Skip to main content
You know the basics. This chapter is about getting genuinely good at it. The difference between a result that kind of works and one that nails it every time usually comes down to how much context the agent has and how clearly you write the instructions.

​
Specific beats vague

This is the single most important skill. Compare two instructions for a weekly report:
Summarize my SEO.
The agent has to guess at the site, the timeframe, the metrics, and the format.
For example.com, compare last week vs the prior week in Search Console: clicks, impressions, average position. List the 5 biggest query gains and 5 biggest drops. Flag any page that fell more than 3 positions. Format as a short bulleted report.
Same task — but now it knows exactly what to produce. The same principle applies to any job: “put together a report” gets you something very different from “give me a one-page summary of Q1 revenue by region with a comparison to last quarter.”

​
Tell it what to skip

Defining what to exclude is often more valuable than defining what to include. “Ignore branded queries.” “Skip pages with fewer than 10 clicks.” “Exclude internal test accounts.” The agent can figure out what is important — what it cannot guess is what you do not care about.

​
Explain why, not just what

When the agent understands the purpose behind a task, it handles ambiguous situations better. Instead of “flag underperforming pages,” tell it:
This report helps us catch ranking drops before they cost traffic. Flag anything that lost positions on a commercial keyword, and err on the side of flagging — I would rather review a false positive than miss something real.
Now it has a decision framework, not just a checklist.

​
Deliverables that ship

Ottic produces real outputs — reports, audits, briefs, exports — as artifacts you can hand off. The fastest way to make every one consistent is to save your preferred structure as a Skill, so the format is reused instead of re-invented each time.

​
Progressive trust

Start a new Routine in read-only. Run it for a day and review what it would have done in the run history. Promote to approval-required for a week, approving or rejecting each action. Once it is reliable, go autonomous. You can always dial back.

​
Power moves

  • The sandbox runs code and a real browser — Python for data crunching and chart generation, and Playwright for live web research, scraping, and screenshots. If a task involves numbers or the open web, it can handle it.
  • Save winning prompts as Skills — reusable playbooks you launch in one click.
  • Schedule Skills as Routines for recurring work.
  • Every run is logged — your audit trail and debugging tool.

​
What’s next

You can write better instructions and recognize the patterns. The last chapter covers fine-tuning the agent and keeping everything running the way you want.