> ## Documentation Index
> Fetch the complete documentation index at: https://help.casaconect.ai/llms.txt
> Use this file to discover all available pages before exploring further.
# Working well with the agent
> Practical habits for getting accurate, useful work from the agent: briefing, reviewing, and knowing when to use a plan, a skill or a stronger model.
The agent is most useful when you treat it as a capable colleague who is new to the file: able, quick and well read, but dependent on you for the context and the standard. These habits make the biggest difference.
## Brief it like a colleague
A good brief covers four things.
Say what you are trying to achieve, not only the task. *"Draft a reply"* is a task. *"The client is anxious about the delay. Reassure them, and get the missing certificate by Friday"* is a goal.
Reference the documents, or make sure you are in the right project or matter with **Include what you're looking at** switched on. Do not describe a document that you could point to.
Language, length, audience, format. If there is a model to follow, reference it: *"in the style of the letter we sent to the Anderssons"*.
Say what it should **not** do, and when to stop and ask. *"Don't contact anyone. If the NIE is missing, ask me."*
```text Weak theme={null}
Check the contract.
```
```text Strong theme={null}
Review the seller's draft arras contract (referenced) against our
buyer-side checklist skill. I act for the buyers. List every point
where the draft departs from our standard position, quote the
clause, and suggest replacement wording as tracked changes. Don't
touch the price or dates. Flag those for me instead.
```
## Match the tool to the job
| Situation | Reach for |
| :------------------------------------------- | :------------------------------------------------------------------------------------------------------------ |
| A large or unfamiliar task | **Plan mode**. Approve the approach before anything changes. See [Conversations](/agent/conversations#modes). |
| Work you do every week | A [skill](/agent/skills) |
| A preference you keep repeating | [Guidance](/agent/guidance) |
| Filling a template from documents | [Populate](/workflows/populate), not a conversation |
| Demanding drafting or reasoning | A [stronger model](/models/choosing-a-model) |
| A piece of work with a deadline and an owner | A [matter](/matters/overview) |
## Review like a supervisor
The agent's work is a draft until you have checked it. The product is built to make checking quick.
* **Tracked changes.** Accept and reject individually. Read the deletions as carefully as the insertions.
* **Sources.** When the agent cites a document, open it. When [Populate](/workflows/populate) shows *Contradicted* or *Insufficient*, treat the value as unverified.
* **What was touched.** After each turn, open the items the agent lists as edited, moved or filed.
* **Recipients.** On every drafted email, check To and Cc before the text.
The agent can be wrong in a fluent, confident voice. It can misread a scan, merge two people with similar names, or state a legal position that was superseded. It does not replace your professional judgment, and nothing it produces should reach a client or a registry unreviewed.
## Correct it, then teach it
When the agent gets something wrong, fix the instance, then ask whether it will come up again.
* **A one-off:** just say so. *"That's the wrong Pérez. The buyer is María Pérez Gil."*
* **A preference:** *"Remember that I write dates as 14 March 2026"*, or add it to your guidance.
* **A procedure:** *"Save that as a skill."*
## Keep conversations focused
* **One subject per conversation.** The agent re-reads the conversation on every turn. A long, mixed conversation is slower, costs more and blurs the instructions.
* **Start fresh when the subject changes.** Your memory and guidance carry over. The clutter does not.
* **Use matters for long-running work.** A matter keeps the description, documents, conversations and history together in a way a single chat cannot.
## Work safely
* Keep the default [permission level](/agent/permission-levels), **Approve for me**, unless you have a reason to change it.
* For anything irreversible, such as a deletion, a submission or a message to the other side, have the agent prepare and do the final step yourself.
* Do not paste passwords or card numbers into a conversation. For websites that need a login, use a [browser session](/agent/browser-sessions) and sign in yourself.
* If something the agent read appears to be giving it orders (*"ignore your previous instructions…"*), that is a red flag about the **document**, not a fault in the agent. Tell your admin.
## When it is not working
It probably lacks your material. Reference the documents and say who the reader is.
Check where the document lives. The agent sees what you can see, in the project you are working in. A file in your private vault is not visible from a project conversation unless you reference it. Newly uploaded files take a short while to become [searchable](/documents/search).
You are probably on *Ask for approval*. Switch to **Approve for me**, or use **Always allow in this thread** for the action in question.
It reached its step or time limit. Reply *"continue"*. For large jobs, ask for a plan and approve it in stages. See [Limits](/agent/cloud-execution#limits).
Try a [stronger model](/models/choosing-a-model) for drafting, and give it an example of writing you like.