Checklist

Before you let an AI agent run an outcome

A short human prep checklist for delegating real-world goals safely. Give the agent the outcome, boundaries, approvals, measurement access, examples, and stop conditions before asking it to move a number.

1. Name the outcome

  • Write the one number you want improved: traffic, signups, revenue, activation, reply rate, support load, or something else measurable.
  • Include where that number lives: GA4, PostHog, Stripe, HubSpot, GitHub, a database, or a report.
  • Give the current baseline if you have it. If not, tell the agent where to collect it first.

2. Set the hard boundaries

  • List what the agent must never do: spam, fake traffic, paid spend, credential changes, destructive edits, off-brand claims, or legal/compliance-sensitive actions.
  • Define quality standards in plain language so the agent knows what counts as useful, not just what is technically allowed.
  • Write the brand, tone, and audience constraints that should survive every run.

3. Decide approval rules

  • Say what the agent can do alone: research, draft, edit local files, run scripts, analyze metrics, or prepare assets.
  • Say what needs human approval: publishing, deploying, sending emails, posting on social, opening public PRs, spending money, or changing accounts.
  • Tell the agent what evidence to show you at the approval boundary: diff, preview, exact post copy, target URL, or recipient list.

4. Provide measurement access

  • Point to the scripts, dashboards, APIs, exports, or saved reports the agent should use to read results.
  • Keep credentials out of prompts and repos; provide access through approved local files, environment variables, or existing connectors.
  • Name the secondary metrics that help explain the result, such as source, page, campaign, conversion, or referral domain.

5. Give examples of good and bad actions

  • Good examples teach taste: useful resource page, specific community answer, targeted pitch, helpful docs fix, relevant integration page.
  • Bad examples prevent damage: generic link drops, thin SEO pages, fake engagement, irrelevant outreach, vague social posts, or misleading claims.
  • If there are past attempts, show which ones worked, which failed, and why.

6. Define stop conditions

  • Tell the agent when to stop and ask: missing access, unclear ownership, risky external action, weak evidence, or a change outside the agreed goal.
  • Set budget, time, and blast-radius limits so one run stays small enough to review.
  • Require a short record of what changed, where it shipped, and what result should be checked later.

The prep sheet

Outcome:
Metric source:
Baseline:
Never do:
Needs approval:
Measurement access:
Good examples:
Bad examples:
Stop and ask when:

Keep the loop from resetting

Superdense helps agents remember each run, connect shipped work to measured reward, and use what happened last time when choosing the next action.

npm i -g @nimrobo/superdense
Read outcome docs →View GitHub repo →