interfluence: Refine Voice Profile

interfluence: Refine Voice Profile

You are helping the user review and refine their voice profile through interactive dialogue.

Process

1. Load current state:

   • Get available voices using profile_list
   • Get the base profile using profile_get
   • Get config using config_get (for voice routing patterns)
   • Get accumulated learnings using learnings_get_raw

2. If there are accumulated learnings:

   • Analyze the edit diffs to extract stylistic patterns
   • Resolve context for each learning: Learnings are tagged CONTEXT:unknown. Use the file path from each entry and match it against the current config's voices array (first match wins) to determine which context the learning belongs to.
   • Group learnings by resolved context (blog, docs, base/unresolved)
   • If any learnings can't be resolved, present them to the user: "3 learnings from unknown contexts — assign them to: [blog] [docs] [base] [skip]?"
   • For each pattern found, propose a voice profile update:
     • "Based on your edits to blog posts, you seem to prefer X over Y. Should I add this to the blog voice?"
     • "This pattern appears across contexts — should it go in the base profile?"
   • Present proposed updates for user approval

3. Show the current profile(s) and ask the user what they'd like to adjust:

   • Show which voices exist (base, blog, docs, etc.)
   • Are there patterns that feel wrong or overstated?
   • Are there voice traits missing from the profile?
   • Any specific "never do this" rules to add?
   • Should any refinement apply to a specific voice or the base?

4. For each refinement:

   • Show the before/after for the relevant profile section
   • Clarify which voice it applies to (base affects all contexts, delta affects only that context)
   • Confirm with the user before saving

5. Save the updated profile(s):

   • Base updates: profile_save(projectDir, content)
   • Voice delta updates: profile_save(projectDir, content, "<voice>")

6. Clear processed learnings using learnings_clear_raw

7. Optionally demonstrate the refined profile by showing a before/after of a short passage

Important

• The user's explicit preferences always override automated learnings
• Don't consolidate or remove profile entries without asking
• Keep the profile readable and specific — resist vague generalizations
• If the user says "that's not how I'd say it" about specific text, extract the lesson and add it
• When adding to a delta vs base, ask: "Does this apply everywhere, or just in [context]?"