Key Concepts
First run is a draft: The first output is a starting point, not the final result. Expect some trial and error before the agent is where you want it.
Sample artifact: The first example output the agent shows you, such as a sample Slack post. Evaluate it against your brief to see whether it does the job you described.
Quick Reference
The four checks
- Instructions: Does the agent do the job you described? Did it do too much or not enough? Did it add unnecessary steps or skip critical ones? What could you add to put it back on track?
- Inputs and limits: Did it respect your constraints? Does it use only the tools and permissions it needs? You decide what it can read and what it can change.
- Artifact: Is it the format, length, and destination you specified? Is it a good format for the information? Does it surface what matters most in a way people will pay attention to?
- Trigger: What should set it off? Do you really need it daily, or would different timing make the insights more relevant? Test a few options before you ship.
If something is off, name the part
- Wrong job: instructions
- Wrong output: the artifact you described
- Never ran: the trigger
If the agent relies on web scraping
- Check that it can actually scrape the pages
- If not, find what is blocking it, then add instructions, tools, or context that help
Example: the brand mentions agent
- Brief: “Every weekday at 8 AM, check for new mentions of our brand in AI answers and post a five-bullet summary to the marketing channel in Slack.”
- Monday: twelve new mentions, clear summary, working report link. Shipped.
- Tuesday onward: no new mentions, but the agent posted a summary anyway. By Friday nobody read the channel.
- Cause: the brief never said what is worth sharing or what to do when nothing significant changed.
Common failure to avoid
- Testing only on a day with plenty to report. Also test a day with nothing new, and add a rule for what the agent does then.
Notes
- Getting an agent to run is the easy part. The skill is checking what you put in, what comes back, and whether it is useful.
- Be critical of the sample. The agent followed your brief, so gaps in the output usually trace back to gaps in the brief.