[sellke]

/prototype now opens with divergence. An idea that needs looking at before it needs specifying gets a few genuinely different, code-portable drafts, a recorded choice, and then the chosen direction built — without ever writing a spec artifact.

Why

Writ had no pre-spec divergence step anywhere in 32 commands. /design could not provide one: every mode routes output into .writ/specs/<spec-name>/mockups/, so it structurally requires a spec that does not exist yet. /prototype was already positioned as the pre-spec front door — its own usage table claims "exploring an approach before committing to it" — but delivered one unexamined canvas mockup and nothing retained.

Evidence base: .writ/research/2026-10-19-prototype-visual-drafts-research.md (on docs/product-reconcile-and-research, not this branch).

What changed

The contract is amended, not stretched. The old outcome said the repository gained no planning artifact, which forbade retaining anything from an exploration. ADR-020 allows one outcome per command, so rather than bolting a second on, the outcome is now the idea made real — drafts to choose between, working code, or both — with shipped code as a terminal state rather than the definition of done. A run ending at three drafts, a decision and no code is complete, and Step 5 and ## Completion both say so. The spec boundary moves from "no planning artifact" to "no spec artifact", which is what exit_criteria[0] already said correctly.

Drafts are render-only and must port cheaply. A draft is a self-contained page built from the project's own component names, framework and tokens — never a screenshot, an abstract wireframe, or running code. The portability test is stated inline: a coding agent reading the draft can name the components to create or change without guessing, or the draft is a picture and not worth keeping. Step 3 hands the chosen draft to the coding agent as the direction to build.

Divergence is constructed, not requested. Asking one agent for "a few options" yields one idea three times — models fixate on their own early output and repeated samples converge (research F4). Step 2.5 declares two or three orthogonal axes before generating, runs one subagent per draft with no shared context, has each generator commit to its axis before rendering, uses ordinary rather than visionary personas, and caps at three.

Rendering delegates to the platform — native artboards, then an artifact/canvas pane, then a written structural description. Never degrades silently. Writ ships no renderer.

Gated, so cost stays proportional. UI surface or an unsettled direction turns Step 2.5 on; --diverge / --no-diverge override. A bug fix with one obvious shape runs exactly as it does today.

/create-spec --from-prototype reads the record. The chosen draft becomes the Deliverable; a new Directions Considered section carries one line per rejected draft and why it lost.

The correctness fix a reviewer should look at first

--from-prototype previously required a diff and marked Story 1 Completed ✅ unconditionally. A /prototype run can now end at a chosen direction with nothing built — so that path would have produced a spec claiming work was done because a mockup existed, sending /implement-spec past code nobody wrote.

Step 0 now crosses exploration-record against diff in a source matrix. Never mark Story 1 complete on an exploration record alone. Step 1's opening framing and Step 3's completion rules branch on the same matrix. create-spec.lean.md carries the identical change — this is a core flow correction, not an opt-in arm like Jev or the review panel, so the lean twin is not exempt.

Two judgment calls worth reviewing

Divergence rules are inline, not in skills/design-divergence/SKILL.md. Referencing a skill that does not exist is the dead-end-path defect the goldilocks assessment catalogued. Extraction is the follow-up, once a second caller exists. The research doc still describes it as a skill — a known, deliberate doc/code divergence.

create-spec.md grew, and it is already the tree's worst byte violator. test_governor_enforcement's ratchet blocked the first version of this change. Following the precedent recorded on that very entry ("two adjacent sentences were trimmed to offset"), 448 bytes were recovered by collapsing Step 1's five anchor questions to one sentence and Step 3's six bullets to the prose the lean twin already proves is sufficient. The residual is the matrix itself: recorded overage repinned 28123 → 29304 with disclosure, acknowledged and not exempted. That file wants a real reduction pass, which is not this change.

eval.sh gains a referenced_paths_allowlist row for exploration.md, which /prototype writes at runtime — the referenced-paths check caught the dangling reference before it landed. test_lean_commands sha and KEPT_HEADINGS repinned for the renamed Step 3 heading.

Verification

• bash scripts/eval.sh — 0 findings, 0 run errors
• uv run --python 3.9 pytest — 2275 passed, 2 skipped
• All bash tests pass
• prototype.md 18,234 bytes, 6,726 under the per-invocation budget

Not in this PR

• skills/design-divergence/SKILL.md extraction
• The two platform defects found during the research, filed as issues on docs/product-reconcile-and-research: cursor-ide-browser translated by no adapter, and Writ's /design colliding with Claude Code's native /design
• Whether this warrants an ADR — a command's purpose changing is arguably a direction change

🤖 Generated with Claude Code