The build-opening prompts live in prompts/ as markdown (#1347)

[suleimansh]

Closes #1347.

prompts/README.md promised that nothing agent-facing is written in TypeScript, but steps.ts still built three prompts as string arrays (the issue counted four — improvePrompt was already removed with the production-grade gate). Changing what we tell an agent in those cases meant editing code (#1224 was a one-word prompt edit through a TypeScript file).

What changed

• New prompts/build_prompt.md, prompts/extend_prompt.md, prompts/scaffold_prompt.md — the prompt text moved verbatim, with ${{ tf.prompt }} for the user's intent (the same name the system prompt gives it). Each with its SPEC.md sibling carrying the rationale that lived in the code comments (Support running against an existing project (not just from-scratch) #185's naming-the-workspace point, Loop can't recover an empty/unbuilt workspace: 'improve' framing blocks scaffolding #182's don't-refuse retry, [The Framework] Remove meta-select + the architect #545's stack-is-the-agent's-call).
• steps.ts keeps the three functions and isWorkspaceEmpty — the functions are now one-line renderTemplate fills over the generated constants, so call sites are untouched.
• Rendered output is byte-identical to the old strings — verified by direct comparison, and the existing steps.test.ts content assertions all still pass unchanged. Break-checked: editing extend_prompt.md's first line fails the test.
• prompts/README.md table row added; prompts/SPEC.md and steps.SPEC.md updated — the prompt-content spec now lives with the prompts, steps.SPEC keeps only the choice + fill and the emptiness check.

Tests

Full suite green: 1586 node tests + 817 dashboard tests; pnpm typecheck clean. No behavior change intended, and the byte-identity check says none happened.

🤖 automated

[suleimansh]

DRY check — two duplications in this PR:

1. The per-prompt SPEC.md files paraphrase the prompts they sit next to. build_prompt.SPEC.md's business-logic paragraph restates build_prompt.md nearly clause for clause (scaffold the whole project, package manifest with scripts, install dependencies, one-paragraph summary), and scaffold_prompt.SPEC.md does the same for its prompt. The prompt file is already prose — the reviewable artifact — so this is the same content in two places, and the copy in the SPEC will silently drift on the next prompt edit. The fix: each SPEC keeps only what the prompt itself does not say — when it is sent, and the rationale (the dropped how-to-behave rules, the stack being the agent's call, why the retry insists an empty directory is expected) — and stops re-narrating the prompt body.

2. steps.ts builds the identical render context three times. All three functions are renderTemplate(X, { tf: { prompt: intent } }) — the context shape is repeated per function. A local fill(template, intent) helper would leave one place that knows how a build-opening prompt renders.

Not counting: the prompts/SPEC.md bullet appearing in both the TL;DR and the "What lives here" list — every other category in that file already appears in both, so that repetition is the file's existing pattern, not new.

🤖 curated

[suleimansh]

Pushed the dedup: the three SPECs now carry only when-sent + rationale (the prompt file beside each is the prose of what the agent is told), and steps.ts renders through one fill helper. Steps tests + typecheck green.

🤖 curated

[suleimansh]

@brillout prompt files moved (#1347) — review welcome, merging

[brillout]

I don't understand the goal of this prompt. Can we remove it and tree-shake (remove all code related to it)?

[brillout]

Let's remove all of the greenfield stuff for now (and tree-shake)

[brillout]

@suleimansh Up for reviewing the prompts? I'll do a second review after yours.

[suleimansh]

It was the ai-autopilot greenfield flow (empty-folder builds) — dead path today. Agreed — removal PR incoming. Prompts review next.