Write a bounded agent brief

Use this when a task may need an agent and you want its job defined before building it. Start with its permitted information, actions, stopping point, and human reviewer.

Prompt template

Help me specify a bounded AI agent for [task].

Context: [audience, process, approved information, and current constraints].
Desired outcome: [concrete artifact or result].
Available tools: [only the tools actually available; use “unknown” where needed].
Allowed access and actions: [explicit read/write boundaries].
Human review: [the person or role who reviews consequential actions].

Produce:
1. A plain-language explanation of why an agent would help, and whether a fixed workflow could do the job.
2. The proposed inputs, state, tools, and decision points.
3. Validation and permission checks the application must enforce.
4. A definition of done, a step/time/usage limit, and stop or escalation conditions.
5. Representative success, failure, and adversarial test cases.

Separate known capabilities from assumptions. Do not invent integrations, credentials, performance figures, or authority. Ask for the most consequential missing information. This is a design brief; do not connect accounts or execute actions.

Boundary: Produce only the design brief, including proposed test cases. Preserve the supplied task, facts, and access boundaries. Leave source material, accounts, and systems unchanged.

Fill in this prompt in the builder

What to supply

  • The task, context, and concrete outcome.
  • Confirmed tools and explicit read and write permissions.
  • A human reviewer and the consequential actions they would review.

Worked example

Illustrative example: the input and result below are invented.

Input

Eli, 28, wants a helper to assemble a troubleshooting brief from a selected set of public help articles. It may search that set and draft a summary. It cannot read tickets, change settings, or send messages. Eli reviews the brief.

Filled-in prompt

Show the complete prompt
Help me specify a bounded AI agent for assembling a troubleshooting brief from selected public help articles.

Context: help-desk preparation using approved public material only.
Desired outcome: a draft brief with article locators and unresolved questions.
Available tools: search within the supplied article set; no other tools confirmed.
Allowed access and actions: read selected public articles and draft text; no tickets, settings changes, or messages.
Human review: Eli reviews the complete draft.

Produce:
1. A plain-language explanation of why an agent would help, and whether a fixed workflow could do the job.
2. The proposed inputs, state, tools, and decision points.
3. Validation and permission checks the application must enforce.
4. A definition of done, a step/time/usage limit, and stop or escalation conditions.
5. Representative success, failure, and adversarial test cases.

Separate known capabilities from assumptions. Do not invent integrations, credentials, performance figures, or authority. Ask for the most consequential missing information. This is a design brief; do not connect accounts or execute actions.

Boundary: Produce only the design brief, including proposed test cases. Preserve the supplied task, facts, and access boundaries. Leave source material, accounts, and systems unchanged.

Result

Simpler option: A fixed search-and-summary workflow may be enough for a small, selected article set.
Inputs: Question and approved article set.
State: Article locators, relevant passages, and unresolved conflicts.
Tools: Search of that set only.
Application checks: Restrict retrieval to the selected set and block write or send actions.
Done: One draft brief with supporting locators and open questions.
Proposed limit: Stop after reviewing the selected set or finding a material conflict. Confirm a time and usage limit before implementation.
Escalation: Eli reviews gaps and conflicting instructions. No tool connection is made.

What I’d check or correct

I'd check that the brief considered the fixed-workflow option instead of assuming an agent is necessary. A prompt can describe access rules, but the application still has to enforce them.

What to check

  • Every tool and permission exists and is limited to the approved task.
  • The output stays a design brief, preserving your facts and access boundaries without connecting tools or taking actions.
  • The definition of done, usage limits, and escalation path can be enforced by the application.

Related pages

Sources

Share this page