What you came for

The seven techniques that work.

Direct answer, for whoever has ten seconds: the best prompting technique is the one that survives the following turns — directed iteration — and its natural extension is called session steering. The seven moves follow, honestly: no miracle recipes, no "secret prompts" — seven gestures that genuinely improve a request, each with the exact limit past which it can do nothing more. The limit is not a flaw: it is information. It tells you where the next level begins.

The bench

Seven moves, three beats each

1. Subject precision

The move: Not "tell me about X," but "I want to examine this precise aspect of X, in this context, for this reason." Precision forces you to clarify your own request before phrasing it.
What it gets you: The model knows what kind of answer is expected; you eliminate the generic output calibrated to displease no one.
Where it stops: A precise question inside an undemanding session falls back to average by the second answer. Precision does not maintain itself.

2. Stating the purpose

The move: Say what the answer is for: understanding, drafting, deciding, publishing. Each purpose calls for a different shape.
What it gets you: A usable answer rather than a correct one — the difference between a text you read and a text that serves.
Where it stops: A declared purpose doesn't prevent drift: over the turns, the model recalibrates on what makes you nod, not on what you declared.

3. Examples

The move: Show one or two cases of the expected result — format, tone, level of detail — rather than describing them abstractly.
What it gets you: The most reliable precision gain in the repertoire: the model imitates better than it interprets.
Where it stops: Examples freeze. For a repetitive task that's a strength; for an exploration it's a rail — the model reproduces the form instead of thinking the substance.

4. Explicit constraints

The move: Forbid in advance what you refuse: "no hollow lists, no unrequested caveats, no rephrasing my question before answering."
What it gets you: You defuse the factory reflexes that would have parasitized the exchange — half the comfort of reading comes from this.
Where it stops: Constraints erode. A few turns in, the reflexes come back one by one; nobody warns you, and the initial prompt can do nothing about it.

5. Role and positioning

The move: Less "act as an expert" — a worn-out formula — than signaling your own level: precise vocabulary, a demand for depth, an announced refusal of compliance.
What it gets you: The response regime rises: the model produces for a reader who will detect surface.
Where it stops: A role is a costume. It changes the answer's outfit, not its verification: the "expert" invents references with the same confidence as the base assistant.

6. Directed iteration

The move: Don't accept the first answer: ask again while pointing precisely at what's missing — "you answered on X, not on Y."
What it gets you: This is the prompt technique closest to the level above: without knowing it, you have begun to steer.
Where it stops: Iterating without knowing what to watch is correcting at random. The signals that call for a retake — there are seven — appear in no prompt: they are learned.

7. Requesting verification

The move: Demand sources, ask for the degree of certainty, have what isn't verified marked "unverified."
What it gets you: A first net: the most fragile claims sometimes flag themselves.
Where it stops: Asking the model to verify itself is measuring the instrument with the instrument. The verification that counts is external — and it is your responsibility, not the model's.

The ceiling

What these seven limits have in common

Reread the "where it stops" column: each limit is the same one, seen from seven angles. A prompt acts once; a session lasts. Everything a prompt installs — precision, constraints, regime — degrades over the turns if no one maintains it, and degrades in silence: the fluency stays perfect while the substance leaves. The prompt is a point impulse. The session is a field.

The consequence fits in one sentence, and it is the whole diagnosis: an excellent prompt in a bad session yields an average result; an average prompt in a well-steered session can produce a real elevation. The lever is not in the phrasing. It is in the steering — the opening, the vigilance, the retake, the close.

Move up a level