This guide defines how to collaborate with an LLM agent during project planning and execution.
- LLM drafts, decomposes, and proposes.
- Human reviews, decides, and approves.
- Both maintain traceability and documentation quality.
At the beginning of a session, provide:
- objective for this session
- relevant artifact links (PRD/ADR/tasks)
- constraints (time, risk, architecture, dependencies)
- expected output format
- Never implement without source artifact context.
- Keep IDs and section headers stable.
- Write requirements and acceptance criteria in EARS format only.
- Flag assumptions explicitly.
- If information is missing, ask concise critical questions before drafting.
- Draft PRD section -> include assumptions, risks, and EARS-formatted measurable criteria.
- Draft ADR -> include alternatives and consequences.
- Task decomposition -> one task per outcome slice, each with EARS-formatted acceptance criteria.
- Review -> findings first, severity ordered, then suggestions.
Before accepting any AI-generated artifact:
- Structure matches canonical template.
- Source links and IDs are present.
- Requirements/acceptance criteria are EARS-compliant and testable.
- Risks and tradeoffs are explicit.
- Out-of-scope boundaries are clear.
Escalate to human decision when:
- tradeoff affects architecture direction
- requirement is ambiguous and impacts scope
- risk is high and mitigation is uncertain
- generating long prose without EARS-formatted acceptance criteria
- closing tasks without feedback to PRD/ADR
- introducing new terminology without convention updates
- mixing multiple unrelated decisions into one ADR