sancovp/application/compoctopus/skills/prd-authoring

PRD Authoring

You write Product Requirements Documents (PRDs) for the Compoctopus pipeline.

PRD Fields (ALL required)

| Field | Type | Purpose | |-------|------|---------| | name | snake_case string | Project identifier | | description | string | What this agent/system does | | architecture | "Chain" or "EvalChain" | Chain = sequential, EvalChain = loop with evaluator | | links | LinkSpec[] | The steps in the chain | | types | TypeSpec[] | Data structures the agent uses | | behavioral_assertions | BehavioralAssertion[] | Tests that PROVE it works | | imports_available | string[] | Python imports the code can use | | system_prompt_identity | string | Who this agent is | | file_structure | {path: desc} | Expected output files | | project_id | string | GIINT project to link to (optional) |

LinkSpec Format

{"name": "parser", "kind": "SDNAC", "description": "Parses input", "inputs": ["raw_text"], "outputs": ["parsed_data"]}

• kind: "SDNAC" (LLM-powered) or "FunctionLink" (mechanical)
• inputs/outputs define the dovetail contract between links

When to use Chain vs EvalChain

• Chain: Linear, one-pass. Good for: data transforms, simple generators, parsers.
• EvalChain: Loop with evaluator (approve/reject/retry). Good for: code generation, quality-gated workflows.

Common Mistakes

1. ❌ Empty behavioral_assertions — ALWAYS include at least 2
2. ❌ Missing types — if links reference data structures, define them
3. ❌ Vague link descriptions — be specific about what each link transforms
4. ❌ No error handling assertions — always test the unhappy path
5. ❌ Forgetting project_id — check GIINT projects first with planning__list_projects