Concise Prompting

For a capable model in a good environment, a short instruction steers as well as an exhaustive rulebook — and an over-detailed prompt can actively make the answer worse. This inverts the older "more context is always better" heuristic. The source's reasoning: when the model is intelligent and wrapped in good context, tools, and skills, a single clear instruction "can now steer just as well as spelling out most of the rules by name," while reusing last year's over-prescriptive prompt bloats the input and degrades output. Instead of enumerating "Rule 1: be concise. Rule 2: use bullets. Rule 3: no jargon…", say something like "Lead with the outcome, keep it simple, and pause only when the work truly needs me."

The key reconciliation — and the reason this is not a contradiction with Intent Context ("give it the why") — is add the why, cut the rulebook: the why is real signal the model cannot guess, so it stays; the enumerated procedure is redundant with what a capable model already does, so it goes. Concise prompting is Agentic Simplicity applied at the prompt level: complexity (here, prompt length/rules) should climb only when it demonstrably helps, and for a strong model it often hurts.

Claims

For a capable model, a short instruction can steer as well as an enumerated rulebook — instruction length is not proportional to control. principle — durable: once a model reliably follows one clear direction, added rules are redundant procedure that costs tokens and can distract, so more instruction ≠ more steering.
Cutting an over-detailed prompt can improve output on a strong model; reusing an old, verbose prompt can make the answer worse. best practice — context: capable, well-scaffolded models (the source attributes this to Fable 5's guidance); on weaker or unscaffolded models more explicit instruction may still help, so "say less" is contingent on model capability and environment.
Reconcile brevity with intent: keep the why (signal the model can't infer), cut the rulebook (procedure it already follows). best practice — context: shortening prompts without losing steering; the discriminator is "is this signal or redundant procedure?" See Intent Context.
Prefer encoding durable guidance once in skills/agents/memory files over restating it in every prompt. best practice — context: reusable setups; standing config keeps individual prompts short while preserving the behavior. See Reusable Workflow Library, Evidence-Gated Completion.
The output-side companion: a standing skill can make the model write less code for the same result, and lower cost is a side effect of writing "only what the task needs, never golfed" — not a "fewest tokens" target. best practice — context: a second source (Chase AI) demonstrates the ponytail skill, whose README rule is "write only what the task needs, and never cut validation, error handling, security, or accessibility." Its benchmark (vs a no-skill baseline) claims −54% LOC, −22% tokens, −20% cost, −27% time at 100% safety, biggest where there's a real over-build trap; the source measured ~22% cheaper running it on Fable 5 at medium effort. Caveat stated in the source's own chart: a terse reasoning model that spends thinking tokens deliberating the guidelines can go the other way (tokens up), and a terse-prose control (caveman) cut LOC but raised tokens/cost — so an output-brevity skill is contingent on the model actually following it, not a guaranteed saving. Figures reproduced as the source presents them, flagged for grounding.

Intent Context — the paired half: keep the why, cut the rules. "Say less, but say the right thing."
Agentic Simplicity — the parent discipline; this is that thesis applied to prompt length. Already cross-linked from there.
Reasoning Effort Control — concise prompting tunes how much you say; effort control tunes how hard the model thinks. Both are lighter-touch for a capable model.
Reusable Workflow Library — encoding standing guidance once (so prompts stay short) is the reuse discipline.
Safety Routing Fallback — "stop asking it to show its reasoning" is a concrete instruction to cut, which also avoids a reroute.
Reasoning Effort Control — the effort dial and output-brevity skills are two levers on the same bill; both cut token spend for a capable model.
Distillate: How Anthropic Engineers Actually Prompt Fable 5 — six habits for a smarter, lighter touch
Distillate: Make Fable 5 80% Cheaper (& Other Usage Cheat Codes) — five levers for spending a premium model less — the output-side companion: token- reduction skills (ponytail, caveman) that make the model write leaner code.
Distillate: Building Great Agent Skills: The Missing Manual — the same "say less, cut redundant procedure" instinct applied to a skill's own text (Minimal Skill Surface, Skill Pruning): shave every word, and compress paragraphs into Leading Words.

Concise Prompting

Claims

Related

Linked from