preference-intake

Preference Intake

Use this skill when the user wants the assistant to learn them better, capture durable preferences, or turn a conversation into a reusable user-profile.md / USER.md style document.

This is a classification and synthesis workflow:

• ask targeted questions first,
• identify durable preferences vs temporary context,
• write a clean profile document,
• update memory only for stable facts,
• keep transient details out of long-term storage.

When to use

• The user says: "learn me", "remember my preferences", "make a user profile", "build a USER.md", or similar.
• The user wants the assistant to adapt style, tone, verbosity, or decision-making to them.
• The user wants a durable personalization file for future sessions.

Core workflow

1. Start with a short intake interview.

   • Ask only the questions needed to learn stable preferences.
   • Prefer concise numbered questions.
   • Let the user answer in free form if they want.

2. Separate durable facts from temporary state.

   • Durable: name, language preference, tone, working style, environment, recurring goals, tool preferences.
   • Temporary: exam dates, one-off projects, short-lived experiments, session-specific context.
   • Do not force temporary items into the profile.

3. Synthesize into a markdown profile.

   • Use clear headings.
   • Group by: address, language, response style, domains, environment, decision style, feedback style, boundaries.
   • Keep it readable and easy to skim.

4. Store only stable facts in long-term memory.

   • Save durable preferences that will matter later.
   • Exclude secrets, credentials, and short-term tasks.

5. If the user asked for it, provide a separate summary of their ideas.

   • Put the summary in a distinct message or section.
   • Do not mix the idea-summary with the durable profile if that would make the profile noisy.

Question bank

Use only the subset that is relevant. Good defaults:

• How should I address you?
• What language should I use by default?
• How concise or detailed should I be?
• What do you want help with most?
• What are your current high-level goals?
• What response style do you dislike?
• What tools or environments should I know?
• What decision style do you prefer?
• What should I avoid remembering or mentioning?

Output shape

When creating the profile, include:

• user-profile.md content with headings
• a short bullet summary of what was learned
• a note about what was intentionally not saved if the user asked for exclusions

Pitfalls

• Do not overwrite temporary context into durable memory. Example: an exam deadline or a single project milestone may be important right now but should usually stay out of the permanent profile unless the user explicitly wants it remembered.

• Do not turn the profile into a transcript. The profile should be a compact operating manual for future sessions.

• Do not accept vague preferences as final if they need clarification. Example: "be helpful" is too broad; ask for how the user wants help in practice.

• Do not store secrets, passwords, card data, or one-time sensitive info.

Verification

Before finalizing:

• Check that the profile reflects stable preferences only.
• Check that sensitive or temporary items were excluded.
• Check that the wording is usable as a future-session reference.

Support files

• templates/user-profile.md — reusable markdown skeleton for a personal profile file.