Working with an AI assistant step by step

Last updated October 21, 2026Auf Deutsch lesen

Once the connector is set up, the question is what to actually say. This is the walkthrough: one brief, start to finish, with the prompts that work and the checks worth doing.

First, give your assistant the context file

An assistant that has never seen uplads will feel its way through the tools. It works, but it wastes turns and sometimes leaves targeting in a note instead of on the ad set.

We publish a context file that fixes that:

https://uplads.com/uplads-mcp.md

Save it into your project:

Claude Code, Cursor, Codex: save it as .claude/skills/uplads/SKILL.md, or paste it into your CLAUDE.md / AGENTS.md.
Claude or ChatGPT in the browser: attach it to the conversation, or paste it once at the start of a project.

It describes the staging model, the right order of operations, the whole tool inventory, and the daily limits. With it in context, a brief usually runs end to end without a wrong turn.

The one rule that shapes everything

Your assistant prepares. You launch.

Campaigns and ad sets it creates are paused. Staging a draft costs nothing, on Meta, TikTok, Google Ads and Snapchat alike. Two tools publish ads, launching a draft and retrying a launch that did not finish, and both refuse unless you have switched on autonomous launching in Configuration → MCP. Everything else hands you a review link instead.

The same holds for campaigns that already run. Pausing and lowering a budget need no permission. Switching something on, raising a budget, loosening a bid or letting a rule do that on its own needs that switch and your yes in the chat. Everything assistants put live or raise in a day also has to fit under the daily ceiling you set next to the switch, all of it together.

The switch and the ceiling are for assistants connected from outside over MCP, such as Claude or ChatGPT. In the assistant inside uplads, pressing GO on its confirmation card is your yes for everything it does there.

That means you can let it work without watching every step.

Step 1: hand over the brief

Write it the way you would write it for a colleague. Concrete beats polite.

Launch our summer sale on Meta in the Bio Stefan account. The videos are in this Dropbox folder: . 50 € a day, Germany and Austria, running until 31 August, linking to /summer-sale. Conversion campaign on the purchase pixel.

What makes a brief run cleanly:

Name the ad account if you have more than one, otherwise the first question back will be which one.
Link the creatives rather than describing them. A Dropbox or Google Drive folder link imports the whole folder.
Say the money and the dates. Budget, start, end.
Say the goal, not just the format. "Conversions on the purchase pixel" tells it which objective and optimisation to pick.
Keep your Brand Context current. Your assistant reads it before it writes a word, so the copy uses your products, your audiences and your words.

Step 2: let it import the creatives first

The first thing a well-behaved assistant does is pull the assets in, before building anything. If a folder link is private or a file is the wrong format, you find out in the first minute rather than after three ad sets exist.

You can watch this happen in Media library while it works.

Step 3: check the targeting, not the prose

This is the step worth your attention. Every interest, behaviour, job title, city and pixel has to be resolved to a real platform id before it lands on the ad set. A good assistant does that with the search tools and tells you when a term has no match.

A useful nudge if you are not sure:

Which geo and interest ids did you actually put on the ad set? List them.

If the answer is a description rather than ids, the targeting is not really there yet.

Step 4: read the review link

When the draft is staged you get a link into the uplads Ad Launcher, opened on your draft with a notice above the cards saying an assistant prepared it. The same draft sits in the Launch Drafts tab under the launcher, so you can come back to it later.

Check the four things that cost money if they are wrong:

Budget and schedule on the ad set.
Geo and audience.
Which creatives are paired with which text.
The destination URL, including UTM parameters.

Then launch it yourself. Nothing before this step has spent anything.

Want something different? Say it in the chat, for example "use five text variants instead of two". Your assistant changes the draft under the same review link. Open the link again before you launch: a tab that was already open keeps the old version, and uplads refuses to launch from it.

Step 5: come back for the numbers

Later, the same connection reads performance:

How did the summer sale ads do over the last 7 days? Compare them to the evergreen campaign.

Ask for the numbers before asking for advice. An assistant that has pulled real spend, CPA and ROAS gives a different quality of answer than one reasoning from the brief alone.

Then you can act on it in the same conversation:

Pause everything under 1.5 ROAS, and raise the budget on the top ad set to 80 € a day.

The pauses go through right away. The raise needs the Autonomous launching switch in Configuration → MCP and your yes in the chat. Without the switch it is refused, and you raise the budget yourself in uplads.

Things it can do that are easy to miss

Turn an organic post into an ad, keeping its existing likes and comments.
Duplicate an ad across ad sets instead of rebuilding it, or an ad set or a whole campaign, created paused.
Change an ad set's targeting or placements without rebuilding it.
Swap the image or video of a running ad, or edit carousel cards. The ad gets a new post, so likes and comments start over and Meta reviews it again.
Stage launches on TikTok, Google Ads and Snapchat the same way, including a Spark Ad on your own TikTok post or a Performance Max asset group built from your best Meta ads. See What your AI assistant can do in uplads.
Manage TikTok, Google Ads, Snapchat and LinkedIn: pause, budgets, bids, names and copies, plus keywords and ad texts on Google. See What your AI assistant can change in running campaigns.
Write and rewrite copy in bulk, then show you a preview of the rendered ad.
Set up automated rules, for example pausing an ad set when CPA goes over a threshold, and replay a Meta rule over the last days before it goes live.
Run an Optimize audit and apply the findings you pick.
Tell you which creatives win across every ad that uses them.
Schedule a budget change for a date, or every weekend, rather than remembering to do it.
Follow a launch through: read it ad by ad, retry what failed, and check that every Meta ad came out the way it was sent.
Keep the media library tidy with tags, boards, placement crops and video covers.

When it stops

"Daily MCP limit reached." There are daily budgets on writes: 40 structure and money changes, 300 edits, 40 drafts, 100 imports, and 20 Optimize audits. They exist to stop a runaway agent, and a full brief uses a fraction. Your assistant should tell you and stop, not retry. Carry on in the uplads UI or wait a few hours.

It reads but cannot create. A trial that ended without a plan turns the workspace read-only, exactly as the interface does.

It cannot see new tools. Clients cache the tool list. Disconnect and reconnect the connector.