Saying the unsaid: vision, scope and taste

Why your first attempt is always disappointing

You know what you want. You can picture it. You ask for it, and what comes back is technically what you said and completely not what you meant.

That gap has a name in every creative industry: the brief was underspecified. You handed over 20% of the picture and the other 80% got filled in with the world's average answer. Not a wrong answer - an average one. Average is the enemy, and average is the default.

Climb all four rungs. Watch the three runs stop disagreeing. That convergence is the entire goal: not a better model, a smaller gap between your head and your words.

The unsaid stuff, catalogued

When you ask a person for something, an enormous amount travels silently: who you are, what your company is like, what happened last week, what you'd obviously never do. None of that travels to a model. Here is what goes missing most often, in rough order of damage:

1. The actual purpose. You asked for a summary. You wanted it because you have four minutes before a meeting with a sceptical CFO. Those are two completely different summaries. Say why.

2. The audience. Experts or beginners? Friendly or formal? People who already agree, or people who need convincing?

3. What "good" means here. Fast and rough, or careful and complete? Both are correct answers to "write me a plan" and you only wanted one.

4. What you'd hate. Emoji. Buzzwords. Bullet points. Being told the obvious. Say it - bans are the cheapest quality control there is.

5. The constraints of your actual life. Budget, deadline, the fact your team is two people, that you can't change the database. Advice that ignores your constraints is worse than no advice.

6. What's already been tried. Otherwise you get suggestion #1 again, the one that failed in March.

Scope: the skill that separates good and bad AI users

A scope is a fence. It says: inside here, do your best; outside here, stop and ask me.

┌──────────────────────────── SCOPE ─────────────────────────────┐
│  DO      rewrite the three product pages                        │
│  DON'T   touch pricing, invent features, change the guarantee   │
│  ASK ME  if a claim needs a source I haven't given you          │
│  DONE    each page under 200 words, one CTA, my banned list     │
└─────────────────────────────────────────────────────────────────┘

Four lines. DO, DON'T, ASK ME, DONE. That template works for a chat message, a project brief and an autonomous agent, unchanged. The ASK ME line is the one people forget and the one that prevents the most damage - it converts a confident guess into a five-second question.

How to say the vision when you can only half-see it yourself

Often you cannot describe what you want because you'd only recognise it on sight. That's normal - it's how taste works. Three moves that work anyway:

Use comparisons. "Like a Wirecutter review, not like a press release." "The tone of a good podcast intro." Comparisons carry a thousand implicit rules and cost you eight words.

Describe the feeling of the reader. "They should feel relieved, not impressed." Extraordinary how much this steers.

Point at the negative space. "Not this - this feels salesy." You are allowed to define by rejection. Do a first pass, react to it, feed the reaction back. You will describe your taste faster by critiquing three drafts than by introspecting for an hour.

Iteration is not failure, it is the method. The pros do not write one perfect prompt. They write a decent one, look at what came back, and each time convert a reaction into a rule they add to the brief. After four rounds you own a brief you can reuse forever.

The reusable brief - your actual takeaway

Once a brief works, keep it. This is what companies mean by "prompt library", and there is nothing stopping you doing it in a notes app:

CONTEXT   who I am / what this is for
AUDIENCE  who reads it and what they know
GOAL      what should happen after they read it
VOICE     two comparisons + three banned words
FORMAT    structure and length
NEVER     the hard lines
EXAMPLE   one thing I loved

Fill it once per recurring task. Paste it forever. You have just built the same thing a professional AI system builds - a system prompt - with no code at all.

Where things stand (as of 2026)

Nothing about scoping, briefing, or taste is model-specific, which is exactly why this lesson doesn't go stale the way pricing or benchmark numbers do - a clearer brief helps every model, past, present, and future. What has genuinely improved is that models are somewhat better at asking a clarifying question instead of silently guessing when a request is ambiguous, though this is still inconsistent and not something to rely on. What hasn't changed at all: models still have no access to your unstated context, your company's history, or what you'd hate, unless you say it. If you read a claim that "the new model just understands what you mean now," treat it with real skepticism and test it yourself rather than assuming it applies to your specific kind of ambiguity - and note the date, because this is one of the areas most prone to overclaiming.

How to read the docs and look things up

There isn't much of a "spec" for taste and scoping - the useful primary sources here are worked examples from people who actually build with these tools regularly (engineering blogs showing real prompts and iterations), rather than generic listicles of "50 prompt tips." A giveaway that you're looking at marketing rather than a real technique: it promises a shortcut around iteration entirely, when in reality (as this lesson argues) iteration is the method, not a failure. To test whether a "prompting framework" someone is selling actually helps: apply it to a real fuzzy task of yours, look at the result with your own reaction the way you did in the warm-up, and see if it saved you a round of back-and-forth. Keep your reusable briefs in a dated notes file - unlike most AI trivia, these barely go stale, because they encode your taste, not the model's current quirks.

Where this leaves you

You now know, honestly, what these things are:

• Traditional software is exact rules a human wrote, brilliant and brittle.
• A language model is a next-token machine with no memory, fluent and unreliable.
• An agent is the loop that joins them: the model decides, the code acts and enforces.
• The limits are compounding errors, cost curves, context rot, retrieval quality and trust.
• Prompting is briefing. Scope is a fence. Taste is stated as bans and comparisons.

That is a more accurate mental model than most people who talk about AI for a living have. The stack will change every six months. None of the six things above will.

If you want to see the machinery from the other side - the actual parser, the actual agent loop, the actual retrieval systems, running in your browser - the engineering book next door starts at chapter zero and assumes nothing.

Check yourself

• Write the four-line scope fence for something you asked an AI to do this week.
• Which of the six "unsaid" items do you leave out most?
• Why is critiquing three drafts a faster route to your own taste than describing it up front?

Recap: what changed in your head

• You know your first draft disappoints because the brief was underspecified, not because the model is weak.
• You know the six things that travel silently between people but never reach a model.
• You have a four-line scope fence (DO / DON'T / ASK ME / DONE) you can reuse anywhere.
• You know comparisons, feelings, and negative space describe taste faster than adjectives do.
• You know iteration is the method, not a failure - reacting to drafts builds your brief.
• You have a reusable brief template you can fill in once and paste forever.