icp-refinement-loop

ICP Refinement Loop

Periodically refine the ICP scorecard using actual won/lost patterns from ≥30 closed deals. Recomputes tier cutoffs, re-tunes dimension weights, surfaces segment shifts. Hard rule: refusal to refine ICP with <30 closes — insufficient signal.

When to Use

• Quarterly ICP refinement — ≥30 closed deals since last refresh
• Forecast accuracy shows systemic over/under-prediction by segment
• Win rate by tier doesn't match scorecard predictions
• Team expanded into new vertical — does ICP capture this
• User says "refine ICP" or "quarterly ICP audit"

Quick Reference

Concept	Value
Min-closes hard floor	30 (won + lost combined); below = refusal
Refinement scope	Cutoff recalibration / weight re-tuning / segment shift / anti-ICP update / trigger refresh
Tier cutoff rule	>25% wins below T1-cutoff → cutoff too high; >25% losses at T1 → overconfident
Dimension re-tuning	Correlate each dimension score with closed_won outcome; adjust weights preserving total 100
New segment requirement	≥10 wins in segment not captured by original ICP
Confidence upgrade	Hypothesis → Medium (n≥30) / High (n≥100)

Procedure

1. Validate inputs. Pull closed deals; confirm n≥30. Below → refusal with explanation. See ${HERMES_SKILL_DIR}/references/thresholds.md.
2. Score won/lost deals against current rubric. Apply original icp-definition rubric to each closed deal as it was at time of close. Compute distribution.
3. Tier cutoff calibration. Wins scoring <T1 → cutoff may be too high. Losses at T1 → cutoff may be too low. Surface recommendations if >25% misalignment. See ${HERMES_SKILL_DIR}/references/calibration-rules.md.
4. Dimension weight re-tuning. Per dimension: compute correlation with closed_won. High correlation → increase weight. Near-zero → decrease. Preserve total 100. See ${HERMES_SKILL_DIR}/references/weight-tuning.md.
5. Segment shift detection. Group wins by firmographic. If segment not in original ICP captures ≥10 wins → recommend new sub-segment.
6. Lost-reason aggregation. Aggregate lost reasons by tier + segment. lost-to-competitor → competitive-intelligence. no-budget by size → revenue-forecasting.
7. Compose ICP delta recommendations. Rank by impact. Each: evidence + proposed change + estimated impact. See ${HERMES_SKILL_DIR}/references/delta-template.md.
8. Retroactive rescoring proposal. Compute: if new rubric applied to active pipeline, how do tiers shift? Surface counts. Don't auto-apply.
9. Push + emit refinement record. User authorizes which recommendations to apply. Route updates to icp-definition + lead-scoring. See ${HERMES_SKILL_DIR}/scripts/push_to_crm.py.

Pitfalls

• Refining with <30 closes — insufficient signal; refusal is right
• Auto-applying ICP changes — always require user authorization
• Overfitting to recent wins — good rubric generalizes; don't optimize to last 30 deals at expense of generalizability
• Adding segments on n=2 — "we won two Healthcare deals" ≠ Healthcare ICP yet
• Ignoring lost-deal patterns — losses teach which to avoid
• Re-running too often — quarterly minimum; more frequent = noise + whiplash
• Not coordinating with positioning — ICP shifts imply positioning shifts

Verification

1. Min-closes threshold honored
2. Per-deal scores reproducible from input + original rubric
3. Recommendations cite actual deal counts (sample-check 5)
4. Retroactive rescoring numbers add up to pipeline count
5. Lost-reason routing happened
6. User authorization queue surfaced before any rubric update