Prompt Codex — Pipeline Architecture

Industry Standard

origin · foundation

BABOK DMBOK IEEE DORA ISTQB ISO 42001

Canonical framework definitions and domain taxonomy

Baseline question corpus — the questions that must be asked

Traceability chain origin — every downstream artifact traces back here

Taxonomy Manager

corpus governance · gate origin

pack governance triage · rerun gate assignment

Pack import, triage, and rerun pipelines — builds and refines the question corpus

Gate assignments originate here and carry downstream through the entire pipeline unchanged

Question lifecycle management — deprecation, versioning, display IDs vs. stable UUIDs

Pack registry — shelf governance, group assignment, publish · revert

Gate Assignment Matrix — criticality × answer type

Answer Type	Critical	Important	Optional
Unstructured	ai + block	ai + advise	none
Structured	ai + block	auto	none

gates carry downstream · unchanged by spec analysis

Creator

analyse · stage · launch

spec analysis curation stage 1 · 2 · 3 enriched payload

Spec Analysis

Gate-agnostic classification of every question against attached documents. Per question: covered · partial · missing + evidence excerpt, citation, suggested answer (max 500 chars). Gate assignments are not read or modified — two independent dimensions.

Curation — Keep / Drop

Operator reviews spec analysis results in SpecAnalysisPanel. Classification, confidence, evidence, AI prose, and traceKey all visible at curation time. Edit gate: operator may adjust TM gate assignments before launch. Dropped questions exit the pipeline entirely.

AI · Stages

Stage 1 — Gap Landscape Analysis

AI synthesises gap picture across all kept questions, grouped by gate priority (ai+block first, then ai+advise). Input: kept questions + classifications + gate assignments. Output: thematic narrative — not 40 individual rows but a structured coverage summary with risk prioritisation.

AI call skippable — with explicit confirmation

Stage 2 — Gap Resolution Sessions

Per partial question: AI-assisted resolution session. ai+block + partial = must resolve (will block in Orchestrator). ai+advise + partial = optional. Input: question + evidenceExcerpt + suggestedAnswer. Confirmed answers upgrade to covered, drop to lightweight payload tier.

AI call skippable — with explicit confirmation

Stage 3 — Readiness Check

Pre-launch gate prediction. Analysis is rule-based from the routing matrix: ai+block + missing = definite block · ai+block + partial = likely block · ai+advise + missing/partial = advisory. Runs in background; operator reviews or dismisses. AI presents the aggregated summary.

skippable — with explicit confirmation

Enriched Payload → Orchestrator

Covered → question + pre-populated answer · Partial (resolved) → question + confirmed answer · Partial (unresolved) → full gap record · Missing → full gap record

enriched payload · gate assignments · gap records

Orchestrator

execute · produce

act 1 · spec act 2 · prompt pack gap register tracekey

Act 1 Spec Generation

Gated Q&A session — operator answers questions; AI gates evaluate answers against Creator's gap context (not generic evaluation)

ai+block gates: answer must close the identified gap or operator is blocked

ai+advise gates: advisory findings captured in phase draft

Each phase closes → synthesis prompt fires → phase spec draft produced

Gap register: named open items from unresolved blocks, carried across all phases

All phases close → compilation prompt → complete spec in Spec Artifact Schema v1.0.0

→ Output #1: Written Specification

Act 2 Prompt Pack Generation

Operator reviews compiled spec and explicitly launches Act 2 (decision point)

Compiled spec decomposed by architectural layer → one self-contained prompt per layer

Layers: Data model · API layer · UI components · Integration · Infrastructure / config

TraceKey injection opt-in: // [PCX:traceKey] markers at function / component level

Gap register TBD items always flagged: // [PCX:TBD:traceKey] regardless of opt-in

Orchestrator is the conductor — operator runs each layer prompt in external AI tool

→ Output #2: Structured Prompt Pack

operator selects one or both outputs

Output #1 Written Specification

Complete specification in Spec Artifact Schema v1.0.0 — functional requirements, data model, API contracts, UI, business logic, integration points, constraints, glossary

Human-executable — a developer team can build from this document without AI

Gap register appended — unresolved items named explicitly as open questions

human-executable · complete · auditable

Output #2 Structured Prompt Pack

Five ready-to-run prompts — one per architectural layer — each cites the spec sections it draws from

Operator runs in Claude, Cursor, Copilot, or any AI coding tool of their choice

Generated source code carries // [PCX:traceKey] markers (opt-in) — every function, class, and component traces back to the specification requirement that produced it

ai-executable · traceable · layer-by-layer