> ## Documentation Index
> Fetch the complete documentation index at: https://docs.cogniagent.ai/llms.txt
> Use this file to discover all available pages before exploring further.
# Giving instructions and personality
> Two short pieces of writing decide how well your AI coworker does its job. Here's what belongs in each, and what to leave out.
Everything your coworker does traces back to two things you write: what its job is, and how it should carry itself.
## What it does
Open your coworker's **Files** tab and you'll find them both, each with a plain-language label:
| | What it holds |
| ---------------- | --------------------------------------------------------------------------------------------------------------- |
| **Instructions** | The job. What this coworker is responsible for, how to approach it, what good looks like, when to stop and ask. |
| **Persona** | The manner. How it writes, how formal it is, how it talks to a customer versus a colleague. |
Keep them separate. Instructions are about *the work*; persona is about *the voice*. Mixing them makes both harder to change later.
## Writing good instructions
The best instructions read like what you'd tell a capable new hire on their first day. Not a specification — a briefing.
**Say what the job is, not just the topic.**
> Weak: "You handle research."
>
> Better: "You answer questions that need a decision behind them. Every claim carries a source. If the evidence is thin, say so rather than filling the gap."
**Say what good looks like.** Your standard is not obvious. If you want three options and a recommendation rather than an essay, write that down.
**Say when to stop and ask.** This is the most valuable line most people never write. "Ask me before contacting anyone outside the company" prevents more problems than any setting.
**Say what to leave alone.** Boundaries are as useful as instructions. "Don't touch pricing questions — send those to me."
## Writing a good persona
Shorter than you think. Three or four sentences is usually plenty.
> Direct and unfussy. Leads with the answer, then the reasoning. Never pads a reply to seem thorough. With customers, warm but not chatty.
The test: could a colleague read this and imitate the voice? If not, it's too vague.
**Name what you don't want.** "Never opens an email with 'I hope this finds you well'" is a more effective line than any amount of positive description.
## When you'd change them
* The coworker keeps making the same wrong assumption
* You find yourself repeating a correction in every task
* Its output is right but the tone is off for the audience
* The job has genuinely changed
* You want it to ask before something it currently does on its own
## Good to know
**Changes apply to the next task.** A task already running keeps the instructions it started with. Edit, then start a new task to see the difference.
**Correct in the chat first, edit the instructions second.** Telling your coworker "too long — half that next time" in the conversation is faster, and it [remembers](/cowork/control/memory). Move something into the instructions once you've said it three times.
Instructions are for *how this coworker works in general*. A procedure with steps that must happen in order belongs in a [skill](/cowork/skills/overview) instead.
## Instructions, skills, or memory?
All three shape behaviour, and it's worth knowing which to reach for.
| | Use it for | Example |
| ------------------------------------- | ------------------------------------------ | ----------------------------------------- |
| **Instructions** | Standing facts about the job | "You never contact customers directly." |
| **[Skills](/cowork/skills/overview)** | A procedure with a right way of being done | How your team runs a competitive analysis |
| **[Memory](/cowork/control/memory)** | Things it worked out along the way | "This person prefers bullet points." |
## Common questions
A few paragraphs. Long instructions dilute — if everything is emphasised, nothing is. Put procedures in skills and keep instructions to the standing rules.
No. Instructions and memory are separate. Rewriting the job description doesn't erase what it picked up about how you like things.
Not directly, but you can save a tuned coworker as a template and hire from it. See [Retiring, re-hiring and templates](/cowork/coworkers/templates).
A coworker built from scratch with no instructions is a general assistant — capable, but with no view on what matters to you. Ten minutes of writing is the highest-return thing you can do.
## Next steps
For procedures, not standing rules.
What it picks up on its own.
Well-written instructions to borrow from.
Boundaries you enforce rather than ask for.