/counsel:calibrate

Correct and refine expert simulations.

Usage

/counsel:calibrate "Hickey wouldn't say that — he'd question the abstraction entirely"
/counsel:calibrate osmani: he focuses more on user experience than raw performance

Inputs

• $1 — Correction description (required)
• Expert name can be explicit or inferred from recent context

Process

1. Acknowledge correction
2. Ask for specifics if needed: "What would [Expert] say instead?"
3. Log calibration to .claude/logs/counsel-calibrations.jsonl:

   {"ts": "...", "expert": "...", "wrong": "...", "correct": "...", "context": "...", "source": "..."}
   

4. Confirm: future simulations avoid "wrong" pattern, lean toward "correct"

Application

When simulating a calibrated expert: load calibrations, filter by expert, avoid wrong patterns, lean toward correct ones. Note "(adjusted per calibration)" when applied.

Constraints

• Only calibrate for specific, documentable corrections
• Do not change confidence score based on calibrations
• Calibrations are project-local (different projects may have different calibrations)
• Acknowledge uncertainty if correction seems to conflict with documented work