Institutional Prompt Cookbook
Interaction patterns for audit, discovery, sprint, handoff, parallel work, Git/Kai, infrastructure, incidents, QA, deployment, and release.
Interaction patterns for audit, discovery, sprint, handoff, parallel work, Git/Kai, infrastructure, incidents, QA, deployment, and release.
The institutional Prompt Cookbook is not a collection of magic phrases. It provides patterns for giving humans and agents a clear working contract with context, limits, expected outcomes, and evidence.
Interaction patterns
Patterns cover discovery, audit, sprint, handoff, parallel work, Git, infrastructure, incidents, QA, deployment, and release. Prompts should describe engineering work rather than vague commands.
- Objective and context.
- Scope and Non-Goals.
- Constraints and authority.
- Evidence and Definition of Done.
Prompts that reduce risk
Sensitive tasks require audit before change, fact/inference separation, and stopping when required information is missing. Merge or release patterns require state review, checks, and authorization first.
- Audit before change.
- Do not invent results.
- Pre-Merge Audit.
- Rollback or Manual Fallback when relevant.
A useful prompt expresses a work contract
DCP Flow does not treat prompts as magic phrases. A strong prompt combines outcome, sufficient context, requested action, boundaries, expected evidence, and explicit authority. The methodology still lives in DCP Flow and project truth still lives in the repository; the prompt only reduces ambiguity for the current interaction.
- Clear outcome.
- Sufficient context.
- Scope and boundaries.
- Explicit evidence + authority.
The pattern changes with the lifecycle stage
The Cookbook separates patterns for auditing before touching a project, Solution Discovery, sprint preparation, Execution Contracts, handoffs, parallel-agent work, infrastructure analysis, incident diagnosis, Pre-Merge Audit, Deployment Readiness, release, and learning. The goal is not to copy every template but to select the one matching the real state of work.
- Audit before change when truth is uncertain.
- PLAN / BUILD / REVIEW with clear limits.
- Pre-Merge Audit before integration.
- Release and retrospective use different contracts.
Recommendation does not mean authorization
A person must be able to ask for analysis or a recommendation without accidentally authorizing execution. Prompts such as “do everything,” “improve whatever you see,” or “merge when done” leave material decisions implicit and increase risk. A good prompt states what may happen now and what remains behind a Human Gate.
- ASK may end without modification.
- Do not self-authorize merge or release.
- Do not expand scope on agent initiative.
- Prompt expresses intent; evidence proves the result.
Contextual glossary＋
Execution contract defining objective, scope, boundaries, evidence, and acceptance conditions.
Explicit transfer of context, scope, state, and evidence between people or agents.
Pre-integration review for conflicts, regressions, known debt, and missing evidence.
Explicit, verifiable manual procedure used when Kai cannot execute an action directly.