Build a Reusable AI Task Workbook: From Brief to Evaluation

Choose one job worth repeating

This guide uses a workshop follow-up draft as an illustrative project. The goal is to turn confirmed notes into a concise participant email containing a recap, exercises and verified resource links. The output remains a draft for review. Choose a similarly bounded task from your own work: a document summary, an enquiry classification or a first version of a checklist.

Write down the manual process before involving a model. Identify the source of truth, the person responsible for the result and the errors that would make the output unusable. This gives you a baseline and prevents the exercise from becoming a search for impressive wording without a practical outcome.

Write the task contract

The contract should state the audience, input material, desired output and limits. For the workshop example, the audience is people who attended an introductory session. The sources are the approved notes and resource list. The draft must not invent a future session date, a certificate promise or an exercise that was not provided.

Task: Draft a participant follow-up email.
Audience: Beginner workshop participants.
Use: Approved notes and resource list below.
Return: Subject, recap, exercises, verified links, open questions.
Missing information: Flag it; do not invent it.
Action boundary: Prepare a draft only.

Treat this as a starting contract, not a universal prompt. Replace the fields with the actual task and review criteria. Keep instructions separate from copied notes so readers and maintainers can understand which text defines the job and which text supplies evidence.

Build a representative case set

Create a small set of synthetic or appropriately anonymized inputs. Include complete notes, notes with missing links, contradictory dates and an unrelated instruction quoted inside the source. Write the expected behaviour before generating outputs. The incomplete case should produce a useful draft with an open question, rather than a fabricated detail or an unnecessary refusal to do the entire task.

Keep some cases out of routine editing. If every change is optimized against the same familiar inputs, your workbook may reward memorizing examples rather than handling new work. Version the cases and record why each one exists, especially when it reproduces a previous failure.

Use a rubric that separates quality dimensions

Review important failures individually. A high average score should not conceal an invented commitment. If two reviewers disagree, inspect the output and refine the rubric instead of immediately averaging their scores. The disagreement may reveal that the task itself is underspecified.

Compare one change at a time

Save the baseline prompt, model configuration and case set. Change one meaningful element, such as the output structure or the handling of missing information. Run both versions on the same cases and record improvements and regressions. Because generated outputs can vary, avoid treating one favourable response as proof of a durable improvement.

For the follow-up example, a shorter prompt might improve readability but omit action owners. A more detailed prompt might preserve every fact while producing an email nobody wants to read. Use the rubric to decide which trade-off serves the task, then test fresh cases before adopting the revision.

Review the concrete draft before acting

The workbook ends with a reviewed artifact, not an automatic send. Show the subject, recipient scope, full body and attachments before any external action. If the draft changes after review, assess whether the earlier approval still applies. Keep sending permissions and recipient validation in the application or operating process rather than relying on the model to remember them.

Record corrections as evidence for future cases. If a reviewer repeatedly fixes the same missing condition, improve the contract or the source material. If the task changes substantially, create a new workbook version rather than accumulating contradictory instructions in the old one.

Your finished workbook

• One task brief naming the audience, sources, output and action boundary.
• A versioned set of normal, incomplete and conflicting inputs.
• A rubric with observable pass and failure conditions.
• A comparison record showing both gains and regressions.
• A review checklist for the exact artifact that will be used.
• A named owner and a trigger for revisiting the workflow.

Use the workbook on a small batch first. Measure the time spent preparing context, checking facts and correcting the output as well as the generation time. The useful outcome is a more dependable task process, not merely a more sophisticated prompt.