agent sidequestsIdeas worth trying

How to check an AI-created file before you use or share it

Open the exact file your agent delivered, compare it with the inputs and approved changes, and try the thing you made it for. A readable preview is a starting point. A printable pattern needs a paper check; a personal font needs your actual words; a little app needs its main interaction. Keep the editable source and export together so a correction reaches the version you will use.

Agree on what you will actually open

Bring the original inputs, the latest agreed changes, the intended app or device, and the destination. Name both the editable material and the usable output: an SVG plus a printable PDF, a timeline plus a video, or a font plus its character map. A picture of a document does not replace an editable document when editing is part of the request.

Use the small acceptance card below before a long run, or fill it in when reviewing a result already delivered. Ask the agent to return the files with the outcome of each check and anything it could not inspect. Keep that report brief enough to compare with the files yourself.

An illustrative acceptance card for a personal handwriting font
What to agreeConcrete answer
Usable outputA TTF for my jam labels, plus editable source and the confirmed character map
Exact trialShow Plum jam 2026 and my two other label texts at their intended sizes
Version and destinationlabel-font-v2.ttf and its matching editable source in one dated folder; the proof uses that TTF
Remaining human checkI open it in my label app and approve a small test print; nothing is published
Review statusPending until I try the exported font; a generated proof is not my approval

Identify the file before judging the preview

Follow the delivered link or attachment and open that output. Check its filename or revision against the delivery note and your last requested change. If three versions are present, ask which one is current and keep the others clearly labeled. A screenshot from yesterday cannot establish what is in today's download.

If the link is unavailable, ask for the saved file in the agreed destination or a supported export. Preserve work that already exists. A failed download does not require the agent to repeat the whole creation task, and an inaccessible link should stay marked inaccessible until it opens.

Compare the export with the editable source

Choose the few details that would make the result unusable if wrong. For a motion film, compare the final wording, scene order and ending against the approved script, then play the exported video. For a font, check that the agreed characters are present and recognizable in the exported font. Inspect your real text, including numbers and punctuation, rather than only an attractive alphabet preview.

Ask the agent which source revision produced that export and whether it inspected the exported file itself. After a source edit, request a fresh export and repeat the affected check. If the agent cannot render, play or otherwise inspect the format, have it name that gap and leave the check for you. Do not turn a proposed check into a passed one.

Try the result where you intend to use it

Use the intended application, screen size or physical setting. A label font gets checked in your label layout. A short film gets played on the device where you plan to show it, with sound if sound matters. An app gets one complete journey from starting input to visible result, using synthetic data and an authorized test environment.

Simon Willison's Agentic manual testing describes having coding agents exercise their output and inspect screenshots, including defects that automated tests missed. He also created the testing tools he discusses. That is firsthand software practice, not proof that an agent can inspect every file format or that this guide's checks catch every defect.

A worked example: the pattern looks right but prints wrong

Imagine an agent has returned an editable apron pattern and a tiled PDF. This example is illustrative and has not been executed. The latest instruction was to preserve the original dimensions and add piece labels. The SVG looks right, but the PDF's calibration square measures differently on paper.

In a September 17 maker account, 糸巻きこけし describes working on AI-assisted pattern files, checking dimensions, assembling paper and sewing before further revisions. The writer runs a sewing business and promotes a YouTube channel. This is a self-report with human corrections, not our reproduction or a model comparison.

  1. Keep the original pattern and both output files. Record exactly which PDF you printed, the paper size and the print settings used.
  2. Compare the calibration square in the SVG, the exported PDF and the paper copy. Ask the agent to identify whether the mismatch is already in the export or appears only during printing.
  3. Repair the affected export or change the relevant print setting. Preserve the design dimensions; do not resize the design merely to make one printer’s result look correct.
  4. Print the calibration page again. Assemble the paper and compare important edges with the original before using fabric. Digital agreement does not establish physical scale or garment fit.

Ask for the smallest repair you can describe

Name the file, what you did, what happened and what should have happened. Include the relevant source detail or screenshot. This gives the agent something to reproduce without discarding the parts you already approved.

An illustrative repair request: In label-font-v2.ttf, the 6 is unreadable in Plum jam 2026 at my label size. Compare it with character 18 on the confirmed source sheet. Preserve the other approved characters, correct the affected outline or spacing, then return a new TTF and a proof made from that exact export. State what you checked and what still needs my review.

Sometimes the input needs repair. In draw-your-font issue #8, prisme49 reports a scan treating template grid lines as character components. The reporter attributes the experiment to geeks.fr; we have not reproduced it. A cleaner capture or corrected character map may be the next step, rather than more cosmetic changes to the final font.

Check the delivered copy and leave an honest finish

Once the output passes its intended-use check, open the copy at the agreed destination. For a private shared file, inspect its access setting and, when appropriate, have the intended recipient confirm access. Do not make it public to resolve a permission problem. Keep creation, your review and sharing as separate states; perform a send or publish only within the authority you gave.

Ask for a short final note naming the current source and export, checks performed, unresolved limitations and the next decision. Anthropic's agent-evaluation guidance distinguishes a conversation record from the resulting state in the environment. Our acceptance card applies that distinction to personal files; it is an original, untested review method. It does not establish compatibility across Muse, Instinct, Dots or Grok Bot.

Copy a prompt and try it

Each idea has a complete starter prompt, the inputs to bring, and the tools it needs.

Common questions

Is a screenshot enough to check an AI-created file?

It can show the captured appearance. Open the actual export to check its contents and intended behavior. A screenshot cannot demonstrate that an editable source exists, a font types correctly or a video plays to the end.

Can I do this if my assistant cannot preview the output?

Yes. Have it return the usable file, editable source and specific checks it could not perform. Open the file yourself and report a concrete mismatch. Ask for a supported fallback format if no usable export can be produced.

Does checking a sample prove the whole file is correct?

No. A sample establishes only what you inspected. Check every critical dimension, required character or promised interaction, and keep untested areas visible. Use a qualified reviewer when the consequences require expertise.

Sources and further reading

These are practical editorial examples. The prompts have not been independently run-tested across every assistant.

Keep exploring

For platform setup, read the Muse, Instinct, Dots and Grok Bot guides.