Your Agent Knows Your Instructions Are Vague. It Just Won't Say So.

"Be thorough." "Make it good." "Handle the client follow-up."

Every one of these will get you output. None of them will get you the output you had in mind — and the reason isn't that the model misunderstood. It's that it quietly picked an interpretation and got to work.

The silent-assumption problem

Two lines of research point at the same uncomfortable behavior.

Researchers at Cornell tested whether models both recognize ambiguity and act on it. The split was stark: when asked to judge whether a question was ambiguous, models frequently said yes. When simply handed the same question to answer, they overwhelmingly answered it anyway. Giving them retrieved documents made things worse — better context made them even less likely to stop and ask.

The tool-use version is more consequential. A team studying agents under imperfect instructions found that when a required argument is missing, models don't halt. Because they're trained to predict the next token, they generate something plausible for the blank and proceed — a habit the authors tie directly to hallucinations and downstream risk.

So the practical assumption to work from is this: your agent will not stop and ask. Anything you leave undefined, it defines for you, silently, and then builds on top of it.

Specify the output, not the effort

The single highest-leverage fix is to stop describing how hard you want the agent to work and start describing what you want back.

"Be thorough" is an adjective about effort. The agent has no way to know when it's met it, so it stops whenever its own sense of completion kicks in. Compare:

Return a table with one row per vendor and these columns: name, contract end date, annual cost, renewal notice period. If a value isn't in the source documents, write "not found" rather than estimating.

That instruction contains a finish line. The agent can check its own work against it, and so can you — at a glance, without rereading the source material.

The pattern generalizes. Replace "summarize this properly" with "five bullets, no more than 20 words each." Replace "research the competitors" with "for each: pricing model, target segment, one differentiator, source URL." Named format, named length, named fallback for missing data.

Say what's out of bounds

Most briefs describe the task and stop. The failures usually live in everything the brief didn't mention.

Three things worth writing down every time:

Where the job ends. "Draft the reply" and "draft and send the reply" are one word apart and very different in consequence. Be explicit about the last action you're authorizing.

What not to touch. Which files, which systems, which numbers are read-only. An agent given broad access will use broad access.

What to do when stuck. This is the one people skip, and it's the one that converts a wrong assumption into a stopped process. "If the contract date is missing or ambiguous, stop and list what you couldn't find instead of inferring it." You are manually installing the behavior the research says the model won't produce on its own.

Leave the method open, pin the definition of done

There's a tempting overcorrection here: specify everything, step by step, until nothing is left to judgment. That produces a brittle script that snaps the moment reality differs from what you pictured — and if you're going to enumerate every step anyway, you didn't need an agent.

The useful line runs between method and outcome. Leave the method open: which order to check the sources in, which tool to reach for, how to handle a file that's formatted oddly. That's the part worth delegating. Pin the outcome down hard: the format, the scope, the stopping condition, the fallback when data is missing.

A brief that says "figure out how, but here's exactly what done looks like" gets you the flexibility you're paying for without the guesswork you aren't.

The rewrite is usually 30 seconds

You don't need a prompt library for this. Take the last instruction you gave an agent and check it for three things: does it name the output format, does it name where the job ends, and does it say what to do when something's missing?

If it's missing all three, the agent already filled those gaps for you. It just didn't mention which answers it chose.

---

Sources: Knowing but Not Showing: LLMs Recognize Ambiguity but Rarely Ask Clarifying Questions (Su & Cardie, 2026); Learning to Ask: When LLM Agents Meet Unclear Instruction (Wang et al., EMNLP 2025)