Pattern 1: outcome-based, not step-based
Every practitioner interviewed for this course named the same mistake first. Not vague prompts. Not missing context. Not bad folder structure. The number-one mistake was over-specifying steps — telling Cowork how to do the work instead of what you want done.
When you dictate the procedure, you hobble the agent. You've committed it to a path you chose, with tools you named, in a sequence you invented. Cowork can't pick a better approach — you've already decided the approach for it.
What over-specifying looks like in practice
Here's the same task written two ways.
Outcome — bad
"Open the folder, find the
.pptxfile, extract the colours with python-pptx, then create a new presentation using those colours, making sure to use the same fonts and layout proportions as the original..."
This prompt specifies the tool (python-pptx), the sequence (extract then apply), and the file type to look for. You've made three decisions that Cowork should make for itself — and at least one of them might be wrong. If python-pptx doesn't handle the specific .pptx variant in your folder, you've locked Cowork into a failing approach.
Outcome — good
"Build a presentation that matches the design of
abp-presentation-template-light.pptxin the same folder."
This prompt describes the outcome (a presentation that matches the design), identifies the reference file, and leaves everything else open. Cowork can now choose whether to use python-pptx, LibreOffice, direct file manipulation, or some combination. It picks the tool that actually works. You get a matching deck regardless of which tool got you there.
What to include — the three things
The outcome. One measurable sentence that says what success looks like. "Build a presentation that matches the design of X." "Produce a summary report of all documents in raw/." "Clean and normalize this CSV into a standard format."
The constraints. What not to do, where not to write, what to leave untouched. "Do not modify the files in raw/." "Write outputs only to out/." "Do not delete anything." Constraints are not limitations on Cowork — they're the guardrails that make delegation safe.
The success criteria. How you'll know it worked. "The output file opens in PowerPoint without errors." "The CSV validates against the schema in schemas/." "Every source document has a corresponding section in the report." When Cowork knows what done looks like, it can self-check before handing off.
What to leave out — the three things
The how. Don't specify the sequence of operations. If you're tempted to write "first do X, then do Y," stop and ask whether you actually know the right sequence — or whether Cowork might know better.
The tool choice. Don't name a specific library or application unless you have a genuine reason (a downstream dependency, a format requirement). When you name python-pptx, you've ruled out LibreOffice. When you name LibreOffice, you've ruled out python-pptx. Let Cowork pick the tool that fits the job.
The sequence of steps. This is a restatement of the first point, but it's worth repeating because it's easy to slip into. Cowork is an agent, not a function executor. It plans its own sequence. When you pre-plan the sequence for it, you're doing half its job badly.
The underlying principle: Cowork has access to multiple tools for almost every task. Outcome-based prompts let it use all of them. Step-based prompts lock it into the one you happened to name.
Key Takeaways
- 1The number-one Cowork mistake is over-specifying steps — telling it how to work instead of what to produce. This hobbles the agent by forcing it onto your chosen path.
- 2Outcome-based prompts include three things: the outcome (what you want), the constraints (what not to do), and the success criteria (how you'll know it worked).
- 3Leave out the how, the tool choice, and the step sequence. Cowork has access to multiple tools for nearly every task — let it pick the one that actually works.
- 4When you name a specific tool (python-pptx, LibreOffice), you commit Cowork to it even if a better option exists. Name outcomes, not tools.
- 5The test: if your prompt reads like a recipe, rewrite it as a destination. One measurable sentence describing what done looks like is almost always enough.