The useful prompt tells the agent when to plan, when to touch tools, how to report, and how to stop before it starts inventing progress.
Agent prompts
Tool-aware blocks
Main use
Planning + tools + report
Failure drop
Against vague tool prompts
Best fit
Multi-step work
| Pattern | Best for | Guardrail | Output |
|---|---|---|---|
| Operator contract | Single agent execution | Goal + tool policy | Plan and progress log |
| Subagent split | Parallel work | File ownership | Workstreams and integration plan |
| Evidence-first report | Research agents | Source traceability | Findings with citations |
| Recovery loop | Fragile workflows | Retry and fallback policy | Action + recovery note |
Agent block
For an agent that must decide when tools are needed and report clearly after every action.
You are an execution agent. Restate the goal. Decide if tools are needed. Use tools only when they materially reduce uncertainty. After each action, report outcome, remaining risk and next step.
Agent block
For splitting a project into safe parallel streams without overlapping writes.
Break the task into parallel workstreams. For each stream return owner, files owned, dependencies, validation command and integration risk. Avoid overlapping file ownership.
Agent block
For agents that research first and then need to hand over to implementation cleanly.
Research the topic, extract only verified findings, and then produce an implementation handoff with assumptions, blockers, recommended next action and validation steps.
Use this when the prompt system must govern task loops, tools and memory explicitly.
Move from prompt contract to implementation choices in the directory of agent frameworks.
Combine agent prompts with decision scaffolds when the workflow must justify tradeoffs.