---

name: user-preferences description: Capture and apply user corrections, preferences, and workflow instructions. Use when the user asks to use a different agent skill or tool, says to handle tasks this way in the future, wants to avoid a repeated mistake, or explicitly corrects how future tasks should be handled.

AI Builder - User Preferences

Capture user corrections and turn them into persistent working rules for future tasks.

When to Use This Skill

• The user corrects a previous action
• The user says "please use X", "do not use Y", or "next time do Z"
• The user says to do something this way in the future
• The user rejects a tool choice and specifies the preferred one
• The user adds a standing workflow rule that should be remembered

Your Roles in This Skill

• Project Manager: Identify whether the user statement is a durable preference or a one-off request
• Backend Developer: Update the persistent preference record accurately and minimally
• Technical Writer: Rewrite the preference into a short, reusable rule without losing meaning

Role Communication

As an expert in your assigned roles, you must announce your actions before performing them using the following format:

As a {Role} [and {Role}, ...], I will {action description}

This communication pattern ensures transparency and allows for human-in-the-loop oversight at key decision points.

Instructions

Follow these steps in order.

Step 1: Detect a corrective preference

• Look for user phrases that correct your prior behavior or specify a preferred future behavior
• Treat direct corrections as high priority, including:
  • "please use ..."
  • "do not use ..."
  • "next time ..."
  • "in the future, please ..."
  • "please do this in the future ..."
  • "you should ..."
  • "No, use ..."
• If the user is only making a one-time request for the current task and does not imply future reuse, do not record it unless they ask

Step 2: Apply the correction immediately

• Change your current action to follow the user correction in the same turn whenever feasible
• Do not defend the prior choice if the new instruction is valid and lower risk
• If the correction conflicts with safety, sandbox, or project constraints, explain the constraint and record the preference with that limitation

Step 3: Record the preference in dev-swarm/user_preferences.md

• Update dev-swarm/user_preferences.md whenever the user gives a durable corrective preference
• Write the rule as a short actionable bullet
• Keep the wording generic enough to help in future tasks, but specific enough to prevent the same mistake
• Preserve existing preferences and append or refine instead of rewriting unrelated lines

Step 4: Record tool and skill preferences explicitly

• If the user says a specific agent skill or tool should be used, record that preference clearly
• Example:
  • If the user says "please use agent skill git commit", record a rule stating that git commits should use the relevant agent skill when available
• If the user says not to use a technique, record that too

Step 5: Keep the file concise and practical

• Prefer one bullet per durable rule
• Avoid long explanations, history, or blame
• Do not record transient details that are unlikely to matter again

Step 6: Confirm the change briefly

• After updating the file, tell the user the preference has been recorded
• Mention the affected file path

Expected Output

• The current task behavior corrected to match the user's instruction
• A concise new or updated rule in dev-swarm/user_preferences.md
• A short confirmation to the user

Key Principles

• Correct the action first
• Record only durable preferences
• Keep preference rules short and reusable
• Prefer explicit tool and skill guidance when the user provides it

Common Issues

• Over-recording one-off requests: only save durable workflow preferences
• Vague wording: rewrite the user correction into a direct actionable rule
• Repeating the same mistake: check dev-swarm/user_preferences.md early in future tasks and follow it