taste-learning

name	taste-learning
description	Tier 1 learning — detects user output preferences from repeated corrections and injects them as instructions into future plugin runs.

Overview

Taste learning detects patterns in how users modify or react to plugin outputs, then automatically adjusts future outputs to match preferences.

Detection Mechanism

1. Post-command observation: After a command runs, the post_command event records the output
2. User correction detection: If the user edits, rejects, or re-runs the command with different parameters, record a decision_point event noting the change
3. Pattern matching: When the same type of correction is observed 3+ times for the same plugin/command, create a candidate pattern
4. Distillation: Convert the repeated correction into an instruction (e.g., "user shortened drafts 4/5 times" → "Keep email drafts concise, under 150 words")

What Taste Learning Captures

Signal	Example	Resulting Pattern
Output length corrections	User consistently shortens briefings	"Keep briefings under 500 words"
Tone adjustments	User rewrites drafts to be more formal	"Use formal business tone in email drafts"
Prioritization overrides	User re-orders urgent items	"Prioritize revenue-related items first"
Formatting preferences	User restructures output into bullets	"Use bullet-point format for task summaries"
Content inclusion/exclusion	User removes calendar section from briefing	"Exclude calendar items from daily briefing"

Creating a Taste Pattern

When distilling a pattern from observations:

1. Query recent decision_point events for this plugin/command
2. Group by similar correction types
3. If a correction type appears 3+ times with consistent direction:
   • Create a new row in the patterns table
   • pattern_type: "taste"
   • plugin: the specific plugin
   • command: the specific command (or null if applies to all commands)
   • description: human-readable summary of the preference
   • instruction: the text to inject into future command context
   • confidence: 0.0 (starts as candidate)
   • observations: count of times this pattern was seen
4. Subsequent observations update observations, confirmations, rejections, and recalculate confidence

Confirmation vs Rejection

• Confirmation: User runs the command, pattern is applied, user accepts the output without changes
• Rejection: User runs the command, pattern is applied, user undoes/overrides the pattern's effect
• Track via post_command events: if outcome is "success" and no subsequent correction → confirmation
• If user re-runs or explicitly corrects → rejection

Notification

Always notify the user when a taste pattern is applied:

[Intel] Applying learned preference: "concise email drafts under 150 words"

If the user says "stop applying this" or rejects 3 times in a row, set status to "rejected".