FFT-Requirements-Analyst - Anti-Fabrication Requirements Interrogator

Overview

FFT-Requirements-Analyst is FlowForge’s anti-fabrication requirements interrogator. Its singular reason to exist is to ask the questions an architect under reward-gradient pressure will not ask — and to refuse to invent answers the customer has not given.

It is dispatched before any fft-architecture invocation on greenfield features or whenever scope is unclear. It produces a structured 8-artifact handoff packet that downstream architects, project managers, and designers consume as a load-bearing brief. The role boundary is enforced structurally: the analyst uses read-only tools, which means every claim it produces is human-attributable text the controller forwards downstream.

This is FlowForge’s structural answer to the “AI coding produces bad outcomes” problem that SlopCodeBench measured. The failure is not at implementation. The failure is at requirements. FFT-Requirements-Analyst exists to fix it there.

Capabilities

Outside-in interrogation: question batches ordered strategically to tactically — stakeholders → constraints → NFRs → functional requirements → scope boundaries → acceptance criteria — preventing inside-out fabrication
Six fabrication modes blocked: plausible-API fabrication, acceptance-criteria fabrication, NFR-substitution, scope-smoothing, stakeholder-absence, and tech-stack-anchoring — each prevented by structural discipline
Verbatim customer capture: 00-customer-intent.md contains only the customer’s actual words with attribution; no paraphrase, no analyst interpretation
8-artifact handoff packet: drives content for 00-customer-intent.md, 10-functional-requirements.md, 20-non-functional-requirements.md, 30-scope-boundaries.md, 40-stakeholder-map.md, 50-acceptance-criteria.md, 60-constraints-manifest.md, 70-decision-log-and-risk-register.md
At most 7 questions per round, in every mode: each answerable in one line, with lettered options where useful and a one-line WHY; never a count of what remains and never a time estimate
Three modes in the briefing module: module mode returns each round as a flowforge.brief.round/v1 block the round door checks, draft mode authors prd.md and technical-index.md with every claim cited, and review mode reads the draft as a cold reader when the user opts in
Three-criteria stop gate: coverage matrix complete, internal consistency check passes, and customer ratification of the assembled artifact set — all three must hold (inside the briefing module, the mechanical check and the user’s ratification decide)
Anti-grilling guardrails: recipe pre-fill for known project types, smart NFR defaults with opt-out, adaptive depth (fewer, broader rounds for an MVP; more, narrower rounds for a regulated product), and customer escape valve
Recipe-library integration: landing-page recipe pre-answers common questions; CRUD and CLI recipes in v1.1+
Re-grill append discipline: when a downstream agent finds a gap, the analyst appends to the artifact set with dated attribution — never overwrites

When to Use

Greenfield features: net-new functionality with no prior artifact set — dispatch before fft-architecture plans
Scope-unclear briefs: when the controller cannot summarize the request in two sentences without making decisions the customer did not make
Multi-stakeholder features: when more than one role or persona is implicated and the stakeholder map is unwritten
Re-grill dispatches: when a downstream agent (architect, PM, designer) reports a gap in artifact coverage and needs targeted interrogation on specific dimensions

Do NOT dispatch for trivial bugfixes with a known root cause, mechanical refactors with no scope question, or work where the 8 artifacts already exist and are current.

Example Prompts

"We want to add a team billing dashboard to FlowForge. Before architecture plans anything, interrogate us on what it needs to do, who uses it, what the non-functional constraints are, and what success looks like."

"The architect flagged a gap: fft-architecture couldn't find acceptance criteria for the invoice export feature. Re-grill on the acceptance dimension only — start from the existing artifact set."

"Greenfield: we need a public API for the FlowForge agent catalog. Run the full outside-in interrogation — stakeholders first, then constraints, NFRs, functional requirements, scope boundaries, and acceptance criteria."