Small AI insights that change how you work
Eighteen practical observations about prompts, images, research, and agents, each with something concrete to try.
Original editorial observations with a concrete action to try.
Ask for the missing input
When the answer feels generic, the missing piece may be a fact only you know: the audience, budget, source, or deadline. Ask what would most change the answer, then supply it.
Try this: Ask: “What one piece of information would help you tailor this?”
Keep the sentence you like
A rewrite can erase the best part of a draft. Name the sentence that should stay and explain why it works. You are directing an edit, not surrendering the whole piece.
Try this: Choose one sentence to preserve before your next edit.
An example needs a label
Without a label, sample facts can sneak into the new output. Tell the assistant whether your example demonstrates tone, format, reasoning, or actual source facts.
Try this: Add: “Copy the structure; the names and dates are fictional.”
The reference has a job
A visual reference can specify composition, color, subject, or mood. Saying which one matters reduces the chance that the assistant copies the wrong feature.
Try this: Describe the exact feature you want from your reference.
Leave room for the words
A striking image can fail as a banner because the subject occupies the headline space. Composition is part of the task, not a finishing detail.
Try this: State where the image will appear and where text needs to fit.
A confident number still needs units
“The result is 12” might mean hours, dollars, percentage points, or rows. Keep units and the time period beside a calculation so the result is interpretable.
Try this: Ask for inputs, units, formula, and a calculation you can repeat.
A real source can support the wrong claim
The link may exist and still refer to a different year or population. Check the sentence against the passage, not just the domain name.
Try this: Open one citation and compare the scope word by word.
Missing is a useful output
An honest blank tells you where the work needs input. An invented owner or date makes a table look finished while creating a new problem.
Try this: Specify “not stated” for missing information.
Ask for different premises
Ten slogans may be ten versions of one idea. Ask for a practical angle, a story angle, and a question angle to explore a wider set of choices.
Try this: Compare approaches before choosing wording.
You can rehearse the awkward bit
A conversation can be useful even when it produces no document. Practice the moment you find difficult, then ask for feedback on one behavior.
Try this: Role-play a project update and practice answering a skeptical question.
Use your own explanation as the starting point
Showing what you currently understand gives the assistant a specific gap to address. It can be more productive than asking for another broad explanation.
Try this: Write “Here is how I understand it. Where does this break down?”
A long chat needs a clean handoff
Old decisions, dead ends, and updated facts can coexist in a conversation. A checked handoff makes the current state easier to find.
Try this: Save the goal, artifact, decisions, open questions, and next step.
A smaller request can be more ambitious
One finished, checked component moves a big project forward. A sweeping request can leave you with lots of partial work and no usable piece.
Try this: Define one artifact that would make the project meaningfully better.
The review is part of the time cost
Fast generation can hide slow cleanup. Count the time to gather inputs, inspect the result, correct mistakes, and get it into its final format.
Try this: Time the next task from preparation to usable output.
A demo is a beginning
One successful example is worth exploring. Try missing information, conflicting inputs, and a case you did not use while writing the prompt.
Try this: Keep one fresh test case aside.
Read a creative result at its destination
A sentence that looks good in a chat might be too long on a button. An image might lose its subject at mobile size. Review where the output will actually live.
Try this: Paste the draft into the real layout before approving it.
A prompt is not a lock
Telling an agent not to touch a folder expresses your intent. Restricting its tool access is a separate control. Use both when scope matters.
Try this: Check the app’s actual permissions before delegating.
Keep a small record of what worked
Your own successful briefs become a useful reference when they include the situation, result, and corrections. A prompt without that context is harder to reuse well.
Try this: Save the request with a sentence about when it works and when it fails.
Related Power of AI pages
Keep reading with Start here, everyday uses, the tool guide, the writing workshop, and sources and standards.