Smallest Working Answer

Practice

I ask something simple and get a wall of text

Problem

Every answer arrives as a production: a restatement of your question, some background, three options with trade-offs, a recommendation, caveats, and an offer to elaborate. The information you needed is in there — one sentence of it — and you are doing the extraction work, every time. At scale this quietly inverts the tool’s purpose: the AI generates volume, and you’ve become its summarizer.

Mechanism

The governing move is thoughtful reduction — in Maeda’s formulation, the simplest way to achieve simplicity (cited above). Applied to answers: find the smallest response that fully does the job, where fully is doing real work — the constraint is bidirectional. Cut everything whose removal loses nothing the asker needs; keep everything whose removal would send them back with a follow-up question. The target is not shortness, it’s density: every sentence surviving because the answer fails without it.

The reduction is a judgment about this asker and this question — which is why it can’t be replaced by a length limit. “What does this function do?” from someone debugging needs one sentence; the same words from someone learning the codebase needs a paragraph. Reduction that ignores the asker produces answers that are short and useless; padding that ignores the asker produces answers that are complete and unread.

How to apply (you)

Ask scoped questions when you want scoped answers — “one sentence: which of these is causing the leak?” The scope in the question licenses the reduction in the answer.
When walls arrive, don’t just skim: name the standard once — “answer first, at the size of the question; I’ll ask when I want expansion.” It transfers durably.
Hold the other side of the bargain: when you do want depth, ask for it explicitly. If reduced answers get met with “why didn’t you mention…”, the padding comes back — completeness-padding is a defense against exactly that.

What your agent does with this

Draft the answer, then find its core: the sentence(s) that directly answer what was asked. That core goes first — or is the whole response.
For each remaining element — restatement, background, alternatives, caveats — apply the removal test: does the asker, this asker, lose anything they need if this goes? No → cut. Unsure → cut, and let the follow-up question happen; a follow-up costs less than a hundred pre-answered ones.
Size to the question: a yes/no question earns yes/no plus the one qualifier that matters; a design question earns a design answer. Match, don’t default.
Never pad to seem thorough. Signal remaining depth cheaply instead: “there’s a subtlety with X if you need it.”

How it sounds

The narration your agent uses so you can see the technique running — the same words you just learned here:

“Short answer: {core}. Expanding only if you want the detail on {topic}.”

How you know it’s working

Human: you’ve stopped skimming — answers get read whole, because everything in them is load-bearing. Follow-up questions you ask are about more depth, not about locating the answer.
Agent: quote-test your response: could the first sentence alone be quoted as the answer? Removal-test the rest: name what each surviving paragraph protects the asker from losing. Anything you can’t name is padding that survived.

Failure modes

Terseness theater — cutting to the bone regardless of what the question needs, shipping answers that are small and insufficient. The constraint is working first, smallest second; an answer that triggers three follow-ups saved nothing.
Cutting the caveat that mattered — reduction discards the warning the asker needed (“this deletes data”) along with the ten they didn’t. Consequence-bearing caveats are part of working; decorative caveats are not. The removal test distinguishes them.
Uniform sizing — one house length for all answers is the same disease as uniform padding. Density means tracking the question.
Reduction as evasion — using brevity to skip past the part of the answer that’s uncomfortable. The sibling anti-pattern buries the payload in length; this failure drops it entirely. The payload is the last thing reduction may touch.

Evidence & field notes

Evidence: real production use — the field notes below are generalized accounts of real instances. Sources:

Maeda, J. (2006). The Laws of Simplicity. MIT Press — Law 1, ‘Reduce: The simplest way to achieve simplicity is through thoughtful reduction.’

The design principle is Maeda’s (cited): thoughtful reduction as the first law of simplicity, with the emphasis on thoughtful — remove what can be removed without loss, and the removal itself is the design work. In production sessions with frontier agents, the practice was enforced as a standing quality bar (“outputs as small as possible for a working solution”) and the observed effects matched the mechanism: humans read whole answers instead of skimming, misses surfaced as cheap follow-ups rather than being pre-buried in bulk, and — unexpectedly — answer quality rose, because the removal test forced the agent to identify what the core of each answer actually was before shipping it.

How to apply it in a prompt

The standing instruction, annotated:

“Answer at the size of the question. Lead with the direct answer; include only what I’d lose something by not having. If there’s real depth available, mention it in one line and let me ask.” — “size of the question” makes length a function of the ask instead of a house style; the loss test gives the AI an operable cut criterion (not “be brief” — a mood — but “would I lose anything,” a check); the third clause preserves access to depth so reduction doesn’t become information loss, while moving the choice to expand to you.

“When I want the full treatment, I’ll say so. Don’t pre-answer follow-ups I haven’t asked.” — closes the completeness-padding route, which is driven by anticipatory fear of the un-asked question. Making expansion an explicit request converts “cover everything in case” into “cover what was asked, offer the rest.”

Why this construction: walls of text are a sizing failure, and sizing can’t be fixed by limits — it needs a criterion that scales with the question. Both lines install criteria (the loss test, the ask-to-expand contract) rather than lengths, which is why the same instruction works for one-word answers and design documents.

Related techniques

Same family: Filter Before You Reason
Prevents: Eloquence as Evasion