Make Your AI Quote Your Instruction Verbatim
One line stops your AI from confidently solving the wrong problem.
An agent rarely tells you it misunderstood. It quietly reinterprets your request, fills in the ambiguity with its own assumption, and then builds a clean, well-tested version of something you never asked for. Because the output looks polished, you only notice the drift after the code is written, when unwinding it is expensive. The cheapest place to catch a misread is before the first file changes.
Why agents drift from your request
Natural-language instructions are ambiguous, and a model resolves that ambiguity silently by picking the interpretation that best fits its priors. It does not flag the choice; it just proceeds. And you cannot rely on the agent to self-diagnose the drift, because asking a model "did you understand me correctly?" gets you a confident "yes" whether or not it did. Self-evaluation does not surface the gap. Forcing the model to reproduce your exact words does.
The mechanism: quote first, code second
Before the agent touches any code, make it quote your instruction back to you word for word, then state its plan against that quote.
Before writing any code:
1. Quote my instruction back verbatim, exactly as I wrote it.
2. In one sentence, state what you will build to satisfy it.
3. Wait for my confirmation before editing anything.The verbatim quote is the trick. Paraphrasing hides drift, because a paraphrase is already the model's interpretation. An exact quote sitting next to the model's one-sentence plan makes any mismatch obvious to you at a glance. If the plan solves a different problem than the quoted words describe, you see it in one sentence.
Paraphrase vs verbatim quote
| Approach | What you see | Catches drift? |
|---|---|---|
| "Summarize what I asked" | the model's interpretation, already drifted | No — the summary hides the gap |
| "Are you clear on the task?" | a confident "yes" | No — self-eval is unreliable |
| "Quote my instruction verbatim, then state your plan" | your exact words beside its plan | Yes — mismatch is visible instantly |
Power moves
- Demand the literal string, not a restatement. The moment the agent rewords your request, it has already substituted its interpretation for your intent.
- Pair the quote with a one-sentence plan. The quote alone proves it read you; the plan beside it proves it understood you. The mismatch between the two is where drift lives.
- Gate on confirmation. Make the agent wait for your "yes" after quoting, so a misread costs one sentence instead of a full implementation.
- Reuse it for long tasks. Have the agent re-quote the current subtask at each major step, so drift can't accumulate silently across a long session.
Resources
- Anthropic — Prompt engineering overview
- OpenAI — Prompt engineering guide
- Anthropic — Building effective agents
- Simon Willison — Notes on coding with LLMs
Shipping AI or cloud systems? Yeda AI audits and hardens production LLM and container pipelines. Talk to us · Read the blog