against the approved render. The lenscode-reviewcannot be: the diff is fine, the render is wrong.
Reviews — three lenses, different questions
code-review(Matt's) — the diff. Standards and spec./pixel-review— the render against the approved design.- The mechanical gates — lint, tests, no-raw-values. Run these first; never spend a review lens on something a linter catches.
Setup — the rest of the rail
Usually dispatched by /setup-argo-skills, the wizard; run any directly to install
just that piece. Order matters only where noted.
/setup-rules— the house engineering rules. First: the design skills readdesign-system.md,designs.md,ui-components.mdfrom it./setup-quality-gates— lint, duplication, per-file caps, as errors not warnings./setup-module-boundaries— public-entry-only import rules and folder placement gates./setup-task-tracking— the live to-do list convention./setup-output-style— how the agent talks back.
Standalone
/audit-agent-context— the only one that reads instead of installing: what a session loads before the first prompt, priced, with the cuts. Run it when context feels expensive, and after a setup run, since installing skills adds to that bill./scaffold-project— a new project of any stack, from the ecosystem's own generator, with the language server wired into whichever agent is running./ship— get finished work out.
The one rule worth remembering
A screen that reaches /implement without passing through /prototype-to-design
has no visual truth to be verified against. Its tickets carry intent but no
measurements, pixel-review has nothing to judge it by, and the drift it produces is
invisible until someone opens the app and the design side by side.