ToolFlowKit
Writing / practical guide

AI Task Instructions for Small Teams: Define the Work and Handoff

Write the action, required output, and acceptance checks for one AI task after gathering its evidence. A complete instruction tells the model what to produce from the context pack. Save repeated jobs as a task kit rather than rebuilding the request each time.

Tell the tool what to do with the supplied material, what the deliverable must contain, and when it should stop for missing evidence or human review. Pair these instructions with a separate context pack.

In this guide
  1. Name the job and its acceptance criteria
  2. What is a complete instruction for one task?
  3. Write a procedure the next run can follow
  4. Use a small example to find ambiguity
  5. Make the handoff useful to a reviewer

Name the job and its acceptance criteria

Use one task and one intended reader. Specify the deliverable's format and the facts or conditions it must preserve.

For a weekly report, complete dates, source totals, unknown fields, and a separation of observations from explanations are part of the acceptance criteria. For a meeting summary, source positions and conditional commitments matter.

What is a complete instruction for one task?

Use the context pack as input, then specify the action and output. This is one drafting request; the task-kit guide covers saving the repeated job.

Reusable structure
Use P1 and P2 to draft a 90-word fit explanation for TRAY-S.
Start with external dimensions and ask the buyer to measure the space.
Return: draft; source IDs for claims; one included-parts question.
Keep P2 open instead of borrowing another variant's packing list.

Write a procedure the next run can follow

Reusable structure
PURPOSE: [Recurring job and reader]
INPUTS: [Required files/fields and source IDs]
PROCEDURE:
1. Check that required inputs exist and match the expected scope.
2. Prepare the requested output using the supplied evidence.
3. Check the important values, claims, and conditions against originals.
4. List unresolved items and the evidence/person needed to settle them.
DELIVERABLE: [Exact structure]
FINISH CONDITION: [Checks completed and review decision recorded]
STOP/REVIEW CONDITION: [Missing source, conflicting values, external action]
REUSE NOTES: [What changes on the next run]

Use a small example to find ambiguity

  1. Choose a sample with one normal case and one unresolved case, such as missing shipping or an uncertain deadline.
  2. Run the instructions in your own chosen service with the sample context.
  3. Compare the result to the originals. Check which step allowed an assumption or omitted a required field.
  4. Revise the instruction that caused the issue and repeat only the affected check.

Make the handoff useful to a reviewer

Ask for the output, source mapping, unresolved questions, and proposed next action. Keep a person responsible for approvals that change accounts, publish material, or commit the team.

Save the successful procedure with examples and recovery notes rather than relying on a long conversation history.

References

Continue the task

A bounded instruction becomes easier to use when the output contract, evidence checks, and execution limits are clear.