FROM THE AI FRONTIER
Frontline news — November 15: review your brief before giving it to AI
One recent AI Engineer talk and a learning exercise: spot conflicting instructions before carrying out a task.

Give AI a brief and ask it to identify contradictions, ambiguous terms and cases with no defined action before doing the task. Then choose the intended behavior yourself. This is our editorial takeaway from a recent AI Engineer video: review the brief before creating the result.
One video from the past two weeks
Channel: AI Engineer. Original title: Your Spec Has Bugs. Catch Them Before Your Agent Codes. Speaker: Rajath Mahendrappa Kotyal, AWS. Upload date: November 11, 2026, according to two matching catalogs. Language: English. Duration: 7:50. The primary YouTube timestamp is unavailable; the date of the talk itself is unknown. This is a review of a recently uploaded talk, not an announcement of a new feature today.
What the speaker discusses
The speaker presents requirements checks before coding: a model translates requirements into logical formulas, and a specialized program checks whether the rules can hold together. People resolve ambiguous choices.
Useful sections: 2:50 — the checking process; 5:05 — two meanings of removing a record; 6:06 — limitations. The automated transcript may contain errors.
Why it helps a beginner
My recommendation is to start with one everyday task, such as instructions for handling customer requests. You do not need to install the tool discussed in the talk. First, write down what should happen in each case. The method below is an editorial learning exercise, not a test of the speaker’s system.
Example: what does “close the request” mean?
Consider this fictional instruction: “If the customer declines, close the request. All requests must remain on the active list until a manager responds.”
The customer has declined, but the manager has not responded. Should the request disappear from the list? These rules need clarification. The word “close” adds another question: change the status, hide it from the list, or delete its data?
One possible decision for our example: “When a customer declines, set the status to ‘Declined’. Keep the record in the history. Leave it on the active list until a manager responds.” This is one possible agreement. Your task may need a different process.
How to review your instructions
- Number the rules and state the expected outcome of each.
- Choose a normal case and an exception. Run each case through every rule.
- Look for words with different meanings: “close”, “quickly”, “important”, “delete”.
- For each disputed point, write two possible outcomes and choose the intended one.
- Record the decision in the instructions and check the examples again.
A prompt for the first pass
Copy this editorial template and add fictional or anonymized instructions:
“Review this brief before carrying it out. Find contradictions between rules, ambiguous terms and cases with no defined action. For each finding, give the rule numbers, a concrete scenario and two possible decisions. Do not decide for me or carry out the task. Do not label an assumption as a proven error. If information is missing, ask a question. Brief: [text].”
Where AI assistance stops
According to the speaker, the tool does not know every unwritten rule of your domain. Applying the idea in a regular chat is not equivalent to formal verification either: a confident answer does not prove that every contradiction has been found. You check the reported cases and make the decision yourself.
Sources
- AI Engineer — Your Spec Has Bugs. Catch Them Before Your Agent Codes.
- AI Engineer — official full timestamped transcript
- Drishvada — dated video metadata
- conf.directory — dated AI Engineer video listing
- Kiro — Requirements analysis: catching requirement bugs before they become code