Evidence first
Ground every requirement
Stable citation and requirement identifiers preserve the origin and disposition of every approved product claim.
Governed multi-agent product documentation
From fragmented evidence to planning-ready work.
PRD Orchestrator transforms stakeholder briefs, transcripts, decisions, and templates into a corrected PRD, a traceable Epic/Feature/User Story hierarchy, and advisory sizing—without hiding uncertainty or bypassing human approval.
6authoritative inputs
8logical stages
12/12baseline tests
3planning artifacts
145/145citation dispositions
0unsupported claims
Design thesis
Automation with evidence, judgment, and release control
Human accountable
Keep decisions visible
Conflicts and gaps are surfaced for approval; the system does not silently manufacture business judgment.
Fail closed
Release only reconciled work
Deterministic controls, semantic evaluation, and the Agreement Gate protect the final export boundary.
Business problem
Manual documentation loses context between evidence and delivery
The operational risk is not merely slow writing; it is untraceable interpretation.
Fragmentation
Evidence lives in many formats
Briefs, transcripts, notes, clarifications, and templates must be reconciled before product work can begin.
Ambiguity
Missing detail becomes accidental invention
Traditional drafting can turn uncertainty into confident prose without preserving what remains unresolved.
Traceability
Delivery loses its source lineage
Stories and acceptance criteria are difficult to defend when the chain back to stakeholder evidence is broken.
Alignment
Contradictions surface too late
Conflicting goals, dates, scope, and technical constraints create rework after planning has started.
Consistency
Decomposition varies by author
PRDs, features, stories, and acceptance criteria drift in structure and completeness across teams.
Governance
Generation can outrun approval
Without a release boundary, plausible-looking outputs can be distributed before human decisions are recorded.
Solution approach
Each failure mode maps to a governed system capability
Ground
Normalize evidence, hash current-run files, and assign exact citation identifiers.
Analyze
Extract requirements and expose gaps, conflicts, dependencies, and clarification needs.
Approve
Record signed decisions and constrain generation to the approved scope.
Validate
Reconcile sets, score semantic quality, and fail closed before delivery.
System architecture
One run identity connects source, decisions, artifacts, and evaluation
A sequential topology keeps handoffs inspectable and makes rollback practical.
Google Drive · Inputs6 approved source files · content + hashes
v0.3.8 Main OrchestratorManual trigger · run identity · governed stage contracts
Google Drive · OutputsPRD · stories · sizing · lineage evidence
01Drive + ClarificationRead, normalize, hash, segment
02Requirement ExtractorItems, evidence, citation classes
03Gap AnalyzerGaps, conflicts, clarification needs
04Human ApprovalSigned decisions, approved scope
05Production PRDApproved evidence to ten sections
06Story BreakdownEpic to feature to story to criteria
07Validator + ExportReconciliation and Agreement Gate
Post-stageStory SizingAdvisory, non-blocking sizes
OpenAI intelligence layerConstrained extraction, analysis, generation, decomposition, and sizing
Langfuse observability layerTraces, code and LLM evaluators, latency, tokens, loaded cost
Governance and trust layerGrounding, approval, set equality, orphan prevention, fail closed
Specialized roles
Separation of duties improves control and diagnosability
Requirement intelligence
Preserves exact evidence and distinguishes requirements from risks, decisions, and open questions.
Gap intelligence
Finds contradictions and missing decisions without resolving them autonomously.
Document generation
Builds the PRD and delivery structure only from approved scope and source identifiers.
Assurance
Uses deterministic reconciliation and semantic judges to protect the export boundary.
Governance & evaluation
Quality controls target different failure modes
Code checks enforce contracts; semantic judges test whether the meaning remains faithful and reasonable.
Grounding
Exact evidence references
Current-run source segments receive stable citation IDs that propagate through requirements and outputs.
Decision boundary
Human approval is explicit
Contradictions, scope decisions, and unresolved questions remain visible until an accountable owner approves them.
Release control
Agreement Gate fails closed
Export requires reconciled citation, PRD, story, and acceptance-criteria sets with no orphan delivery items.
Operational baseline
Fully loaded cost makes the quality strategy measurable
545,467total tokens
$1.474409accepted-run loaded cost
43generation + LLM calls
84.5%semantic evaluation cost share
1×governed run per week
~$6.38monthly operating baseline
15.5% generation84.5% semantic evaluation
Outcomes & artifacts
Three planning artifacts emerge from one governed evidence chain
Each output is usable independently while retaining forward and reverse traceability.
Product definition
Corrected ten-section PRD
Includes requirements, NFRs, feature-level acceptance criteria, dependencies, risks, scope, and source references.
Delivery planning
Epic / Feature / User Story
Provides a dynamic hierarchy with story-specific acceptance criteria, priority, and evidence lineage.
Estimation support
Advisory story sizing
Adds proposed size, confidence, refinement flags, and planning guidance without creating commitments.
Bidirectional traceability
Every delivered story can answer “why does this exist?”
Approved stakeholder evidence
Exact source segment
Classified product need
Feature and criteria
Epic and feature
Story and testable criteria
145citations indexed
145/145terminal dispositions
0orphan citations
0orphan PRD elements
0orphan delivery items
100%grounded output
Public resources
Inspect the implementation and validated release package
Repository
Version history, documentation, and implementation package.
v0.3.8 workflows
Nine versioned n8n workflow exports and release notes.
Validated evidence
PRD, story breakdown, sizing, and lineage artifacts.