GUIDE 05
Prompt for outcomes,
not steps.
A step-by-step method for getting much more out of an AI model by changing what you ask for. Most people use AI to run a process they have done by hand for years, one instruction at a time. This guide replaces that with a four-block brief: the job, why it matters, the guardrails, and what done looks like. The model works out the steps, and it usually finds a faster route than the one you would have given it.
What you need.
- Any current AI model with a chat or agent interface. The method was written against Claude, and the prompts here use its phrasing, but the four blocks work anywhere.
- One real task you have been doing by hand, or walking the AI through step by step. A reconciliation, a report, a clean-up, a summary. Something with a clear result.
- If the task involves files, an agent that can read a folder, such as Claude Code or a desktop AI app with file access. A chat window works for everything else.
Catch yourself writing steps
The symptom is easy to spot once you know it. “Remove the blank rows. Now move this column so it lines up with the other report. Now filter for anything over 90 days.” One instruction at a time, the same clicks you would otherwise make by hand, just dictated.
It feels productive, and it is a little faster than doing it yourself. It also throws away most of what the model can do. An AI with tools can write and run code, read a whole folder at once, and try a route you would never have taken. Walk it through your manual steps and you forbid all of that.
Before the next task, write down the result you want in one sentence. If the sentence starts with a verb like open, filter, or copy, you are describing steps. Rewrite it until it describes the end state.
- One sentence that describes the result, with no steps in it.
Give it the whole job
Hand over the complete task in one request and let the model run it end to end. Current models are built for this; managing them stage by stage mostly adds your time to theirs.
A worked example. A utilities audit: checking that costs paid upfront had been correctly billed back, across a year of ledgers and rent rolls where the amounts did not always line up one to one. Given the folder and the outcome, an agent produced a reconciliation schedule with items flagged for review, a summary of findings, and a list of follow-up questions for the vendor. It took 15 minutes and a few dollars of usage, against about two days by hand. The flagged items still needed a person to look at them, but that is a different job from building the reconciliation.
Write the job as one or two sentences that describe that kind of result: the thing that should exist when the model is finished.
- The Job block: what you want done, as an outcome, in one or two sentences.
Say why
A model that knows why a task matters makes better judgment calls on everything you did not specify. Tell it the larger task this is part of, who the output is for, and what they will do with it. Three short clauses are enough: “I am working on [the larger task] for [who it is for]. They need [what the output enables].”
This is the block people skip, and it is the one that changes the most about the result. A summary for a board is a different document from a summary for an engineer, and the model can only know that if you say so.
- The Why block: the larger task, who it is for, what the output enables.
Set the guardrails
Guardrails are the few things the model must not do, plus a rule for how to handle the decisions in between. Three lines cover most tasks.
- The scope: only touch these files, this folder, this range. Leave everything else alone.
- The must-nots: anything that must not change, be sent, or be deleted. Name them.
- The judgment rule: make routine judgment calls yourself, and ask me only if the answer would change the whole result.
That last line is what lets the model finish without a dozen check-ins, while still bringing you the one question that matters. Guardrails are reasons and boundaries, not a list of every forbidden action; see step seven for why.
- The Guardrails block: scope, must-nots, and the judgment rule.
Define done
If you say what finished looks like, you get it. If you do not, you get the model’s guess. Define done in three parts: how you will both know it is finished, how big the deliverable should be, and how the model should report back.
For size, name sections or a word count, or say “as short as covers the substance”. For the report, ask for where the result is and three short bullets on what was done, and nothing more. That last instruction is what keeps a long task from ending in a wall of narration.
Surface the unknowns here too. If there is something you do not know yet, say so, and say whether the model should decide, ask, or leave it open.
- The Done Means block: the finished test, the size, and the three-bullet report-back.
Let the model write the brief
You do not have to fill the four blocks in yourself. Two shortcuts, and both produce better briefs than most people write by hand.
The first is to give the model the context, paste the template, and ask it to fill the template in for you. Read what comes back, correct anything it assumed, and run it. The second is to have the model interview you first: it asks the questions a good analyst would ask before starting, you answer in plain language, and it writes the brief from your answers. Use the interview for anything bigger than an afternoon.
Before you do anything, interview me about this task. Ask the questions a good analyst would ask before starting: what the result is for, who will use it, what is in and out of scope, what must not change, how we will know it is finished, and anything I have not thought about. One question at a time, up to eight. Then write the brief in four blocks: THE JOB, THE WHY, THE GUARDRAILS, DONE MEANS. Show me the brief and wait for my yes before starting.
TIPThe one-line version, for a quick task: paste the four-block template, add two sentences of context, and say “fill this in for me, then run it”.
- A completed brief the model wrote and you corrected.
Give reasons, not rules, in your standing instructions
Every AI tool has somewhere for standing instructions: a project file for a coding agent, custom instructions in a chat app, an identity file for a personal agent. Most of what people write there is a list of rules: never do this, always do that. Rules get pattern-matched, and a rule that does not quite fit the situation gets applied anyway, or ignored.
Replace each rule with its reason. “Never use an em dash” becomes “em dashes read as machine-written in my voice; use a full stop or a comma instead”. “Always ask before sending” becomes “sending is irreversible and goes out under my name, so stage a draft and ask”. The model can then handle the cases the rule never anticipated.
Here are my standing instructions: [paste them]. For each rule written as a prohibition or a command, rewrite it as the underlying reason, in one sentence, so that a reader who understands the reason would make the right call in a situation the rule did not anticipate. Keep anything that is a genuine hard limit (money, deletion, sending, safety) as a hard limit, and say why it is one. Show the before and after side by side.
- Standing instructions where each line carries its reason.
Stop adding “double-check”, and fix the voice once
Two habits from older models are now wasted words. The first is telling the model to double-check its work, verify everything, or review its own output before finishing. Current models already do this, and the extra instruction mostly adds length and hedging. If an output genuinely matters, ask for the evidence and the checks that were run, not for reassurance.
The second is correcting the tone in every conversation. If you keep asking for shorter answers, no headers, or plainer language, put that in the standing instructions once, or in an output style if your tool has one, and stop repeating it.
One more habit worth adopting: when a task involves documents, put the documents before the question. Context first, then what you want done with it.
- No “double-check” lines in your prompts.
- Your voice and format preferences written down once.
Keep the template and reuse it
The four blocks are the whole method. Keep the template somewhere you can paste it from, use it for every task that would take more than a few minutes by hand, and let the model fill it in when you are short on time. After a few weeks, the shape will be in your head and you will write outcome-first without the scaffold.
THE JOB [What you want done, as an outcome, not as steps. One or two sentences.] THE WHY I'm working on [the larger task] for [who it's for]. They need [what the output enables]. THE GUARDRAILS - Only touch [the scope]. Leave everything else alone. - [Anything that must not change, be sent, or be deleted.] - Make routine judgment calls yourself. Ask me only if the answer would change the whole result. DONE MEANS - [How we both know it's finished.] - Keep the deliverable to [sections, a word count, or "as short as covers the substance"]. - When you finish, tell me where the result is and give me 3 short bullets on what you did. Nothing more.
- The template saved where you can paste it.
- One real task run with it, and the result compared with how you would have done it by hand.
Where this comes from.
The four-block template is the one from an Alpha Infra post in September 2026, adapted from Ben van Sprundel’s Ben AI video on Anthropic’s prompting guidance for Claude 5 models, which is also the source of the rules in steps two, six, seven, and eight: give the whole job, interview yourself first, reasons not rules, avoid double-checking, and fix the voice once. Anthropic’s own prompt engineering documentation is the primary source and worth reading in full.
The utilities audit in step two is a real piece of professional work, described without the client, property, vendor, or figures involved; the times are the actual run. The interview prompt, the rewrite prompt, and the step order are this guide’s own.
LET’S TALK
Would rather have it built
with you?
Bring the idea, or the job you want off your desk. We can run the same steps together, or I can build it and hand it over.