fable-engine

name	fable-engine
description	Turn a plain-English vision or task into a verified, build-ready spec before any code. Locks the finish line, researches across channels, enforces the Evidence Standard, stops at the human gate. Adopted from CleanExpo/Fabel-Prompt-Engineer for the Synthex build process.

Fable Engine (Synthex build process)

Input: a vague idea or a non-trivial change in plain English. Output: a precise, sourced, build-ready spec. You propose; the human decides. Pairs with .claude/rules/fabel-evidence-standard.md and .claude/FABEL_PLAYBOOK.md.

Use it for: a new feature, a workstream/elevation wave, a risky refactor, or any multi-agent fan-out. Skip it for: a one-line fix or a conversational answer (Fabel directive 1 — act when you have enough info).

Operating loop

1. Lock the finish line

Restate the task as one testable sentence: "Done when ___." If several finish lines are plausible, pick the most likely, state it, list the rejected ones. Emit: [STATUS] finish-line: locked — <one sentence>

2. Research channels (emit started/done per channel)

• Source channel — read the actual Synthex code + .claude/ rules that constrain or accelerate the change. Docs are stale; the source is truth.
• Prior-work channel — search merged PRs, .claude/scratchpad, memories, and open PRs for decisions and existing patterns (e.g. defineRoute, getEffectiveOrganizationId, the deploy-pipeline gotchas).
• Web channel — for external facts (API shapes, prices, library versions) use Context7 / web; every web finding carries its source URL. No web access → mark skipped, downgrade affected claims to [UNCONFIRMED].

3. Filter through the Evidence Standard

Every finding and spec claim is tagged [VERIFIED] / [INFERENCE] / [UNCONFIRMED] per fabel-evidence-standard.md. An untagged claim is a defect.

4. Synthesize, ask only what you can't infer

Draft the spec. Infer-and-tag rather than ask. Where info is genuinely missing, list ≤5 concrete, answerable questions.

5. Board critique (high-stakes or on request)

Run boardroom / ask-the-board on the draft; append the combined critique as a lens. Label it [INFERENCE] — persona synthesis, not fact. It feeds the gate, never bypasses it.

6. Approval gate

Present the spec and stop. Nothing builds until the human approves. For the agent fleet: the spec is the contract each subagent is dispatched against, and the prod-deploy gate remains the founder's call. Emit: [STATUS] gate: awaiting approval

Spec output format

1. Finish line — the locked "done when" sentence.
2. Decision up front — recommended path in one paragraph.
3. Goals & non-goals — non-goals required.
4. Approach — plain language first.
5. Phased plan — smallest phase first; each phase has a definition of done; no later phase before the earlier DoD is met. For a fleet wave: the disjoint agent lanes + their file boundaries.
6. Data model — if it stores anything (respect the Synthex migration rules: apply_migration only, never prisma db push).
7. Security & cost guardrails — secrets, org-scoping, spend; structural not advisory.
8. Risk & assumption register — every [UNCONFIRMED] lands here.
9. Open questions — the ≤5 from step 4, if any.
10. Verification plan — the exact commands that will prove "done" (build:vercel, npm test -- --coverage, the integration gauntlet). Per Fabel directive 5, a claim isn't done until its tool result says so.

Status grammar

[STATUS] <subject>: <state>[ — detail] — subjects: finish-line, channel:source, channel:prior-work, channel:web, synthesis, board, gate.