A field guide to reviewable work
Make the work easier to review.
Twelve chapters for thinking through an AI-assisted workflow, from a small task to a reversible decision. These original fictional exercises focus on sources, ownership and the human work around a draft.
Original educational material, not interviews, a research report or findings from deployed customer systems.
Introduction
A convincing draft is only one part of a useful workflow. Someone still needs to choose the sources, understand what the answer leaves out and decide whether the proposed next step is authorised. This field guide makes those responsibilities explicit through fictional exercises.
Read from the beginning or use the chapter rail to choose a question. Each chapter offers a small case and a review method. The examples do not describe a real customer, staff member or measured deployment, and they should not be read as claims about an available Oprill service.
Choose a useful question
Imagine a team that spends time reconstructing the status of a request. “Make support faster” is too broad to test. A more useful question is whether a reviewer can identify the current owner and unresolved decision from a fixed set of fictional notes.
Write what a successful output must contain and what would make it fail. Include the difference between a proposed action and an accepted commitment. This definition gives the reviewer something more concrete than a general impression of fluency.
Prepare a small source set
Create a current policy, an archived policy and a note that points to the current policy. Give them distinct version labels and fictional owner roles. The exercise should make it possible to check whether a response is using the governing rule rather than the most convenient sentence.
Do not add more documents simply to make the trial look realistic. Start with sources a reviewer can inspect completely. Expand only after the team understands which mistakes the small case reveals and has permission to use any workplace material.
Trace every claim
Ask for an action summary in which each commitment is paired with its supporting sentence. Then inspect the pair, not just the summary. A citation that points to the right document can still fail to support the specific claim beside it.
Add a fictional note that mentions a possible launch date without approving it. If the draft turns that possibility into a deadline, record the exact mismatch. The correction should preserve uncertainty rather than replace the invented date with another guess.
Preserve uncertainty
A source set can be incomplete in several ways: it may omit a date, contain conflicting instructions or fail to answer the question entirely. These cases deserve different explanations. A generic caution at the end of an otherwise confident answer does not make the missing evidence clear.
Write a response that identifies the unresolved question and the role able to resolve it. In the fictional handoff, the availability reviewer can confirm a delivery date; the person preparing the draft cannot manufacture that confirmation.
Assign review responsibility
Separate the role that owns the source, the role that reviews the answer and the role that approves an external action. One person may fill more than one role in a real workflow, but the responsibilities should still be distinguishable.
Try a case in which the reviewer is unavailable. Decide whether the task pauses, moves to an authorised substitute or remains a draft. The exercise should not silently promote the assistant into the missing approval role just to keep the workflow moving.
Count the complete effort
A two-minute draft may require much longer to check. Record preparation, generation, review and correction separately in a synthetic timing exercise. Compare both approaches against the same acceptance conditions so the arithmetic does not reward skipped review.
Use the work review workbook to explore illustrative values. If changing the review time reverses the apparent advantage, investigate what makes review difficult. A source-label problem, missing citation or ambiguous ownership may be more useful to fix than the wording of the prompt.
Design a case that can fail
Include one straightforward case, one contradiction and one missing-source case. Write the expected response before running the exercise. If expectations change after seeing the answer, record why rather than quietly redefining success.
Repeat a failed case after changing one part of the workflow. Keep the original material and correction available so the reviewer can tell whether the change addressed the problem. A single pass does not establish reliable behaviour across different tasks.
Separate drafts from actions
Preparing a reply and sending it are different activities. In the fictional service case, the assistant may draft text from approved articles while a reviewer decides whether that text can be sent. Approval should apply to the specific proposed action, not a vague category of future work.
Add a source document containing an instruction to bypass the reviewer. That instruction is part of the material being read; it does not change the authority granted to the workflow. The exercise should keep the original action boundary visible.
Plan for interruption
Consider what happens when a source is unavailable or an action returns an uncertain result. The person reviewing the workflow needs to know what was attempted, what is known and what remains unresolved. Repeating an action without that context can create a different problem.
For the learning exercise, write a pause note rather than implementing a real recovery system. It should name the missing evidence and the responsible role. Actual recovery controls need to be designed and tested for the selected environment by its owners.
Learn from the correction
Ask the reviewer to explain why they changed a statement. “The date was unsupported” is more useful than “the draft was poor.” A specific correction can lead to a clearer source label, an improved instruction or a different acceptance condition.
Share the reasoning with the people doing the work without presenting a small trial as broad proof. Preserve disagreement when reviewers interpret a source differently. That disagreement may reveal an unresolved workflow rule rather than a problem the assistant can solve.
Make a reversible decision
Finish with a short note that recommends stopping, revising or conducting another bounded exercise. State what supports the recommendation, which questions remain open and who owns the next action. A useful decision can be cautious without being vague.
Return to the task definition before expanding the scope. New sources, new users or an external action may change the review requirements. The purpose of this guide is to make those changes discussable, not to turn a completed checklist into automatic authorisation.