Posts

Give an AI Editor a Stop Rule for Repeated Revisions

An editing session can continue indefinitely when every new version introduces a new preference. Define the acceptance checks before asking for another revision. Separate necessary corrections from optional refinements. Try a bounded example For a fictional announcement, the required checks might be correct facts, working registration link, plain-language opening and a stated length limit. A different adjective is optional unless it affects the intended meaning or audience. Ask the assistant to identify which check the current draft fails and propose the smallest correction. Avoid requesting a full rewrite when one sentence needs clarification. Keep the previous version so an improvement in tone does not hide a lost fact. Check before accepting After the correction, run the original checks again. If all pass, stop unless you discover a concrete new issue. Record optional ideas separately rather than treating them as reasons that the task remains incomplete. Review the final publishable...
Post a Comment
Read more

Check an AI-Generated Diagram Before Using It to Explain a Workflow

A diagram can be attractive while showing the wrong sequence. Start from the workflow's actual states and transitions, then review the image or diagram against that list. Decorative layout should come after the logic is sound. Try a bounded example For a fictional publishing process, the states might be Draft, Reviewed, Scheduled and Published. A scheduled item has not yet been publicly released. Write the condition for every arrow, including what happens when review fails or a schedule is cancelled. Ask the assistant to represent only those states. Inspect labels for invented stages and arrows that skip required decisions. Check whether a loop has an exit and whether the diagram distinguishes an optional path from a mandatory one. Check before accepting Walk through one ordinary case and one failure case using the finished diagram. If either path requires an unlabelled jump, revise the logic before adjusting colours or icons. A visually balanced arrangement does not prove the proc...
Post a Comment
Read more

Compare Two AI Drafts Without Choosing by Confidence Alone

A confident tone can make one draft feel more trustworthy than another. Compare the substance against a fixed checklist before choosing. Keep factual accuracy, completeness and readability as separate review questions. Try a bounded example For a fictional how-to note, write the required outcome and three facts that must survive. Read both drafts without looking at which tool produced them if practical. Mark unsupported details and missing steps directly in the text. Check whether each draft actually answers the reader's question. A longer introduction is not evidence of a more complete explanation. A concise answer may still omit the one condition that makes the procedure work. Check before accepting Use the original sources and a worked example to resolve factual differences. Do not ask one assistant to declare the other correct and treat that vote as independent verification. Its assessment is another output to inspect. Choose the stronger draft or combine specific verified pass...
Post a Comment
Read more

Use a Prompt Template Without Carrying Yesterday’s Assumptions Forward

A reusable prompt should preserve the useful structure of a task while allowing its facts to change. Before reusing one, identify the fields that belong to the new assignment. Dates, audience, destination and source material are common places for stale assumptions. Try a bounded example For a fictional event notice, keep the format instructions but replace the event facts in a clearly labelled input block. Do not bury the new date halfway through an old example. Mark examples as examples so they are not mistaken for the current event. Create a short preflight list: correct input, correct audience, correct output format and no obsolete facts. If a saved instruction refers to a specific product edition or account, verify that it still applies rather than copying it automatically. Check before accepting Run one small example after a significant change. Check whether the output repeats an old name or introduces details from the template's demonstration. If it does, simplify the templat...
Post a Comment
Read more

Build an AI Research Question That Has a Checkable Answer

A broad request such as “Research productivity tools” invites a broad answer that may be difficult to use. Define the decision first. Then write a question that can be answered with inspectable evidence and a clear stopping point. Try a bounded example For a fictional student workflow, the decision might be whether a notes application can produce a file accepted by an existing course system. The research question becomes “Which documented export formats match the required upload format?” That is more useful than asking which app is best. List the sources that could establish the answer: the destination's requirements and the vendor's export documentation. Record the platform and edition when they matter. Keep pricing, design preferences and unrelated features outside the first research pass. Check before accepting Ask the assistant to return Answer, Evidence and Unresolved questions. Verify the cited passages directly. If the documentation does not settle compatibility, plan an...
Post a Comment
Read more

Keep Research Notes Separate From Publishable AI Copy

Research notes contain questions, provisional interpretations and reminders. Publishable copy should make clear claims supported by the available evidence. Keep the two layers separate throughout an AI-assisted writing task. Set up a practical example Use a fictional article folder with Sources, Working notes and Final copy. In the working notes, label a statement as Confirmed, Interpretation or Pending. Give the assistant only the notes needed for the current writing step and explain which labels may appear as factual claims. Ask for unresolved questions in a separate editor section. Do not let the assistant turn “check whether mobile export exists” into a sentence asserting that the feature is available. A reminder is not a source. Verify before proceeding Before publication, search the final copy for placeholders, internal comments and language that overstates certainty. Open links and compare the article's numbers and names with the checked notes. If the evidence remains incomp...
Post a Comment
Read more

Review an AI Checklist by Trying the Steps in Order

A checklist is useful only if the intended reader can follow it. Test the sequence with a small authorised example before presenting it as a working procedure. Look for missing prerequisites, vague actions and steps that cannot be verified. Set up a practical example For a fictional file-export workflow, “Export and share the document” hides several decisions: format, destination, filename and permission. Ask the assistant to separate them and explain what result each step should produce. Keep the scope narrow enough to test. Walk through the checklist in order on a copy or disposable sample. Record where you had to guess, leave the instructions or repair an unexpected result. Do not quietly fill those gaps and then describe the original checklist as successful. Verify before proceeding Revise the failed step with a concrete instruction and an observable check. “Open the exported file and confirm that all three sections are present” is easier to assess than “Make sure everything is fin...
Post a Comment
Read more