Task-oriented execution model for multi-step agent work
Task-oriented execution model
Rule
For any multi-step agent task (audit, review, RCA, refactor, prototype, design), treat every requirement as an explicit trackable task. Assign stable IDs and emit as checklists.
Pattern
- Every requirement gets a stable ID (TASK-1.1, TASK-1.2, …, TASK-N.M) where N = phase, M = step within phase.
- Group tasks under headings that match phases — preserve traceability across output regenerations.
- Emit as Markdown checklists.
- Preserve scope exactly as written — do not drop or add requirements.
- Include code only in fenced blocks when required.
Why
- Prevents silent dropped requirements. In a 20-item spec, agent skipping 3 items is invisible unless output is checklist-shaped.
- Enables partial completion. Human reviewer sees which tasks got checked; unchecked = next-up work.
- Cross-regeneration stability. TASK-1.1 in run 1 = TASK-1.1 in run 2, even if the prose around it changes.
- Composes with delegation. Sub-agent per phase; each returns its checklist chunk.
When to use
- Any prompt from prompts.chat's "Agent Role" family (RCA, refactor, api-design, dependency, env-config, etc.).
- Any
/code-review,/security-review,/verifyinvocation with >5 concerns. - Any multi-phase design task where user might want to resume mid-way.
When NOT to use
- Trivial single-step tasks — checklist for one item is noise.
- Conversational Q&A — no requirement to track.
- Streaming code generation — checklist would fragment the flow.
Anti-patterns
- ❌ Renaming TASK IDs between runs — breaks resume
- ❌ Skipping IDs that "seem covered by another" — mark them checked, don't hide them
- ❌ Adding tasks the user didn't request — inflates scope
- ❌ Emitting prose without checkbox structure
Cross-refs
- [[scope-preservation-verbatim]] — companion rule on not adding/dropping requirements
- [[karpathy-guidelines]] — goal-driven execution + verifiable steps
- [[minimum-everything]] — checklist stays minimum
- Upstream: shared header used across all "Agent Role" prompts in f/awesome-chatgpt-prompts, 2026-07-03 harvest