glebis Prototype
Announce at start: "I'm using the humane:prototype skill to make this idea lookable and clickable at the cheapest fidelity that answers the question."
A prototype exists to answer exactly one open question — is this the right structure, does the navigation make sense, does the screen read at a glance — before anything expensive is committed. This skill produces the cheapest artifact that answers that question and stops.
Quick Reference
| Need | Where |
|---|---|
| Format contracts per rung: file layout, hotspot markup, self-containment, the SVG text traps | references/formats.md |
| Where the output file goes | setup/references/paths.md — setup owns the path table |
| The editable design-file exit: availability, style archetypes, the active-document pre-flight | references/design-file.md |
The one question
Every prototype names, before anything is drawn, the single question it will answer. A broad request ("mock up the app") contains several candidate questions — extract them, and ask the user to pick one; a prototype chasing two questions answers neither. Record the chosen question in the artifact's notes block (see Output).
The question is settled only by the user saying so, or by a named
reviewer's output — a respondent-panel read, a walkthrough result. The
maker's own impression that a structure "looks right" settles nothing; this
skill never reviews its own output, and it does not certify its own handoffs
either. Until the question is settled, it stays open, and escalation waits.
Where it sits
The home case is greenfield: after jtbd, before design-tokens. The corpus
says what job the interface serves; the prototype proposes a structure for it;
only a structure that survives being looked at or clicked through deserves a
token set.
jtbd is preferred, not required. Entered directly, with no corpus: offer to
run jtbd first, and if the user declines, proceed — with "no corpus;
structure is a guess" stated in the artifact's notes block.
Prototyping inside an existing system is also legitimate — a new screen for a
product that already has tokens. The existing tokens are then context, and the
token-faithful tier renders under them. What this skill still never does there
is re-litigate the system: a finding about the tokens themselves belongs to
design-tokens, and defects in the eventual real build to layout-rules.
The fidelity ladder
Three formats, in escalation order. Name the open question first, then pick the lowest rung that can answer it. The user's explicit format request always wins.
| Rung | Format | Answers | Never claims |
|---|---|---|---|
| 1 | ASCII — box-drawing sketch in a fenced code block | Is this the right structure? What goes where? | Proportion, color, feel |
| 2 | SVG click-dummy — screens as SVG frames, hotspots linking them, in one self-contained HTML file | Does the navigation make sense? Can you get from A to B? | Visual polish, real rendering behavior |
| 3 | HTML — real markup and CSS, self-contained file | Does the screen read? Do the interactions feel right? | That it is the product |
Escalate only when the current rung's question is settled (see "The one question" — settled by the user or a named reviewer, never by the maker) or when the open question lives on a higher rung. Do not skip rung 1 for a new structure — an ASCII sketch that is wrong costs one message to redraw; an HTML page that is wrong costs an argument.
HTML fidelity is the user's choice. Before building rung 3, ask which tier:
- Wireframe — grayscale, system font, real copy, no brand. Default when no token set exists (the normal case at this point in the cycle).
- Token-faithful — reads the project's compiled
design-tokensCSS if one exists and renders under it. Offer this only when the tokens exist; never invent a palette to fake it — inventing one isbrandkit's job.
Format contracts — file layout, hotspot markup, the self-containment rules —
live in references/formats.md. Read it before producing rung 2 or 3.
The editable exit: a design file
When the host exposes a design-file backend, a prototype can be produced as an
editable design file (.pen) instead of a rung-3 HTML page. This is not
rung 4 — the ladder produces disposable artifacts that answer one question; a
design file is a living document someone keeps editing. Take this exit when the
user wants to carry on designing, not when they want an answer to a question.
The design_tool setting (setup) gates it: auto uses a backend when the
host exposes one, none pins it off. Say which of the two artifact kinds you
produced — "here is your prototype" means different things for a file that
opens and a file that needs an app.
The full contract — availability and the doctor caveat, why style archetypes
are scaffolding and not brand decisions, and the mandatory active-document
pre-flight before any build (the backend writes into whatever document is open)
— lives in references/design-file.md. Read it before taking this exit.
Copy in a prototype
The prototype drafts its own strings freely — buttons need words before
ux-writing has run, and lorem ipsum hides layout truths. Ground the wording
in the JTBD corpus where it exists. But drafted copy is scaffolding: if any
prototype string is about to be reused in the real product, hand it to
ux-writing — that skill owns the source wording of every user-facing string,
and prototype copy that skips the handoff ships unreviewed.
Output
Every ladder prototype is a file the user can open, not a paste into chat
(rung 1 may also be shown inline). It is self-contained: no CDN, no
external fonts, no build step. Double-click opens it. The design-file exit is
the stated exception — references/design-file.md owns its output rules.
The destination is setup's rule, not this skill's: resolve
artifact_path(name, "prototype", "html") — it returns the complete file path
(.design/prototype-<name>.html beside the project) — and never write to the
working directory. The full path table, defaults, and prohibitions live in
setup/references/paths.md. If the user names a place, use it.
Claude Code extras: offer to publish the file as an Artifact when the user wants a shareable link; the file on disk remains the source of truth.
Every prototype carries a notes block inside the file itself — a visible footer (rung 1: a paragraph under the sketch) — so the context travels with the artifact when it is opened alone. Chat may repeat it; it must never be the only copy. Three fields:
- The question this prototype answers, in one sentence.
- The job it serves, cited from the corpus (
jtbd.jsonoutcome or force) — or "no corpus; structure is a guess" whenjtbdhas not run. - What is fake — drafted copy, invented data, dead controls — so a reviewer does not report scaffolding as findings.
Smoke test before handoff
Producing the file includes proving it meets its own format contract. This is
production QA, not review — reviewing the design stays with respondent-panel,
walkthrough, and the rest. Before handing the artifact over:
- open the file with no hash and confirm the first screen shows;
- follow every hotspot and every back link once;
- confirm zero external requests (no CDN, fonts, or linked assets);
- check every string against its container (the SVG text traps in
references/formats.md).
Anything on this list that cannot be run is reported as Not verified, never silently assumed.
What this skill does not own
| Concern | Owner |
|---|---|
| Whether the structure serves a real job | jtbd — cite it, don't re-derive it |
| The token set, colors, and contrast measurement | design-tokens |
| Typeface selection | type-specimen |
| Final wording of any string that outlives the prototype | ux-writing |
| Structural defect classes in the eventual real build | layout-rules |
| How the prototype lands with strangers | respondent-panel |
| Whether a person can complete a task on it | walkthrough |
| Fitting the settled structure to an existing design system | design-frameworks |
A prototype is never reviewed by this skill. It is handed to the reviewers as a first-class artifact and judged there, once, by the skill that owns each rule.
Conflict precedence
The user's explicit words > this ruleset > the project's existing system > personal taste. If the project already has screens the prototype contradicts, flag the conflict; don't silently redesign the product.