By KattebTeam, AI Agents Platform
Most disappointing AI agent runs fail before the first click. The agent was not bad; the instruction was. "Find me some leads" or "keep an eye on competitors" reads fine to a person, but to an agent it has no finish line, no definition of good, and no idea what it is not allowed to do.
After watching a lot of agent runs, we think a useful goal always answers five questions. Below are the five, followed by six copy-paste templates.
The five parts of a finishable goal
- Outcome - what exists when the job is done? A sheet with 20 rows, three replies posted, one report. Countable beats vague.
- Source - where should the agent look? Name the sites, the inbox, the dashboard, the file.
- Quality bar - what makes one result acceptable and another not? "Companies with 10-50 employees in Germany that sell B2B software" is a filter; "good companies" is not.
- Limits - what must it never do? No purchases, no more than N messages, no deleting, ask before publishing.
- Report - how should it prove what it did? Links, row counts, screenshots, and a plain list of what failed.
If you can't fill in all five, the agent can't either.
Six templates
1. Lead research
Find 20 companies in [niche] in [country] with [size] employees. Use [directory / search engine]. For each: name, website, one decision-maker's public profile, and one sentence on why they fit. Skip companies already in [sheet]. Do not contact anyone. Report the rows as a table.
2. Competitor price watch
Open the product pages at [list of URLs] every morning. Record price, stock status and any promotion banner. Flag any change greater than 5% compared to yesterday's row. Report only the changes.
3. Review replies
Read new reviews on [platform] since the last run. Draft a reply to each one in the reviewer's language, thanking them and addressing the specific point they raised. Post replies to 4 and 5 star reviews; for 1-3 stars, save the draft and ask me before posting.
4. Invoice follow-up
Check [billing tool] for invoices more than 14 days overdue. Send one polite reminder per client using [template], at most 10 emails per run. Never change an invoice amount. Report who was reminded and when they were last contacted.
5. Content publishing
Publish the article in [doc] to [platform] with the title, tags and cover image given. After publishing, open the public URL while signed out and confirm it loads and contains the link to [site]. Report the URL and the check result.
6. Weekly SEO check
For the keywords in [sheet], record our position and the top three competitors on [search engine]. Compare with last week. List keywords that dropped more than 3 positions and the page that replaced us.
Why the "report" part matters most
An agent that says "Done!" is not useful. An agent that says "Posted 3 of 4; the fourth platform asked for a phone number, so I stopped" is. Ask for evidence and for failures explicitly, and you will be able to trust the agent with longer jobs over time.
Try it
Take one task from your week, fill in the five parts, and run it. If the first report surprises you, the goal was missing a part; fix that part rather than rewriting everything.
We build browser-based AI agents for small teams at https://katteb.com. Questions about writing agent goals are welcome in the comments.
Top comments (0)