Stop Telling AI What to Do. Tell It What Done Looks Like.
The full series
- The Prompt Is the Least Important Part
- Three Files That Replace a Chief of Staff
- You Don't Have a Productivity Problem. You Have a Setup Problem.
- Stop Telling AI What to Do. Tell It What Done Looks Like. (you are here)
- The _MANIFEST.md: Five Minutes That Save Hours
- Plugins Are Composable. Stack Them.
- The Monday Briefing That Runs Without You
- Custom Skills: Build Once, Run Forever
- AI as a Team, Not a Chatbot
- Cowork Is Not Chat. Stop Using It Like Chat.
Most people write prompts the way bad managers give assignments, with every step prescribed and every decision pre-made. The people getting the best results from Cowork describe the destination and let the agent find the route.
There are two kinds of managers in every organization.
The first writes task assignments that read like instruction manuals, with every decision pre-made, every step prescribed, and every judgment removed from the person doing the work. The employee receiving that assignment has no room to think and no latitude to improve on the instructions. They follow the steps and produce exactly what was described.
The second manager writes three sentences: here is what the finished product looks like, here are the boundaries, use your judgment for everything else.
The second manager gets better work, every time, because delegation invites the kind of thinking that step-by-step instructions prevent.
Delegation. Everything else is micromanagement.
Most people prompt Cowork like the first manager, writing 70 words of step-by-step instruction and pre-making every decision from column ranges to file names to word counts for bullet points. The output follows your instructions precisely, but it doesn't think. No insight the agent noticed but wasn't asked about, no anomaly it spotted but had no instruction to flag, no formatting improvement it could have made but wasn't permitted to try.
Compliant output and good output are not the same thing.
Describe the destination, not the route
The shift is straightforward, and it mirrors how good delegation works with human teams.
Stop describing the process and start describing the deliverable.
A micromanager prompt for a quarterly analysis tells the agent exactly which spreadsheet to open, which columns to calculate, which format to use for headers, how many bullet points to write, and what to name the output file. Every step dictated, every choice removed.
A delegator prompt for the same task says: analyze the Q4 revenue data in this folder, identify trends and anomalies worth attention, deliver a one-page executive summary with supporting data, flag anything that looks off, and save to outputs.
Half the words, twice the latitude.
The delegator prompt describes what done looks like, not how to get there. It sets a quality bar and handles uncertainty, then trusts the agent to figure out which columns matter and which format serves the reader best.
We've watched this pattern hold across sales, consulting, product management, and operations. The delegated version consistently produces output with observations that weren't requested, formatting decisions that fit the content better than any prescribed template, and quality that matches or beats the micromanaged version.
Describe the destination. Not the route.
Three questions
Every effective task prompt in Cowork answers three questions, and the approach works whether your task takes five minutes or five hours.
What does done look like? A one-page executive summary, a client-ready proposal, a comparison table with a recommendation. Name the deliverable and be specific about the finished product, not about the steps required to produce it.
What are the constraints? Under 500 words, matching the format in the templates folder, drawing only from this quarter's data. Constraints are the guardrails, not the steering wheel, and two or three are usually enough.
What should Claude do when uncertain? Flag the uncertainty and ask before proceeding, put ambiguous items in a needs-review folder, or mark them with VERIFY. This is the question most people skip, and in our experience it matters more than the other two combined.
Define the output. Define the constraints. Define the ambiguity. Walk away.
Three questions in plain language, typically under 50 words. Your context files handle everything else: the voice, the formatting, the behavioral rules, and the project specifics.
Short prompts with good context outperform long prompts with no context. We've seen this hold in session after session. The people writing ten-word task descriptions on top of well-built context files get better results than the people writing 200-word prompts into a blank workspace.
Your prompt handles one thing: what does done look like, right now, for this specific task.
Let the agent ask
Cowork has a feature that most users don't trigger, and it changes the entire dynamic of how you work with an agent.
When Claude encounters ambiguity in a task, it can stop and ask structured questions before proceeding, not a generic request for clarification but specific, targeted questions that eliminate guessing before work begins.
This feature activates when you give the agent room to use it. Micromanager prompts leave no room because every decision is already made and there's nothing left to clarify. Delegator prompts leave deliberate gaps, and the agent fills those gaps by asking the right questions at the right time.
Add one line to your working-preferences file: "When uncertain, stop and ask instead of guessing." That single instruction changes how every future session handles ambiguity across your entire workspace. The agent stops producing confident wrong answers and starts producing targeted questions that take ten seconds to answer.
We've seen practitioners report that three clarifying questions at the start of a task consistently eliminate the revision cycle at the end, replacing twenty minutes of rework with ten seconds of input.
Plan before execute
There is one more pattern worth adding to your system, and it costs thirty seconds per task.
Before the agent executes anything, ask it to show you a plan. Add this line to your Global Instructions: "Show a plan before executing and wait for approval."
A brief outline appears. You scan it, approve it, and the agent proceeds with confidence that the direction is right.
The alternative is twenty minutes of autonomous work heading in the wrong direction, followed by a full restart. We've seen both paths play out across hundreds of sessions. The thirty-second plan saves more time in a single week than any prompt technique ever published.
This isn't a Cowork-specific insight. Good managers do the same thing with human teams, asking to see the approach before the deck is built, the analysis is run, or the draft is written. The principle transfers directly to how you work with an AI agent.
The word count test
Run this diagnostic on every prompt you write.
Count the words. If your prompt is over 50, read it again and ask whether you are describing the deliverable or describing the steps. If you find step-by-step instructions embedded in the prompt, cut them and replace them with the deliverable description and a couple of constraints.
The context files handle the voice. The working-preferences file handles the behavior. Folder Instructions handle the project specifics. The prompt handles one job.
What does done look like, right now, for this specific task.
The best prompts in Cowork are short, not because brevity is a virtue, but because everything else is already in the system, and the prompt doesn't need to carry what the context files already hold.
Do This Today
- Take your next Cowork task and rewrite it using the three-question format. What does done look like? What are the constraints? What should Claude do when uncertain?
- Add "Show a plan before executing. Wait for approval." to your Global Instructions.
- Add "When uncertain, stop and ask instead of guessing" to your working-preferences file.
- Run the task. Count how many words your prompt has. If it's over 50, cut it to the deliverable description and constraints only.
- Compare the output to your last long-prompt attempt. Note what the agent found that you didn't ask for.
---
Next in the series: when Cowork reads a project folder, it reads everything, including the draft you scrapped three weeks ago and the notes from the meeting that changed direction. One file fixes this permanently, and it takes five minutes to create.
Read Part 5: The _MANIFEST.md: Five Minutes That Save Hours →