How to talk to it

Practical patterns that get good output, and the ones that waste a turn.

Be specific about the artifact

Vague: "help me with onboarding emails."

Specific: "draft a 4-email welcome sequence for new Ahrefs Starter signups, focused on getting them to set up their first project. One CTA per email, each under 120 words."

The agent treats the second one as a brief and produces drafts. The first one triggers a clarifying questionnaire, which is fine, but costs you a round trip.

Tell it the format you want back

• "Reply in a table with columns: angle, audience, channel, owner."
• "Give me the JSON only."
• "Markdown with headings, no preamble."

Default outputs are conversational. Format hints save edits.

Push work into artifacts, not the chat

If you want something you will reuse, ask for a file or an app, not a chat reply. The chat scrolls away. A file in your workspace, a page in the Console, or a job on a schedule does not.

Examples:

• "Save this as ~/workspace/notes/q1-launches.md."
• "Build a Console page that shows our top 20 organic pages by traffic, refreshed daily."
• "Schedule a job that emails me every Friday with new keywords my competitors started ranking for."

Use "remember this"

The agent reads ~/workspace/.memory.md at the start of every chat. When you say "remember this," it appends. Good things to remember:

• Your role and team.
• Recurring vocabulary (product names, internal acronyms).
• Stylistic rules ("never use em dashes," "always ship in two variants").
• Long-term goals so it can connect dots later.

Bad things to remember: anything sensitive, anything that changes weekly.

Approve once, not per turn

Some actions surface an approval card: installing a Python package, scheduling a job, allow listing a new domain, connecting a third-party API. Owner or admin clicks once; the agent retries automatically.

If you keep getting stalled on approvals, the fix is usually to batch the request: "set up the full Slack integration, request all the scopes you need in one go." One card, one click.

When to ask for a plan vs jump to execution

Ask for a plan when:

• The work touches multiple systems.
• You are not sure what is feasible.
• You want options to choose between.

Jump to execution when:

• The output is small and clearly scoped.
• You have a template in mind ("same as last time, but for X").

The agent has a bias toward asking before big work. You can override with "skip the questionnaire, just build it."

Things that do not work well

• Pasting a screenshot and saying "you know what to do."
• "Make it better." (Better than what, in which dimension?)
• "Do everything in one shot." (For multi-system work, the agent will plan first whether you ask or not.)
• Conversations that run for hundreds of turns. Start a fresh chat for new tasks; memory carries the important context across.

Under the hood. Each chat has a finite context window. The agent prefers writing intermediate results to disk rather than echoing them in chat. If you ask "show me all 5,000 rows," it will instead write them to a file and show you a sample plus a count. This is intentional and keeps long sessions usable.