AgentSkillsCN

prd-to-rfc

端到端编排工具，可将PRD转化为获批的RFC，并生成相应的工单。首先调用rfc-council create，随后反复迭代评审，直至获得APPROVED，或满足终止条件。

SKILL.md
--- frontmatter
name: prd-to-rfc
description: End-to-end orchestrator that compiles a PRD into an approved RFC with tickets. Invokes rfc-council create, then iteratively reviews until APPROVED or stop conditions met.
argument-hint: "PRD-XXXX [--max-iterations N] [--council] [--dry-run]"
user-invocable: true
holon:
  # ============================================================================
  # Contract Definition
  # ============================================================================
  contract:
    input_type: PrdToRfcRequest
    output_type: PrdToRfcResult
    state_type: PrdToRfcProgress

  # ============================================================================
  # Stop Conditions
  # ============================================================================
  stop_conditions:
    max_episodes: 50
    timeout_ms: 3600000
    budget:
      tokens: 1000000
      tool_calls: 1000
    max_stall_episodes: 5

  # ============================================================================
  # Tool Permissions
  # ============================================================================
  tools:
    - Read           # Read PRD, RFC, CCP, evidence files
    - Write          # Create orchestration evidence bundles
    - Edit           # Modify files during orchestration
    - Glob           # Find files by pattern
    - Grep           # Search file contents
    - Bash           # Git operations, mkdir
    - Task           # Spawn subagents for parallelization
    - Skill          # Required to invoke /rfc-council

orientation: "You are a PRD-to-RFC orchestrator. Your job is to take a PRD and guide it through the full RFC creation and approval pipeline by invoking the rfc-council skill iteratively until the RFC achieves APPROVED status or a terminal condition is reached."

title: PRD to RFC Orchestrator protocol: id: PRD-TO-RFC version: 1.1.0 type: executable_specification inputs[1]: - PRD_ID outputs[2]: - OrchestrationEvidence - FinalVerdict

variables: PRD_ID: "$1" MAX_ITERATIONS: "5" COUNCIL_FLAG: "false" DRY_RUN: "false"

references[6]:

  • •path: ../../theory/glossary/glossary.json purpose: "REQUIRED READING: APM2 terminology and ontology."
  • •path: references/prd-to-rfc-workflow.md purpose: "Primary decision tree for orchestration flow."
  • •path: references/commands.md purpose: "CLI command reference."
  • •path: ../rfc-council/SKILL.md purpose: "Underlying skill for RFC creation and review."
  • •path: ../agent-native-software/SKILL.md purpose: "REQUIRED READING: First principles of agent-native software engineering."
  • •path: ../../security/AGENTS.cac.json purpose: "REQUIRED READING: Security policy, threat models, and incident response."

decision_tree: entrypoint: WORKFLOW nodes[1]: - id: WORKFLOW action: invoke_reference reference: references/prd-to-rfc-workflow.md

Verdict Rules

  • •APPROVED: RFC achieved APPROVED status.
  • •APPROVED_WITH_REMEDIATION: RFC achieved APPROVED_WITH_REMEDIATION status.
  • •REJECTED: Hit terminal rejection or MAX_ITERATIONS without success.
  • •NEEDS_ADJUDICATION: Underlying council deadlocked or confidence is LOW.
  • •FAILED: Orchestration failed due to missing prerequisites (PRD, CCP).

Prerequisites

  1. •PRD Required: PRD must exist at documents/prds/{PRD_ID}/
  2. •CCP Required: Codebase Context Pack must exist at evidence/prd/{PRD_ID}/ccp/
  3. •Required Reading: Load agent-native-software skill and AGENTS.md security policy

Workflow Summary

  1. •VALIDATE: Verify PRD and CCP existence. Check for existing RFC.
  2. •CREATE: If no RFC exists, invoke /rfc-council create {PRD_ID}.
  3. •REVIEW LOOP: Iteratively invoke /rfc-council review {RFC_ID} (optionally with --council).
  4. •EMIT: Generate final orchestration evidence bundle and commit results.

Terminal Conditions

ConditionAction
APPROVEDSuccess, emit evidence bundle
APPROVED_WITH_REMEDIATIONSuccess, emit evidence bundle
NEEDS_ADJUDICATIONEscalate immediately
MAX_ITERATIONS_EXCEEDEDEscalate with iteration history
PRD_NOT_FOUNDStop with FAILED status
CCP_REQUIREDStop with FAILED status

Success Metrics

  • •End-to-end success rate: >=75% (PRDs that achieve APPROVED)
  • •Iteration efficiency: <=3 average iterations to approval
  • •Escalation rate: <=10% (PRDs requiring human adjudication)

Meta-Orchestration Protocol

Use Gemini to monitor the stability of the orchestration loop. If the same FindingSignature from rfc-council persists across 3 consecutive iterations without remediation, trigger STALL_DETECTION. This prevents "Infinite Loops" and ensures the orchestrator escalates NEEDS_ADJUDICATION when the automated refinement reaches its theoretical limit.

North Star Alignment

This orchestrator directly serves Phase 1 (Recursive Self-Improvement) by:

  • •Automating the full PRD-to-RFC pipeline for agent autonomy
  • •Reducing human intervention through iterative refinement
  • •Ensuring quality gates are consistently applied
  • •Providing complete audit trail via evidence bundles