[MendixMau]

Direct lane (changes to skills/, bug-logs/, bin/, project-bin/, project-tests/, pipelines)

What changed and why (one paragraph): Refs #729. New on-demand, experimental skill skills/poc-build.md. A POC build is a build type (Build type: full | poc in PROJECT.md), chosen by the kickoff interview's first question (intake Q16, "Which kind of build?": Full build (Recommended) or POC build, rendered per interview-protocol.md §3) and independent of the entry mode and of Pipeline mode:. It mirrors how Pipeline mode (intake Q15, #700) was added: the question lives in bin/lib/intake-template.sh, a Build type: full line is in the PROJECT.md scaffold, the Stage P row names it. No line means full; Interview mode: auto never picks poc; an existing mid-build project is not asked. Flow: after the build-type answer ONE message with the design question ("Do you have a prototype or a design?", plus three short questions on the Nothing path); design artifacts first; ONE human stop on the clickable prototype; then model work (security and theme first, wireframe-built pages one at a time, looked at on 1280 and 390); a journey that covers the happy path and every refusal; a markdown hand-over so the POC can continue through existing-app-change.md. It names what a POC build skips and says to state that in the hand-over. Size is warn-only. Second proof build fixes in this push: the skill now tells the lead to write a fixed Unverified — how to verify: not asked, POC build … under the intake questions a POC build skips (the escape form intake.md already offers), so Stage P stops reporting FAIL; the RESUME starter says to ask Q16 first; re-seed without an invented command; page pre-flight precedence (skip Step 0 and Step 3 of ui-preflight-pages.md); one-line warnings on token naming and the bottom-bar nav.tabs. No gate or script logic changed (one line of the RESUME starter text in bin/lib/resume-template.sh).

Impact

Row	Answer
Process — human steps or gates added/removed; does anything new block; new stage artifacts	One new kickoff question (intake Q16, asked first), plus the POC flow's one human stop. Nothing blocks existing projects. A POC build does not run the stage gates. Checked by running gate-check.sh on a scaffold with Q16 answered and Q1–Q15 carrying the Unverified line: Stage P PASS (markers present, says nothing about the answers), Stages 0–4, 6, 7 PENDING, Stage 5 MANUAL; with one question unmarked Stage P is FAIL. The skill tells the user this is expected and not to waive stages to turn it green.
Dependencies — shared files touched; order vs other open PRs; new tools/platform assumptions; new artifact reads + their producer	bin/lib/intake-template.sh (Q16, appended; Q9 not renumbered), bin/init-project.sh (two PROJECT.md scaffold lines), bin/lib/resume-template.sh (step 2 of the RESUME starter), skills/conversion-runbook.md (Stage P row), skills/poc-build.md, bin/lib/skill-routing.tsv (one on-demand row) and its renders (README.md, ROUTING.md, four agent stubs), CHANGELOG.md. No ordering constraint with open PRs. Producer of Build type:: the kickoff answer (Q16). Fixtures grepping the changed lines (test-stage-p.sh, test-bug12-sync.sh T10): read, unaffected; none greps the RESUME starter text; not run.
Context cost — always-loaded words, base → head, baseline total vs 80,000	Baseline 66,014 → 66,025 (+11, the Stage P row's "Which kind of build, asked first (full or POC, intake Q16)") of 80,000, from bin/render-routing.sh --check; unchanged by the second push. poc-build.md is on demand (about 2,020 words).
Behaviour — A/B result for agent-read text, or "NOT RUN: "	Two driven builds plus a blind comparison with a no-toolkit build (below); no A/B of the skill text itself.
Runtime — seconds/resources added or saved	n/a: no check or gate added. Measured for the flow: about 36 min (first proof build), about 64 min (second, run in parallel with another build, so inflated).

Field evidence — fictional test apps only, Mendix 11.13, mxcli v0.28, 2026-11-10/12:

• Six earlier test builds of one small loan-desk app (two rounds) shaped the flow: stock-look builds scored about 2.0–2.2 on visual design whatever the process; design-system-first scored 3.8 and reproducing a supplied prototype 4.0, both about 30 min.
• First proof build driven by this text (visitor passes; 2 roles, 5 pages): about 36 min; journey 22/22; independent grade: function 26 of 29 testable, visual about 4.1 of 5 (one grader, not blind), follow-up docs 17 of 24; "nearly demo-ready".
• Second proof build (final text, kickoff question asked first and answered POC; 6 pages): about 64 min, journey 37 of 37 on three cold runs. A blind grader compared it with the same requirements built in plain plan mode with no toolkit: function 25 of 25 against 24 of 24 testable; refusals in-page 7 of 7 against 2 of 6; must-haves 10 of 10 against 5 of 10; visual mean 3.8 against 3.6; hand-over 22 of 24 against 12 of 24; build time about 64 min against about 27 (both ran in parallel, inflated). The measured gain is in the must-have list and the hand-over, not the look, and it costs time.
• The second builder listed 15 process-text problems; each is fixed in the skill, fixed in the RESUME starter, or left as a follow-up below.
• Limits: one model, small apps, nothing measured above 6 pages and 2 roles; the approval was answered from a prepared file, not a person.

Follow-ups, not in this PR: (B) assemble-prototype.js hides any class starting wf-; (C) friendlier defaults in check-design-reaches-app.sh (token names that differ from the theme's) and check-page-shell.sh (a phone bottom bar nav.tabs read as a top bar); (D) the page-shell/class rules of ui-preflight-pages.md split from its StyleGallery steps; (E) session-start steps that assume a chat and a writable toolkit clone (the git pull ritual, "state in chat which commit"); (F) a hand-over block for brd-generation; (G) a few mxcli facts for the learned-rules files; (H) the project wiring lists ./bin/exec.sh <script> without --commit-first, and does not mention FINDINGS.md/brain capture for a POC build.

• No client data anywhere in the diff
• Size cap: one new skill, one routing row and its renders, one changelog line
• Test tier reached: static checks (render-routing.sh --check, check-scripts.sh, check-no-private-citations.sh, mdl-rules.sh --check, check-portability.sh) plus two driven builds, an independent grade and a blind comparison
• For a new/changed instrument: n/a, no instrument
• For a new skill: routing row added (bin/lib/skill-routing.tsv + bin/render-routing.sh)
• CHANGELOG.md line appended in this PR, crediting the source
• New bug entries: n/a

---

Generated by Claude Code

[MendixMau]

Review: MERGE-AFTER-FIX: the skill asks "Normal build or POC build?" whenever PROJECT.md has no Build type: line, which is true of every existing project. Its routing row also fires on "a demo this week". Together, a full-pipeline project in the middle of its build can be steered into skipping its gates.

Value: this takes #729's light mode as triage option 2: an on-demand skill, with no hook or gate changes. It also folds in #730's measurements:

• one journey step per rule (skills/poc-build.md:30);
• a buildability check on the wireframes before approval (:25, [idea] Light mode: requirements or an idea to a running app in a 2-hour workshop, tested (27 min, $2.41, journey 15/15) #729 finding 4);
• a second demo user (:27).

Text only, about 1,575 words, and not overbuilt. "Generic first" holds: it lives in skills/, reads no source, and every script it cites is in install-manifest.sh. Minor: :56 uses the test app's Visit_Host as its example; a placeholder would be better.

Regression risk: no script, gate, template or hook changes, so sync-project.sh brings no new FAIL, PENDING, refusal or exit-code change. Both layouts and bash 3.2 are unaffected. The four agent-stub diffs are generated on-demand rows only. The risk is in the process path:

• :8 uses "no Build type: line" as a trigger, and no intake ever writes that line (grep finds it nowhere outside this PR).
• :37 lets the skill override CLAUDE.local.md's "Before any page MDL" rule and the stage gates.
• So a Stage 5 project whose user mentions a demo could switch to POC rules in the middle of its build.

Correctness:

• (a) The CHANGELOG.md line says "follow-up BRDs", but :31 says the hand-over is markdown, not BRD JSON.
• (b) "Done when" (:62) asks for "one deliberately broken control run fails", but step 8 (:30) never asks for one.
• (c) "Skip intake.md" (:8) is missing from the Skipped list (:35). The precedence clause (:37) covers only Skipped items, so this skip has no authority over the runbook's intake.
• (d) The routing row's roles leave out gate and review, the agents that will see every stage as PENDING.

The PR adds no fixtures, and none were needed.

Context cost: 66,014 → 66,014 baseline words (budget 80,000), measured with render-routing.sh --check on origin/master 03ab940 and on the head; in sync, exit 0.

Time cost: none per build row. The flow itself was reported at about 36 minutes for one build (not re-measured).

Exact fixes required:

1. skills/poc-build.md:8: treat a missing Build type: as full in any project that already has a stage register line or a built model. Ask the question only at project birth, or when the user explicitly asks for a POC.
2. CHANGELOG.md: replace "follow-up BRDs" with "a markdown hand-over".
3. skills/poc-build.md:30: add the broken control run to step 8, or drop it from :62.
4. skills/poc-build.md:35: add intake.md to the Skipped list.
5. Optional: add gate,review to the row's roles.

No merge conflict with origin/master.

---

Generated by Claude Code

[MendixMau]

The fixes from the review above are in 0769a5fb and c713680c:

1. In-flight projects are safe. The lead asks "Normal build or POC build?" only when the user asks for a POC build in so many words, or at project birth (the register line is still exactly the scaffold's **Stage P — Kickoff**, in progress. and SHOW ENTITIES lists none). Otherwise a missing Build type: line means full. A project past birth whose user says "we need a demo this week" does not switch: one line of explanation and the normal build continues, optionally on Pipeline mode: high-speed. An explicit switch is confirmed once and recorded as a decision. The Precedence paragraph applies only when PROJECT.md says Build type: poc.
2. CHANGELOG.md says "a markdown hand-over".
3. The broken control run is in step 8, so "Done when" asks only for what a step instructs.
4. intake.md is in the Skipped list.
5. The routing row's roles include gate and review.

Also: the test app's identifier used as an example is replaced by a neutral description. The skill does not cite #730.

Re-review walked six cases (fresh scaffold with and without a POC ask; Stage 5 with "demo this week"; Stage 5 with an explicit switch; Build type: full with "quick pilot"; an existing app with no PROJECT.md): none skips gates, the "Before any page MDL" rule or intake without an explicit choice. Verdict at c713680c: MERGE, no findings. Baseline 66,014 → 66,014.

---

Generated by Claude Code

[MendixMau]

Added since the last note (head 50783168): the kickoff interview now asks the build type, so choosing a POC build no longer depends on how a request is worded.

• Intake Q16, asked first at project birth: "Which kind of build?" with Full build (Recommended) and POC build, each with a one-line explanation. The intake preamble says it is asked before Q1. Recorded as Build type: full | poc; a missing line means full; Interview mode: auto never picks poc.
• Scaffold: PROJECT.md starts with Build type: full, marked as the default until the question is answered.
• Existing projects: tested by scaffolding with the previous state, then syncing with this head. intake.md and PROJECT.md are unchanged, the Stage P gate still passes with its 15 answers, and nobody is asked retroactively. --repair-intake stays opt-in.
• Skill: "Who asks" now points to the kickoff question; a project past kickoff still does not switch on a mention of a demo. The status block lists the kickoff question as not yet run by a fresh agent.

Files beyond the skill: bin/lib/intake-template.sh, bin/init-project.sh (scaffold line and comment), one cell in skills/conversion-runbook.md. Baseline 66,014 → 66,025 (+11). Fixtures targeting the changed scripts: test-stage-p 14/0, test-bug12-sync 64/0, test-init-autocompact 15/0, test-guide-reopen 10/0.

Review at 50783168: MERGE, no findings.

---

Generated by Claude Code

[MendixMau]

Duplicate of the note above: the first attempt to post it timed out and landed late, after the retry. Nothing new here.

---

Generated by Claude Code