Tuning your ICP

Agents are only as good as the bar they score against. The ICP is that bar. It has a prose half and a structured half, plus a feedback loop that sharpens it as you triage.

The prose brief

The free-text brief is what the judge actually reads, so write it the way you'd brief a new SDR: who buys, what they look like when they're about to buy, what's never a fit. Specifics beat adjectives. "10 to 100 people, running outbound, hiring their first RevOps role" gives the judge something to check; "innovative growing companies" doesn't.

The structured knobs

• Company size bounds: enforced in the free prefilter where the source exposes team size, and read by the judge everywhere else.
• Keywords: signals that matter, weighed by the judge.
• Excludes: hard "never a fit" terms, enforced deterministically before any scoring. The cheapest filter you have.
• Persona: who you want to reach at a fitting company ("regional managers", "head of sales"). Read by the judge, the person resolver, and both Claude scouts.

The feedback loop

When you dismiss a lead from the board, leadrack extracts the trait that made it a miss (an industry, a role shape) and can turn it into an exclude rule, so the same kind of miss stops arriving. Accepted and dismissed leads also accumulate into field notes, a digest of what past hunts found and how your team judged it, which the planning and scouting prompts read on later runs. The fleet you have in month two is scoring against more than the brief you wrote on day one.

When to edit the brief itself

If a whole category of misses keeps arriving, don't fight it one dismissal at a time. Say it in the brief. The judge follows the brief closely, and one added sentence ("not agencies; not companies selling to consumers") outweighs a pile of accumulated excludes. Keep the brief the source of truth and let the knobs handle the sharp edges.