Field Note 02 — Less Is the Instruction: A New Way to Prompt Fable 5

abstract

For years, getting good output from an AI model meant building guardrails: assign a role, specify the format, list the rules, forbid the failure modes. That habit was rational when models were weak, the same way bumpers are rational when the bowler is six years old. The engineers who build today’s frontier models now describe the opposite discipline: the model has read more about your domain than any person you could hire, and a prompt stacked with constraints caps its output at the ceiling of your own imagination.

This note describes the method I use with Fable 5 across every Modven build: name the destination, show your taste with one example instead of ten rules, ask for genuine exploration, then iterate against the draft rather than retrofitting constraints into a longer prompt. I include specific before-and-after examples from marketing, analysis, and software, and the short list of constraints that still matter.

The claim is not that instructions are dead. It is that the instruction budget should be spent on where you are going and why it matters, not on how to take each step.

01 · where the guardrail habit came from

Early models earned their reputation for wandering. They lost the thread, invented facts, and ignored formats, so practitioners responded sensibly: they fenced the model in. Prompt-engineering culture grew from that fencing, and by 2024 a “good prompt” meant a role, a persona, a numbered list of rules, three examples, a forbidden-words list, and a required output template. Entire libraries of these prompts circulated like recipes.

The habit worked, and that is exactly the problem. Habits that work stop being examined. The guardrails stayed as the models changed underneath them, and most people are now managing a senior collaborator with a checklist written for an intern.

02 · why the logic inverts with Fable 5

The engineers at Anthropic who work with these models daily make a point that sounds like marketing until you test it: the model knows more than we possibly could. Not more than you about your customers or your company, but more than any one person about the accumulated craft of writing, analysis, code, and design that it learned from. When you hand a model like that fifteen constraints, you are not preventing failure anymore. You are pre-deciding the answer, and your answer is rarely the best one available.

Here is the image I use with clients. A constraint-stacked prompt is a recipe handed to a great chef: you will get exactly the dish you described, cooked competently, and you will never find out what the chef would have made. An open prompt with a clear destination is describing the evening instead: who is coming, what the occasion is, what you loved last time. The chef’s knowledge finally gets used. The same applies to a GPS: you enter the destination, not the turns. When you dictate every turn, the map can no longer route around traffic.

figure 1 · the ceiling problem

The circle is what the model can do; the box is what your constraints allow it to show you. The dot on the right usually sits outside the box on the left, which is the entire argument of this note.

03 · the method

Four moves, in order. First, name the destination: the outcome, the audience, and the stakes, in a sentence or two, the way you would brief a partner you trust. Second, show your taste instead of writing rules: one example of something you love, with a word about why, teaches more than ten adjectives. Third, ask for genuine exploration: “give me the version you’d bet on, and one direction I wouldn’t think of” produces range that no constraint list ever will. Fourth, iterate against the artifact: react to the draft the way an editor reacts to a manuscript, specifically and locally, instead of returning to the prompt to bolt on more rules. The draft is the conversation now.

example 1 · a landing-page headline

before · the guardrail habit

Act as a world-class direct-response copywriter.
Write a headline for our landing page. Rules:
1. Max 9 words. 2. Include the word "effortless".
3. Mention AI. 4. Address pain point of wasted time.
5. No questions. 6. No jargon. 7. Title case.
8. Create urgency. 9. Sound premium, not salesy...

The output will contain the word “effortless,” mention AI, and be forgettable, because the answer was pre-decided by someone who is not a copywriter.

after · destination and taste

Here's our company and who buys from us: [two
sentences]. The landing page has one job: make a
$3-50M owner feel understood in five seconds.
A headline I love for its confidence: Stripe's
"Financial infrastructure for the internet."
Give me the headline you'd bet on, one riskier
direction, and tell me what you'd put under each.

The model gets the job, the buyer, one calibration point for taste, and permission to explore. The second turn is where you edit.

example 2 · a market analysis

before

You are a senior strategy analyst. Analyze this
market using Porter's Five Forces. Format: exec
summary (100 words), then each force as a section
with 3 bullets each, then a 2x2 matrix, then
exactly 5 recommendations ranked by impact...

You will receive a competent Porter’s Five Forces, whether or not Porter is the right lens, because you chose the lens before the analysis.

after

Here's our data and the decision in front of us:
[the decision, the deadline, what happens if we're
wrong]. Before you analyze anything, tell me what
you'd want to know and which lens fits this
decision. Then give me your read, committed to a
recommendation, and the strongest case against it.

The model picks the frame, argues both sides, and commits. You judge the reasoning instead of the formatting.

example 3 · the iteration turn, which is where the real prompting happens

The draft is 80% there. Three reactions:
— The middle section hedges. Commit to one recommendation
  and defend it.
— The tone drifts corporate in paragraph 4; the rest sounds
  like a person. Match the rest.
— The pricing point is the strongest thing here and it's
  buried. Lead with it.

Notice what this is not: it is not a new list of upfront rules. It is an editor reacting to a specific draft. Two or three turns like this outperform any single mega-prompt I have ever seen, because each turn spends your judgment where it is actually needed.

04 · what still deserves a constraint

Some constraints are real, and the method is not an excuse to drop them. Facts the model cannot know belong in the prompt: your numbers, your customers, your history, the things that make you you. Contracts on the output belong there too, when a machine consumes the result: a JSON schema, a file format, a required legal disclaimer. Brand-locked language stays: if your company never says “cheap” and always says “founding equity,” say so. The rule that separates these from the guardrail habit is simple: constrain the output contract and the facts, and stop constraining the thinking. If a rule exists because you are afraid the model will be dumb, delete it. If it exists because the world requires it, keep it.

05 · what changes when you work this way

Three things, in my experience across every build this year. The ideas get better, because you finally see options from outside your own box, and the best one is frequently something you would not have specified. The work gets faster, not because each prompt is shorter but because you stop paying the mega-prompt tax: writing them, maintaining them, and debugging the strange outputs they cause. And your own judgment gets sharper, because iterating against drafts is a muscle: you learn to say what is wrong specifically, which is the same skill that makes someone a good editor, a good manager, and a good client. The people who get the most from these models are not the best prompt writers anymore. They are the best reactors.

Less is the instruction.

This is how every Modven system gets built.

This is how every Modven system gets built.

Start with an Edge Brief

Thanks — I’ll be in touch.