Bokili
Perspectives3 min read

Define Done Before You Delegate Work to AI

Give AI work a five-line definition of done—purpose, evidence, boundaries, format and decision owner—before you write the prompt.

Bokili Editorial· Verified August 17, 2026
ShareX
A work brief passes through five quality gates before and after AI generation.

Many weak AI outputs begin as weak assignments. A manager asks for a “good briefing”, but never defines what the briefing must enable, which evidence it may use or who must approve it. The model fills those gaps with plausible choices. The fix is not a longer prompt. Define done before the task starts.

Quality belongs in the brief

Microsoft’s human-AI interaction research recommends making clear what an AI system can do and how well it can do it, and planning for the moments when it is wrong. NIST’s generative-AI risk profile likewise treats evaluation and risk management as organisational work. For everyday tasks, that principle can become a five-line contract between requester, user and reviewer.

The five-line definition of done

1

Purpose

Name the decision or action this output must support.

2

Evidence

List the sources it may use and what claims must be traceable.

3

Boundaries

State what data, claims, actions and topics are out of scope.

4

Format

Specify length, structure, tone and the fields a reviewer needs.

5

Decision owner

Name the human role that checks the result and decides whether it can be used.

Worked example: a supplier briefing

“Summarise these supplier proposals” sounds efficient but leaves every important choice open. A definition of done makes the assignment reviewable: the brief supports a shortlist meeting; claims must cite the submitted proposals; legal compliance is out of scope for the model; the output is a one-page comparison with unknowns; procurement owns the final recommendation.

Prompt without a definition of done

  • Make a persuasive summary of these proposals.
  • Choose the best supplier.
  • Keep it concise.

Brief with acceptance criteria

  • Support a shortlist decision, not a final award.
  • Use only the proposals and cite each material claim.
  • Separate facts, assumptions and missing evidence.
  • Procurement reviews and owns the decision.

Turn the definition into a prompt

  1. 1

    Write the five lines first

    Do this outside the chat so the standard does not drift with the output.

  2. 2

    Add the task context

    Explain the audience, deadline and supplied materials without changing the acceptance criteria.

  3. 3

    Ask for a self-check

    Require a final table showing how the output meets each line and where evidence is missing.

  4. 4

    Review against the contract

    Judge the work against the five lines, not fluency or confidence.

  5. 5

    Record one failure pattern

    Use it to improve the next practice task, template or guardrail.

Reading is a start. Practice makes it stick.

Start learning
Looks goodIs done
PurposeReads smoothlyEnables the named decision
EvidenceIncludes detailsMaterial claims are traceable
BoundariesSounds completeMarks exclusions and unknowns
OwnershipAI recommendsA named human role decides

Manager’s pre-delegation check

  • Can I name the decision this output supports?
  • Have I listed allowed evidence?
  • Have I set data and action boundaries?
  • Could two reviewers agree on the required format?
  • Is one human role accountable for acceptance?

Put the standard where work begins

A definition of done works best when it is visible before anyone opens an AI tool. Add the five lines to the request form, task card or team template that already starts the work. Keep examples beside the criteria, but do not let an example replace them. If reviewers repeatedly disagree, the criterion is probably too vague. Rewrite it so a colleague can point to observable evidence. This makes the standard portable across models and tools: the technology may change, while the decision, evidence boundary and accountable owner remain stable.

Practise on a reversible task

Write one definition of done
  1. Choose a low-risk task you already understand.
  2. Write the purpose in one sentence.
  3. List the evidence and one important boundary.
  4. Specify a review-friendly format.
  5. Name the decision owner, then ask a colleague whether the five lines are unambiguous.

Acceptance criteria do not make every task suitable for AI

If the work involves prohibited data, rights-sensitive decisions, regulated judgment or an action outside the approved tool boundary, stop and follow the relevant policy.

The central habit is to move judgment to the start. A definition of done gives the AI less room to guess and gives the reviewer something better than instinct. Bokili helps teams practise this habit on realistic, bounded work until clear delegation and verification become part of normal management.

Sources

  1. Guidelines for Human-AI Interaction — Microsoft Research
  2. How to build effective human-AI interaction — Microsoft Research
  3. Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile — NIST
ShareX

Reading is a start. Practice makes it stick.

Bokili turns skills like this into ten-minute missions for your whole team, with instant feedback and progress you can see.

Start learning

Keep reading

All articles