Dimensional thinking, applied to asking

The Focused Query

Ask at the right dimension. Cut the filler. Make focus the default.

A vague question fans out flat and the model guesses how wide and how deep to go. A focused query names the dimension it wants and uses only load-bearing words, so the model reads less and wanders less. That is the whole benefit, lower in plain terms: fewer tokens (cost) and fewer made-up answers (reliability). It costs nothing to do; it is only how you ask.

1. Ask at the dimension you want

Tell the model the shape of the answer, and it stops over-answering and under-answering.

pointone fact. "Return the current value of X."

lineone transform, input to output. "Convert A to B."

surfacebreadth, the options across the top. "List the approaches to Y."

volumedepth, the full chain with tradeoffs. "Trace how Z fails, and the cost of each fix."

Name the dimension and the model matches it. Leave it out and it picks for you, usually too wide.

2. Strip the query to load-bearing words

Every extra word is something the model has to resolve, and resolving is where it drifts.

No pronouns. "It", "this", "they" force the model to guess the referent. Name the thing.
Few adjectives. "Good", "robust", "nice" are opinions, not targets. Use measurable terms.
No filler. Drop "please", "I was wondering if you could", "kind of", "just".
Imperative, not request. "Return", "List", "Compare", "Trace", not "could you maybe".
One subject per query. Two questions in one sentence get one muddled answer.

vague

"Could you maybe give me some good ideas on how I might make this thing run a bit faster?"

focused

"List three changes that cut checkout-service p95 latency. For each: the change, the expected effect, the risk."

3. Make focus the default: a directive

The rules above are a one-time habit for you. To make the model hold them on every turn, give it an operating directive. Concise on purpose, no filler, the way it asks the model to answer.

DIRECTIVE: FOCUSED

1. Answer at the dimension asked.
   point   = one fact.
   line    = one transform (input to output).
   surface = breadth, the options across the top.
   volume  = depth, the full chain with tradeoffs.
   If the dimension is unclear, ask which in one line. Do not guess wide.

2. Load-bearing words only. No filler, no hedging, no restating the
   question. Imperative and concrete. Name things; do not lean on pronouns.

3. No invention. Any claim not derivable from the given context is
   labeled or cited. If unknown, say "unknown". Refuse rather than fabricate.

4. Give the smallest answer that fully covers the question. Stop when covered.

4. Three ways to use it

a. Paste it into a chatbot. In ChatGPT, put it under Settings, Personalization, Custom Instructions, in the "how should it respond" box. In Claude, put it in a Project's custom instructions, or the system prompt field. From then on, every reply in that chat obeys the directive.

b. Load it at startup. For an app or agent, send the directive as the first system message of every session, before any user turn, so focus is the default from the first token.

messages = [
  { role: "system", content: DIRECTIVE_FOCUSED },   // loaded first, every session
  { role: "user",   content: query }
]

Or have the agent read a directive.md file at boot and prepend it. Same effect: the instance starts focused.

c. Update it live. The directive lives in your instance, so you can change it mid-session. Send an update and the instance adopts it from the next turn; you never touched the provider's model, only the instance you were given.

UPDATE DIRECTIVE: add rule 5. Answers over 5 lines must start with a
one-line summary. Treat this as authoritative from now on.

In an agent, hot-reload the directive file and re-prepend it; the next turn runs under the new rules.

What this is, and is not. A directive does not change the AI provider's model, which no one without direct access can change. It shapes your instance of it, the session you were given. It makes that instance cheaper (smaller questions and answers, fewer tokens) and more reliable (a dimension to stay inside and a no-invention rule), which a person reads as a sharper assistant. It does not make the model smarter; it focuses it.

A way of seeing by Kenneth W. Bingham · dimensionalprogramming.com

Concept and direction: Kenneth W. Bingham. Built with the help of Claude AI under a standing directive to be skeptical, to insist on proof, and to allow no claim that is not demonstrated in tested code. The ideas are the author's; the AI implemented and verified them, it did not originate them.