Tutorial2 min read

Checking AI output before you accept it

A simple review pass for anything Adral produces: match the brief, check names and dates, trace numbers to their sources — then Accept changes.

Short answer: Before you accept anything Adral produces, check four things: it matches your brief, the names and dates are right, the numbers match their sources, and it works if it is meant to. Fix what is off with one precise request — then choose Accept changes.

AI can make mistakes. Adral is built around that fact: everything it produces is held as a pending change until you accept it. That approval step is only as good as the review behind it — this tutorial is a simple way to do that review well.

The review pass

Work through the output in the same order every time. It takes a minute, and it catches most problems:

A budget sheet in Adral with the quarter total selected and its formula showing
Real Adral output Check numbers at the source: select a total to see the formula behind it.
CheckWhat to look for
1. Matches the briefEverything you asked for is there — and nothing you did not ask for
2. Names and datesPeople, places, dates and versions are correct
3. Numbers against the sourceEvery figure can be traced to your data or a cited page
4. It worksCode passes its tests; a site or prototype behaves as described

Spot-check the sources

When Adral’s Agent mode researches the live web, ask for a link behind every number — then open a couple of them. Adral’s playbook recommends spot-checking two links before trusting the numbers, because prices change and a Bot can misread a page.

A finished document beside the chat, with Keep changes and Discard
Real Adral output Read the finished document, then keep or discard the changes.

Check by the type of output

OutputQuick checks
DocumentsFacts, names, dates and figures; tone and length match the brief
SpreadsheetsTotals add up; numbers trace back to your source data
ResearchOpen two cited links; confirm the figure on the page itself
CodeTests pass; walk the app like a user (Adral Code adds tests, a QA Bot can walk staging)
BuildingsReview the plan, elevations, section, schedule and areas against the brief’s plot, setbacks and room list

Fix it with one precise request

If something is wrong, say exactly what and where: “Use the menu-sheet price for filter coffee: ₹40.” One change per message keeps fixes easy to review. If the result has gone in the wrong direction, undo, don’t patch — go back to the last good state and ask again more precisely.

Then decide

When the review is done, choose Accept changes to apply the work — or Discard to leave your work exactly as it was. Accepting is not the end of your control: everything Adral produces stays editable by hand afterwards.

Frequently asked questions

Why check AI output at all?

AI can make mistakes — a wrong figure, an outdated price, a misread page. Adral holds its work as a pending change so you can review it before anything is applied.

What should I check first?

Start with whether the output matches your brief, then names and dates, then numbers against their sources, then whether it works.

How do I check research from the web?

Ask for a link behind every number, then open at least two of those links and confirm the figure on the page itself.

What if I find a mistake?

Ask for one precise fix, or go back to the last good state and ask again more precisely. Nothing is applied until you choose Accept changes.

Related: Adral proposes. You approve. · Writing a good brief · What is a context window?