Brief the AI like a brilliant new hire: writing clear, direct instructions

Overview

This is the one prompting habit worth fixing before any other: saying exactly what you mean, in plain words, with no room for a second reading. Why now — Anthropic’s living prompting guide, current through Claude Fable 5 and the Opus 4.x line as of June 2026, still opens its general principles with “be clear and direct” and frames the model as “a brilliant but new employee who lacks context on your norms and workflows.” Newer models reward this more, not less: the guide’s model-specific page for Claude Opus 4.8 notes it “interprets prompts literally and explicitly”, particularly at lower effort levels — so a vague instruction now gets a vague-but-confident answer rather than a helpful guess. By the end you’ll have a fast test for whether any instruction you write is actually clear.

The content

The obvious read is that better prompting means learning clever phrasings — magic words, role incantations, the right template. The overturn: the highest-leverage move isn’t a trick at all. It’s removing ambiguity. Anthropic’s own guide opens its general principles with “be clear and direct”, listing it ahead of context, examples, XML structure, and roles.

The handle to keep is the colleague test. Anthropic states it as the golden rule of clear prompting: “Show your prompt to a colleague with minimal context on the task and ask them to follow it. If they’d be confused, Claude will be too.” It works because the failure mode is the same for both. A human with no context fills your gaps with their assumptions; a model fills them with the statistically likeliest reading. Neither is your intent.

The brilliant-new-hire framing does real work here. You wouldn’t tell a sharp new starter “clean up this report” and expect your version of clean. You’d say what to cut, what to keep, and what the finished thing looks like. Capability was never the gap — shared context was. The guide makes the same point in reverse: if you want “above and beyond” effort, ask for it explicitly rather than hoping the model infers it.

This is where it breaks, though. Clarity removes ambiguity; it cannot supply judgement you haven’t formed. If you don’t actually know what “good” looks like for the task, no amount of clear phrasing rescues you — you’ll write a crisp instruction for the wrong thing. The colleague test catches confusion. It does not catch a confident, well-worded mistake.

Try it

Take a prompt you’re about to send for real work today — not a toy one. Before you run it, run the colleague test on it yourself with this:

Read the instruction below as if you were a sharp new colleague with zero context on
my work, my norms, or what I consider "good". Do not try to be helpful.

1. List every word or phrase you'd have to guess the meaning of (e.g. "clean up",
   "make it punchy", "the usual format").
2. List every decision I left open that you'd have to make for me.
3. Rewrite my instruction so all of that is stated explicitly — same intent, no gaps.

INSTRUCTION:
"""
[paste your real prompt here]
"""

The list in steps 1 and 2 is the part to read closely — each item is a place the model would otherwise have guessed. Where it breaks: if step 1 comes back nearly empty but the output still isn’t what you wanted, your problem isn’t clarity, it’s that you haven’t decided what you want. That’s a thinking task, not a prompting one.

Additional reading

• Prompting best practices — Claude Docs — lists “be clear and direct” as the first general principle, ahead of context, examples, XML structure, and roles; also the golden rule and the brilliant-new-employee analogy (living page; current through Claude Fable 5 and Opus 4.8 as of June 2026).
• Prompt engineering overview — Claude Docs — points to the best-practices guide as the living reference; that’s where the techniques live.
• Prompting Claude Opus 4.8 — Claude Docs — the model-specific source of the “interprets prompts literally and explicitly” note (particularly at lower effort levels).

Editor’s note

Of all the prompting advice that survives a model upgrade, this is the one I’d keep — clarity is the cheap habit that keeps paying. The trap I see in real teams is the opposite assumption: people reach for a role (“act as a senior analyst”) or a longer template when the actual problem was a single undefined word — “summarise” without saying for whom, or “tidy this up” without saying what tidy means. The colleague test is useful because it’s uncomfortable. It reframes AI from a chatbot to a colleague. And as with an interaction with a colleague, a perfectly clear instruction for a thing you haven’t thought through is often a clear instruction for the wrong thing.