The standing set
These are scheduled chats — a saved prompt, a cadence, and a destination. Set them up once from Settings → Scheduled chats, or ask Emma in Agent Mode to create one.
Start with two. The Monday review and the weekday lead run are the ones that carry the system; the rest are additions once those are habitual.
The Monday review
The one hour that makes the other four days worth running. Three questions, in this order.1
What produced qualified pipeline?
Not leads — qualified ones. Pull the lead-source breakdown with average ICP score per source, per measurement.You’re looking for one thing: a source where volume rose and average score fell. That’s the pattern that quietly wastes budget, and it’s invisible if you only watch cost per lead.
2
What did I hear last week?
Go through the conversations you actually had. Objections that repeated, language buyers used that you didn’t, a segment that keeps saying no.Write it into Brand Hub → Market Strategy. Emma re-reads it on every run, so this one edit propagates to lead scoring, ad targeting, content angles, and outreach copy at once. Skipping it is the single most common reason a system stops compounding.
3
What's one thing to change?
One. A tightened ICP, a killed campaign, a new content angle, a raised budget on something working.Changing four things a week means you never learn which one mattered.
The monthly maintenance pass
Thirty minutes, month end. Four checks that stop slow decay.- Review your schedules. A schedule pointed at a source that’s gone quiet keeps producing empty reports indefinitely, and nobody notices because empty reports look like normal reports. Kill anything you’ve stopped reading.
- Check that Market Strategy still matches reality. If your ICP has shifted and the file hasn’t, every downstream output is drifting with it.
- Convert what worked. Any conversation you’ve now had three times should be a work plan or a skill. Do it while the thread is fresh — reconstructing it later from memory produces a worse version.
- Prune the Knowledge Base. Outdated case studies and superseded positioning docs actively degrade output, because Emma treats them as current.
The autonomy ladder
Automation earns trust; it doesn’t start with it. The failure mode isn’t Emma doing something wrong once — it’s scheduling a step you never verified, then not reading the output that would have told you. So move one rung at a time, and only on evidence.Level 1 — Interactive, every item reviewed
You’re in the conversation. Emma drafts, you read every item, you approve each one. Nothing is scheduled. The point of this level isn’t caution for its own sake — it’s calibration. You cannot judge whether Emma’s output is good until you’ve seen enough of it to know what good looks like for your ICP and your voice. Stay here until: you’ve reviewed roughly 50 items of a given kind (leads, drafts, messages) and you can predict what she’ll produce before she produces it.Level 2 — Scheduled drafting, interactive approval
Emma runs on a cadence and produces a batch. The schedule’s final step is “draft and send me the list” — never “send”. You approve and dispatch in a separate, interactive session. This is where most founders should live indefinitely. It removes the work of remembering and preparing, and keeps a human on every outbound action. Before promoting, run both evals:- Human eval — your edit rate. Across the last three batches, what fraction did you approve unchanged? Below about 80% and the output isn’t stable enough to sample rather than read. Track it; don’t estimate it from memory.
- Agent eval — make Emma grade herself. Add a final step to the plan: “Before presenting, score each item 1–5 against these criteria: ICP fit, personalization specificity, factual grounding. Flag anything below 4 and say why.” Then check her scores against your own judgment on a sample. If her self-scores don’t correlate with yours, she isn’t ready to be trusted as a filter, and the fix is sharper criteria rather than more autonomy.
Level 3 — Scheduled drafting, sampled review
Same as Level 2, except you spot-check rather than read everything — review Emma’s flagged items plus a random sample of the rest. Approval is still interactive. What changed is how much you read, not whether a human dispatches. Requires: a sustained edit rate above 80%, agent self-scores that track your judgment, and at least a few hundred reviewed items.Level 4 — Narrow autonomy, low blast radius only
A small set of actions run without you: internal research, enrichment, scoring, report generation, drafting into a queue. Everything at this level shares three properties — reversible, private, and cheap. If an action fails any one of those, it doesn’t belong here regardless of how well Emma has performed.What never reaches Level 4
These stay interactive permanently. Not because Emma performs badly on them, but because the cost of the rare failure is unbounded while the saving is a few minutes.
When to drop back down
Demote immediately, without treating it as a setback:- Your edit rate falls below 80% for two consecutive batches.
- Anything went out that you’d have stopped.
- Your ICP, positioning, or brand voice changed. This is the one people miss. Emma re-reads Market Strategy on every run, so a positioning change alters her output while your trust calibration still reflects the old version. Return to Level 2 and re-baseline.
- A source went quiet and the schedule kept producing empty output nobody noticed.
Where each motion should sit
Scaling the rhythm
As the team grows, the cadence stays and the ownership splits:- Solo founder: you run all of it. Keep the standing set to two or three.
- Founder + first marketer: marketer owns the Monday review and content backlog; founder keeps outreach approval and the conversation-to-strategy loop. That second one shouldn’t be delegated — it depends on being in the sales conversations.
- Small team: move to a shared workspace so schedules, brand context, and lead lists are visible to everyone rather than trapped in one person’s account. See Organization sharing.
Next steps
- The pipeline system — what this rhythm is operating.
- Measurement — the review prompt and the metrics behind it.
- Work plans — turning a repeated conversation into a reusable sequence.
- Agent Mode — creating and managing scheduled chats.