Prompting Like an Operator

Most bad output is a briefing problem, not a model problem. You would not hand a new assistant a sticky note reading "write something about our product" and expect usable work. The five parts:

• Role: who it should be. "You are a direct-response copywriter who writes for skeptical professionals."
• Task: the single verb. Write, rewrite, summarize, critique, outline, compare. One primary task per prompt.
• Context: who it's for, what they already believe, what you're actually trying to cause.
• Format: length, structure, medium. "Three paragraphs, under 150 words, no headers, no bullet points."
• Constraints: the guardrails. What to avoid, what's forbidden, what tone to never use.

Constraints are the part people skip and the part that does the most work. "Don't use the words unlock, elevate, game-changer, or journey. No rhetorical questions. Don't open with a statistic." That one line eliminates most of what makes writing read as machine-made.

ROLE: You are a [specific role] who writes for [specific audience].

TASK: [One verb. Write / rewrite / outline / critique / compare.]

CONTEXT:
- Audience: [who they are, what they already believe, what they're skeptical of]
- Goal: [what I want them to think, feel, or do after reading]
- Background: [the facts, offer, or material to work from]

FORMAT: [length, structure, medium]

CONSTRAINTS:
- Do not use: unlock, elevate, game-changer, journey, dive in, landscape
- No rhetorical questions
- No em-dash-heavy sentence rhythm
- [your own bans]

Produce the work. No preamble, no summary of what you're about to do.

"Make it warm but professional" means almost nothing. Warm like a nurse? Like a bartender? Like a pastor? Adjectives are compression: you're handing over a zipped file and hoping it unzips the same way in someone else's head.

Examples don't compress. Paste two pieces of writing you'd be happy to have written, tell the model that is the target, and the output improves more than any amount of adjective-tuning will produce. This is the single highest-leverage prompting technique there is, and it takes ten seconds.

It works in both directions. A "never like this" example is as useful as a "like this" one, often more, because it pins down the specific failure you keep getting.

The examples don't have to be yours. If you admire how a particular writer opens a post, paste the opening as a structural reference. Study the shape, don't copy the words.

Here are two examples of writing in the voice I want. Study them.

EXAMPLE A:
[paste]

EXAMPLE B:
[paste]

Before you write anything, tell me what you notice about this voice: sentence length and rhythm, vocabulary level, how it opens, how it handles the reader, what it never does.

Then, matching that analysis, write: [your task].

If you find yourself producing something more polished or more corporate than the examples, you've drifted. The examples are the target, not a starting point to improve on.

Beginners get a mediocre first output and start over with a slightly different prompt. That's the slot machine. Operators treat the first output as a draft to steer, and steering is faster than re-rolling.

Three moves cover nearly everything:

• Point at what's wrong, specifically. Not "make it better." Try: "the second paragraph is generic; it could be about any company. Make it specific to what I told you about the audience."
• Keep the good part, explicitly. "The opening line is right. Keep it exactly. Rewrite everything after it."
• Ask for options, not one more attempt. "Give me three different openings, each taking a different angle." Choosing is faster and better than describing.

One more, underused: ask it what it needs. "What would you need to know from me to make this genuinely good?" The questions it asks are frequently the ones you hadn't thought to answer.

That draft isn't there yet. Don't start over. Revise.

KEEP EXACTLY: [the part that works]
FIX: [name the specific problem, e.g. "paragraph 3 is generic," "the close is weak," "it sounds like a brochure"]

Then give me THREE versions of the fixed section, each taking a genuinely different angle. Label them A, B, C and add one line under each explaining the angle you took.

Don't hedge by making them nearly identical.

Models are noticeably better at spotting problems than at avoiding them the first time. You can exploit that directly: ask for a draft, then ask for a hard critique against explicit criteria, then ask for a revision that addresses the critique. Two passes, one extra minute, consistently better output.

The trick is criteria you specify. "Is this good?" gets you flattery. "Score this 1 to 10 on: does the opening earn the second line, is every claim specific enough to be falsifiable, would a skeptical reader roll their eyes anywhere, does it sound like the voice examples" gets you a genuinely useful teardown.

Add one instruction that changes everything: "Be harsh. Assume this draft is going out under my name to people whose respect I need." Models default to agreeable. You have to explicitly turn that off.

Critique the draft above before revising anything.

Score it 1-10 on each, with a one-sentence justification:
1. Does the first line earn the second?
2. Is every claim specific enough to be argued with?
3. Where would a skeptical reader roll their eyes?
4. Does it match the voice examples, or has it drifted corporate?
5. What's the weakest sentence, exactly?

Be harsh. This is going out under my name to people whose respect I need. Flattery costs me more than criticism does.

THEN, and only then, rewrite it addressing every point you scored under 8.