What you'll be able to do after this guide: change what your agents focus on week to week by editing a spreadsheet cell or a document, without touching the scheduled task itself.
Editing an agent's prompt is risky: it's easy to break something, and the prompt and its copies drift apart. Most changes you'll want ("focus on this niche this month", "stop sourcing that category") aren't changes to how the agent works. They're changes to what it works on. Put those choices in data the agent reads every run.
Fernhill's Segment B agent sources solo experts across dozens of niches, listed in a category spreadsheet. The spreadsheet has an Active Focus column:
| Category | Priority | Active Focus |
|---|---|---|
| Bookkeeping services | High | Elevated |
| Interior design | High | Elevated |
| Garden design | Medium | |
| Wedding photography | Medium | Paused |
| Personal styling | Low |
The agent's instructions say:
Re-read the spreadsheet from the file at the start of every run. People edit it.
Pick 2 or 3 categories today.
At least one must be an Elevated category, as long as that category isn't already saturated in the CRM.
Never pick a Paused category.
Otherwise, pick the categories with the fewest companies already in the CRM, rotating so the same ones don't repeat every day.
If an Elevated category is running dry, say so in the summary and suggest broadening it.
To change focus next month, edit the cells. That's the whole mechanism.
Rotation by fewest matches. Counting how many CRM records already carry each category, and picking the least covered, balances coverage automatically. The CRM is the counter.
Calendar rules. "Odd-numbered days: group 1. Even-numbered days: group 2. Weekends: revisit stuck records." These are predictable, so you always know what ran.
A decision log. A dated document recording decisions and withdrawn rules. Tell agents to read it, and that a newer documented decision outranks older instruction text, as long as they say so in the summary. The one exception: a newer decision can never make the agent less cautious about sending or writing; for that, the instructions themselves must change. This once saved us: an agent noticed its own instructions still described a method we had withdrawn, followed the newer decision, and flagged the contradiction.
Pre-approved budgets. "You may use a small number of paid lookups per run (say, five), only for high-priority candidates." This steers spending without needing someone to approve each lookup.
Counts in summaries. Ask agents to report numbers that would justify a change, for example "candidates who would have gone to the social channel if partners were included". The agent gathers the evidence; you decide.
Safety rules, field ownership and anything legal belong in the spec, not in a cell someone might edit casually. Steering dials should change what the agent works on, never whether it follows a rule.
Next: Keep Your Prompts in Sync.
