Negative Prompting
Explicitly tell the model what not to do, and where to stop. A capable model acts on its own more readily — drafting, editing, refactoring, "getting creative" past the literal ask — so the highest-value sentence in many prompts is the one that bounds the action. Negative prompting states the deliverable and the stop condition: "report what you find and stop; don't fix, send, edit, or delete anything until I say go; do the simplest thing that works and skip cleanup I didn't ask for." The mental model the source offers is an intern: you tell a new intern the specific things not to do because they don't yet know the process and will otherwise over-reach.
Notably the source observes this technique's effectiveness has changed with model generation: negative prompting "used to not work as well" than positive specificity, but on recent models it "tends to work pretty well" — a best practice whose value is tied to the current model landscape, not a timeless law.
Claims
- Bounding a capable model's action — stating what not to do and where to stop — prevents the unrequested over-reach that comes from a model acting readily on its own. principle — durable: the more agentic and autonomous the model, the more the binding risk is unwanted action, so an explicit stop condition is load-bearing.
- Negative-prompt the boundaries: name the deliverable, then the "do not" (don't fix/send/edit/delete until told; don't add features; do the simplest thing). best practice — context: agentic or side-effect-capable settings (code, email, files) where an over-eager action is costly to undo.
- Negative prompting is more effective on recent models than it was on older ones. observation — the source reports its own reversal of opinion; a claim about the current model generation, subject to change, not a durable property.
- A stop-and-report boundary is the same pattern as read-only supervision — "the deliverable is your assessment; report and stop." observation — connects negative prompting to human-in-the-loop gating; the boundary keeps the human's approval in the loop before any mutation.
Related
- Intent Context — intent says where you're going; negative prompting says where to stop. Paired lighter-touch steering.
- Evidence-Gated Completion — "report and stop" and "prove it before you claim done" are both ways of gating a model's action/claims on the operator's terms.
- Agent Supervision — a "report, don't act until I say go" boundary is human-in-the-loop control encoded into the prompt (the "AI is like an intern" framing recurs here).
- Agentic Simplicity — "do the simplest thing that works, skip cleanup I didn't ask for" is the simplicity discipline stated as a boundary.
- Distillate: How Anthropic Engineers Actually Prompt Fable 5 — six habits for a smarter, lighter touch