Learn

Personal Preference Learning

Extract preferences, styles, and behavioral patterns from the current conversation and persist them for all future sessions.

Scope: This skill captures how the user works, writes, and wants agents to behave. It does NOT capture code patterns, error fixes, or debugging techniques.

Three-Tier Target Architecture

~/CLAUDE.md should stay under 200 lines. When it grows, move coherent rule groups to ~/.claude/rules/.

Workflow

Step 1: Scan conversation

Review all user messages for extractable patterns. Use the phrase families below as signals — do not infer preferences from silence or context alone.

Explicit rules (highest priority, save immediately):

• "always…", "never…", "from now on…", "I prefer…"
• "remember that I…", "make sure you…"

Corrections (save if repeated, see Step 2):

• "don't do X", "stop doing Y", "that's not right"
• "actually, it should be…", "I told you before…"
• User rewording or editing Claude's output in place

Positive reinforcement (save when the approach was non-obvious):

• User thanking Claude for a specific behavior
• Accepting an unusual choice without pushback (quieter signal — note why it was non-obvious)

Domain context (save as user-tier facts):

• Role, expertise, institutional affiliation, teaching approach
• Tools, frameworks, or workflows the user commits to long-term

Style edits (save as style rules):

• User modifying prose, formatting, or structure Claude produced
• Terminology corrections ("use X not Y")

Non-signals — skip these:

• One-time instructions ("do X now", "for this file only")
• Hypotheticals ("what if I wanted…", "could you also…")
• Third-party preferences ("my coauthor likes…", "reviewers want…")
• Silence (user didn't push back ≠ user approved)
• Code error resolutions and library workarounds (covered by continuous-learning)
• Patterns Claude's auto memory already captured this session

Step 2: Apply the repetition threshold

A single off-hand correction is not a durable preference. Before promoting any pattern to later steps:

• Explicit rule ("always…", "never…", "from now on…") → save immediately
• Repeated correction (same pattern corrected 2+ times this session or referenced as "I told you before") → save as tentative, note the repetition
• One-time correction with no repetition and no explicit framing → skip, do not fabricate a rule
• Inferred from silence or hypothetical → skip unconditionally (see ## Anti-Patterns below)

If a tentative pattern has been corrected 3+ times across sessions (check existing ~/CLAUDE.md for near-duplicates), surface it in Step 6 with a "Confirm as permanent rule?" flag so the user decides explicitly.

Step 3: Read current state

1. Read ~/CLAUDE.md in full. Parse the section hierarchy (it evolves — always re-read).
2. List files in ~/.claude/rules/ to check existing rule files.
3. Read the current project's ~/.claude/projects/<project>/memory/MEMORY.md if it exists.
4. Identify what is already captured across all three tiers to avoid duplication.

Step 4: Detect conflicts with existing preferences

For each extraction, search the already-loaded ~/CLAUDE.md and rule files for contradicting text. If a conflict is found:

1. Do NOT silently overwrite the existing rule.
2. In Step 6's preview, show both entries side-by-side under a ## Conflicts heading and ask which supersedes which.
3. Resolution defaults: more specific scope wins (project > user rule > ~/CLAUDE.md); within the same tier, the newer instruction wins only after explicit approval.

Step 5: Classify each extraction

For each pattern that survived Steps 1–4, decide its tier:

Categories and typical tiers:

Step 6: Present proposed changes

Show a numbered preview grouped by tier:

## Tier 1: User Instructions (~/CLAUDE.md)

### [1] Add to "## Academic Writing > ### Style" (existing section)
+ - Do not use semicolons to join independent clauses in academic prose

### [2] New section: "## Workflow" (after last existing ## section)
+ - When re-running a pipeline, first check what outputs already exist in cache
+ - Prefer flat directory structures keyed by ID over nested semantic hierarchies

## Tier 2: User Rules (~/.claude/rules/)

### [3] New file: ~/.claude/rules/communication.md
+ # Communication Preferences
+ - Lead with the answer, not the reasoning

## Tier 3: Project Memory

### [4] Add to project memory MEMORY.md
+ - TGL project uses SDT framework (autonomy, competence, relatedness) scored 1-7

## Skipped

- "Use precise terminology" → already in ~/CLAUDE.md ### Word Choice

Approve all? Or specify numbers to modify/reject.