Three Files That Replace a Chief of Staff
The full series
- The Prompt Is the Least Important Part
- Three Files That Replace a Chief of Staff (you are here)
- You Don't Have a Productivity Problem. You Have a Setup Problem.
- Stop Telling AI What to Do. Tell It What Done Looks Like.
- 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 their context files once and never open them again. The files work, but they deliver only 40% of what they could.
You created the three files, pointed Cowork at your folder, and the output was immediately better than anything a blank-slate chatbot ever produced. Good enough.
Except "good enough" has a ceiling, and you hit it in the first week. The output sounds roughly right, but not exactly right, a tone that sits close but lands off for certain audiences, formatting that works but doesn't match what a specific client expects. We've seen this pattern across hundreds of practitioners who finished the initial setup and then never touched the files again.
The AI does the work, and you still rewrite thirty percent of every deliverable before sending.
The files aren't wrong. Just vague where they need to be specific, general where they need to be concrete, and silent on the things that matter most.
The diagnostic skill
When output misses the mark, most people rewrite the prompt, adding more instructions and tighter language to compensate. This fixes one session and teaches the system nothing. The prompt disappears, the gap stays, and the same problem resurfaces two days later.
The better move takes ten seconds. Ask yourself one question: Is this a prompt problem, or a context problem?
Prompt problems are task-specific and limited to the wrong scope, deliverables, or audience for this particular request. Fix the prompt and move on.
Context problems occur across multiple sessions while wearing different clothes. Wrong tone on Tuesday, wrong format on Thursday, wrong assumptions the following week. The same class of mistake in different tasks, all pointing back to a gap in one of the three files.
When you spot a context problem, ask which file. Wrong tone, generic phrasing, or words that don't belong in a client deliverable point back to the brand-voice file.
Wrong priorities, wrong assumptions about the role, or output that misses what matters this quarter point to the about-me file.
Wrong behavior, wrong format, or output that ignores the rules points to the working-preferences file.
One line, added to the right file, creates a permanent correction that every future session inherits automatically. We've tracked this pattern, and it holds across marketing, law, finance, and operations.
Weak files vs. strong files
The gap between a mediocre context file and a great one is always the same thing.
Specificity.
A weak about-me contains three words. Works in marketing. A strong about-me specifies the role, the audience, and the current initiatives, something like running content strategy for a B2B SaaS company selling to mid-market CFOs while launching a partner channel and reducing churn. The weak version gives the AI a category, and the strong version gives it a mission with enough detail to produce work you'd actually send.
A weak brand-voice says two words. Professional, friendly. A strong brand-voice specifies short sentences, direct tone, no hedge words, explicit bans on leverage and synergy and circle back, with notes on how tone shifts between client emails, internal Slack, and social media. The weak version is a mood, and the strong version is a manufacturing spec that the AI can follow sentence by sentence.
A weak working preference contains a single vague instruction, such as 'ask before doing anything major'. A strong working preference lists specific rules. Show a plan before executing, ask clarifying questions when the task has more than one interpretation. Never delete files, default to .docx, flag anything below 80% confidence. One guides. The other governs.
The file knows which phrases to avoid.
The file knows how your tone shifts between client emails and internal messages.
The file knows what format your deliverable needs before you mention it.
The file knows when to stop and ask, rather than guessing.
That knowledge came from specific lines that most people don't write.
What you don't want matters more
Positive instructions tell you what to aim for: "Professional tone, clear structure, concise paragraphs." These help.
In our research, negative instructions narrow the range from broad to precise. These help more, and they're almost always missing.
The phrases-you-hate list in the brand-voice file is one of the strongest additions that most people skip. Every banned phrase eliminates an entire category of generic output. Five specific bans consistently do more work than fifty words of positive description. We've watched practitioners add a short ban list and see immediate improvement in the next session.
But bans on individual words aren't the only negative instruction that compounds. We've found anti-examples work the same way. Paste a piece of writing that represents the wrong tone and annotate what makes it wrong, and the AI learns faster from one concrete anti-example than from any amount of aspiration. The best context files are built from "never this," as much as "always that."
The five-minute weekly audit
Context files aren't a setup task. They are a living system.
Every week, spend five minutes on one file. Read the last few Cowork outputs. Was the tone off? Did a priority shift occur since your last update? Did you correct formatting that a single line in your preferences file would prevent?
Find the gap and add one line.
In our experience, the file improves while the prompts stay the same.
Over weeks and months, the gap between what the AI produces and what you would have written closes. Not because the prompts improve, but because the context sharpens with each correction, building on every previous one. We've seen practitioners reach first-draft quality within three or four weeks of weekly refinement, and that timeline surprises anyone who assumes these files are a one-time chore.
Specificity compounds. Vagueness doesn't.
Do This Today
- Open your brand-voice file and add three phrases you never want to see in your output. Not "don't be too formal" but specific bans on specific words.
- Add a "tone by context" section describing how your voice changes in client emails, internal Slack, and social media.
- Open your about-me file and replace vague descriptions with specifics, turning categories into missions with current priorities and real numbers.
- Run a real task. When the output misses, ask which file would fix it. Add one line to that file. Run the task again.
- Set a weekly calendar reminder for a five-minute context file audit. One file per week. One correction. Permanent improvement.
---
Next in the series: Global Instructions are just the first layer. Cowork has a second instruction layer that almost nobody uses, and it changes how things behave when you open a session inside a project folder. We cover Folder Instructions, tool connectors, and the 30-minute setup that eliminates the cold-start tax.
Read Part 3: You Don't Have a Productivity Problem. You Have a Setup Problem. →