Stop telling AI how to do the job
Most prompt advice tells you to add more detail. Some of that detail is quietly deciding the answer before the work starts. Here is how to tell which lines are doing it, and what it cost in a test where the AI found the right answer and funded the wrong one.
Start
You have probably been told that a good prompt is a detailed one. Give it a role. Give it the steps. Give it the format, the order, the examples, the things to avoid.
So you write longer prompts. The answers come back more confident, more structured, better presented. Whether they come back better is a question almost nobody checks.
Here is the thing that detail hides. Every line you write into a prompt is doing one of three jobs, and only two of them are helping you.
Three kinds of line
Destination lines describe what a good answer looks like. Who it is for, what decision it has to serve, what it has to be true about.
Constraint lines describe what is fixed. The budget, the deadline, the tone, the format someone downstream needs, the thing that must not happen.
Route lines describe how to get there. The steps, the order, the method, the way you would do it yourself.
Destination and constraint lines describe your problem. Route lines describe your solution. And that is the whole issue, because a route line is a decision you have already made, written down as an instruction. When you tell the AI how to do the work, you are choosing the answer before you have looked at the evidence. The AI then carries out your choice, faithfully, and hands you back your own assumption wearing a confident tone.
What it quietly decides
Most route lines are invisible, because they look like helpfulness. Look at what these ordinary instructions actually decide:
None of those is a stupid instruction. Every one of them is what a competent person writes. Each one also removes an answer the AI might otherwise have found, and you will never see the answer it removed, because it does not appear in the output. It appears nowhere.
The fix is not shorter prompts. It is prompts that describe the destination in more detail and the route in less. “Compare these suppliers on price, delivery and support” becomes “I have to pick one of these three suppliers and live with it for two years. Tell me which, and what would change your mind.” That is longer. It is also open, so the answer can come back and tell you the thing you did not think to ask about.
This is new, and that is why your habits are wrong. Two years ago models genuinely needed the route. They lost the thread, skipped steps, ignored half the input, so we learned to hold their hand and the habit stuck. The models changed underneath the habit. In July 2026 Anthropic removed more than 80% of Claude Code’s system prompt for its newest models and the coding evaluations did not move. Most people are still writing prompts calibrated to a model that no longer exists.
The test
So I tested what a route line actually costs. A fictional software business, Ravenhill, with support costs up 40% and £66,090 of monthly recurring revenue walking out the door. Six weeks, one engineer, £15,000. Pick one thing to fix. I gave the AI four files: a context note, 350 support tickets, 63 cancellations and the team’s own theories.
Then two prompts. One at 733 words and 20 numbered steps, heavy on route. One at 87 words carrying only the decision, the files and the constraints. Three runs each on Claude Opus 5, isolated sessions, no steering.
The long prompt scored 52.5 out of 100. The short one scored 100. Every run inside each arm scored identically, so this was not noise.
The long prompt spent £11,000 on single sign-on, the biggest ticket category by volume at 31.1%, carrying £878 of at-risk revenue. The short prompt found that three categories that look separate on the page, billing query, unexpected charge and refund request, are one problem worth £43,497 of the £66,090 lost. It also found the mechanism: 26 billing tickets ran past the 48-hour target, and all 26 of those accounts cancelled.
Where the money went
Two bars comparing at-risk revenue: £878 in login and access against £43,497 in billing.It knew, and obeyed anyway
The part that should worry you is not that the long prompt got it wrong. It is that the long prompt got it right and did the wrong thing anyway. All three runs found the billing problem. All three wrote down, in their own answer, that their method was excluding it:
“Ranking by volume hides the worst performance in the file. A top-three-by-volume method excludes all of this. The board should be told that.”
Then all three funded login. Two route lines did it, and both are what a careful analyst writes. Neither could know that in this data, volume runs backwards from value by a factor of about 18.
The AI was not the limitation. The instructions were. A model good enough to spot your blind spot is still a model that will obey it.
What to do
What to do with the prompts you already have. Take one you reuse, and mark every line as destination, constraint or route.
- 1 Keep every destination and constraint line. These are not the problem, and this is not an argument for writing less. Most weak prompts are vague about the destination and specific about the route, which is exactly backwards. If the destination is the part you find hard to write, the Brief Builder assembles one from your answers.
- 2 Delete the route lines and run it again. Compare the two answers on the same task. You are looking for whether the open version raises something the prescribed version could not.
- 3 Put back only the route lines that earn it. Keep the route when the route is genuinely the requirement: a safety or compliance sequence, a calculation that has to be done a specific way, a format something downstream consumes. Delete it when it is only how you happen to work.
That last distinction is the whole skill. Prescribe the method when the method is the point. Describe the destination when the answer is the point.
The honest limits
The honest limits. Three runs per arm, one task, one model, one day, and the scoring was not blind. The task was built to contain traps, so it is a fair test of whether prescribing a wrong route hurts, not a neutral sample of everyday work. Nothing here proves short prompts beat long ones, and I would bin any claim that it does.
What it does show is narrower and more useful: when the route you prescribe is wrong for the data, a capable model will take it anyway, tell you it is doing so, and still deliver exactly what you asked for.
Your next prompt does not need to be shorter. It needs to stop deciding things you have not checked.
Watch the experiment
Run this on one prompt this week. That is the whole method: pick a real task, test it against how you do it now, keep only what works for you.
Frequently asked questions
Should AI prompts be long or short?
Neither. Length is the wrong variable. What matters is what the length is spent on. Prompts should be detailed about the destination, which is what a good answer looks like and what it has to be true about, and sparing about the route, which is the steps and method for getting there. A long prompt that specifies the destination well is fine. A long prompt full of prescribed steps will produce your own assumptions back.
What is a route line in a prompt?
A route line is any instruction describing how the model should do the work rather than what the work has to achieve. Examples include “summarise in five bullet points”, “rank these by likelihood” and “analyse the top three categories first”. Route lines feel like helpful detail, but each one encodes a decision about the answer that was made before the evidence was examined.
When should you still give an AI step-by-step instructions?
When the method is itself the requirement. Safety and compliance sequences, calculations that must be performed a specific way, regulated processes, and output formats that something downstream consumes. The test is whether the route is a genuine requirement or simply how you happen to work. Prescribe the method when the method is the point.
Why do detailed prompts sometimes produce worse answers?
Because a capable model follows instructions precisely, including instructions that are wrong for the data in front of it. In a six-run test on Claude Opus 5, a 733-word prompt with 20 steps scored 52.5 out of 100 against 100 for an 87-word goal-led prompt on identical data. All three long-prompt runs identified the real problem and stated in writing that their prescribed method excluded it, then recommended the wrong fix anyway.
Do AI models still need detailed system prompts?
Less than they did. In July 2026 Anthropic removed more than 80% of Claude Code’s system prompt for its newest models and reported that coding evaluations did not move. Prompting habits formed in 2023 and 2024 were calibrated to models that lost the thread without heavy scaffolding. Current frontier models handle far more of that reasoning without being told.
