| name | collaboration-audit |
| description | Use this to map the full human-AI working loop - prompts in, verification out, understanding kept or leaked. |
Collaboration Audit
Lens 1: Inputs
Judge prompts by functional properties, never by grammar, brevity,
informality, or voice dictation:
- desired outcome present
- relevant context present
- constraints present
- examples present
- success criteria present
- authority boundaries present
- verification requested
- unnecessary instructions
- conflicting instructions
Find the highest-performing prompt patterns in this corpus - the ones
associated with fewer correction loops.
Then list instructions the subject retypes manually again and again.
Each one is a candidate repo context file, skill, test, or guardrail.
Memory is the wrong storage for repeated rules.
Lens 2: Output Handling
After plausible AI output arrives, classify the response: inspect, test,
ask why, request sources, run a reproduction, compare alternatives,
implement immediately, ship immediately.
Measure verification depth against risk level. Hunt both directions:
- AI SOUNDED CORRECT but result was wrong - what check was skipped
- excessive verification burned hours on a reversible, cheap decision
The goal is calibrated checking, not maximum checking.
Lens 3: Understanding Ledger
Map tasks onto delegation levels: autocomplete, implementer, researcher,
debugger, architect, decision adviser, autonomous operator.
Then locate:
- delegation that saved weeks
- delegation that manufactured rework
- where the subject stays intellectually involved and thrives
- where evaluation replaced understanding and cost them later
- concepts repeatedly re-explained in later sessions
- where AI visibly accelerated genuine learning
- where the underlying system became a black box
One-Line Memory
Calibrated checking beats maximum checking.
