| name | blindspot-miner |
| description | Use this when analysis is complete. Mine for findings the subject cannot see from inside any single session. |
Blindspot Miner
The Search Grid
Generate candidates across ten grids:
- LONGITUDINAL - visible only across months
- CROSS-PROJECT - same behavior in unrelated contexts
- CROSS-MODEL - behavior constant despite different assistants
- CONTRADICTIONS - stated preference vs measured behavior
- MIS-CALIBRATIONS - expected vs actual divergence
- HIDDEN BOTTLENECKS - apparent problem is not where time and rework go
- UNDERUSED STRENGTHS - good-outcome behaviors applied inconsistently
- COSTLY DEFAULTS - reasonable sometimes, applied automatically always
- AI-INDUCED BEHAVIOR - specific assistants improving or degrading process
- UNBUILT SYSTEMS - lessons corrected manually that never became tooling
Plus two uncomfortable hunts: productive-feeling-but-empty work (outcome
evidence required) and undervalued high-leverage habits.
The Kill Question
For every candidate ask:
Would this still be interesting if the subject already knew it?
If no, deprioritize. The report lives or dies on non-obviousness.
Evidence Standard
Every significant insight carries all fields:
OBSERVATION / EVIDENCE / FREQUENCY / DATE RANGE / PLATFORMS /
REPRESENTATIVE EPISODES / COUNTEREVIDENCE / ALTERNATIVE EXPLANATIONS /
CONFIDENCE / PRACTICAL CONSEQUENCE
Apply the 3x3x3 principle: a finding across 3 independent tasks, 3 months,
3 different contexts beats one vivid anecdote every time.
Confidence ladder: HIGH, MEDIUM, LOW, SPECULATIVE. Never dress LOW up as
fact. Do not cherry-pick dramatic conversations.
One-Line Memory
If they would nod immediately, it isn't a finding.
