Extract and search human decisions from Claude Code sessions

version
1.4.0
GitHub stars
6

Install

/plugin install superwiser@mu4farooqi

Run it inside Claude Code after adding the marketplace with /plugin marketplace add mu4farooqi/superwiser.

Contents

Discovered from source

Skills 2

Conflict Resolution

When any Superwiser tool returns results containing "⚠️ CONFLICT [id]", you must resolve the conflict before proceeding.

What Is a Conflict?

A conflict occurs when the user has expressed contradictory preferences at different times. For example:

  • "Always use tabs for indentation" vs "Use 2-space indentation"
  • "Prefer functional components" vs "Use class components for state"

How to Resolve

  1. Present the conflict clearly to the user:

    SuperWiser Conflict [abc123]:
    I found conflicting preferences:
    1. "Always use tabs for indentation" (from when you were setting up the project)
    2. "Use 2-space indentation" (from your React work last week)
    
    Which should I follow? Or do they apply in different contexts?
    
  2. Wait for the user's response. They can:

    • Pick a number (e.g., "1" or "use the first one")
    • Say "both" or "all" if both rules should be kept
    • Explain when each applies (e.g., "tabs for Python, spaces for JS")
    • Say "skip" to ignore for now
  3. After the user responds, continue with your work. You do NOT need to:

    • Call any tools to resolve or delete rules
    • Process the response in any special way
    • The conflict resolution is handled automatically in the background

Important

  • Do NOT skip conflicts - they indicate important ambiguity in user preferences
  • Do NOT guess which rule to follow - always ask the user
  • Do NOT call any MCP tools to "resolve" the conflict - just continue after they respond

Superwiser - Load Coding Preferences

When this skill is activated, call the load_preferences MCP tool to retrieve the user's coding preferences.

Usage

Call the tool with context about what the user is working on:

load_preferences(task_context="<describe what user is working on>")

What Gets Loaded

  1. Global preferences - Universal rules that apply everywhere (style, patterns, conventions)
  2. Contextual preferences - Rules specific to the current task (matching semantic similarity)

After Loading

Present the loaded preferences to the user in a clear format, then apply them throughout the session.

Commands 12

Show the user's current Superwiser configuration.

Instructions

Call the get_config MCP tool from superwiser.

Display the results in a readable format showing each setting with its current value and description.

Remove a rule from the database.

Usage

/superwiser:delete <rule_id>

Instructions

The rule_id is a 6-character alphanumeric ID like x7k9m2.

Call the delete_rule MCP tool from superwiser with the provided rule_id.

Confirm the deletion to the user.

Initialize the superwiser plugin manually. Use this if you installed the plugin mid-session.

Instructions

Call the initialize MCP tool from superwiser.

This will:

  • Install any missing dependencies
  • Start the background worker if not running
  • Register the current project
  • Initialize the project database

After running, confirm to the user that superwiser is ready.

Show captured rules with optional limit, sorting, and filtering.

Instructions

Call the list_rules MCP tool from superwiser.

Parameters:

  • limit: Number of rules to show (default: 10). If user says "top 5" or "list 20", extract the number.
  • sort_by: Sorting option
  • preferences_only: If True, only show universal rules without context

Sorting options (pass as sort_by parameter):

  • recent (default) - newest rules first
  • important - highest importance score first
  • hits - most searched rules first

Preferences filter: If user says "preferences", "preferences only", or "without context", pass preferences_only=True.

Examples:

  • /superwiser:list → list_rules() (10 recent)
  • /superwiser:list 5 → list_rules(limit=5)
  • /superwiser:list important → list_rules(sort_by="important")
  • /superwiser:list 20 hits → list_rules(limit=20, sort_by="hits")
  • /superwiser:list preferences → list_rules(preferences_only=True)
  • /superwiser:list 10 important preferences → list_rules(limit=10, sort_by="important", preferences_only=True)

Important: Do not display the raw JSON result from the tool. Present the results in a clean table format with columns for ID, Rule, Context, Confidence, Date, and Score. Also show the total count of rules.

Pause Superwiser preference recording for this project. This persists across sessions.

Instructions

Call the disable_recording MCP tool from superwiser.

Tell the user:

  • Recording is now paused for this project
  • This will persist across sessions
  • Use /superwiser:record to resume recording
  • Existing preferences are preserved and searchable

Resume Superwiser preference recording for this project.

Instructions

Call the enable_recording MCP tool from superwiser.

Tell the user:

  • Recording is now active
  • New preferences will be captured from your messages
  • Use /superwiser:record-stop to pause recording

Search Rules

Search your recorded coding rules and decisions.

Usage

/superwiser:search <query> [limit:N] [preferences]

Examples

/superwiser:search database choice for user sessions
/superwiser:search error handling patterns for REST APIs
/superwiser:search typescript strict mode configuration
/superwiser:search python testing conventions limit:10
/superwiser:search coding style preferences

Instructions

Call the search_rules MCP tool from superwiser with the user's search terms.

Parameters:

  • context: What to search for (required). Use a detailed 1-2 sentence description for best results.
  • limit: Max results to return (default: 5)
  • preferences_only: If True, only return universal rules without context (default: False)

If the user says "top 10" or "limit:10", pass the number as the limit parameter.

If the user says "preferences", "preferences only", or "without context", pass preferences_only=True.

Important: Do not display the raw tool result. Present the results in a clean, readable format with the rule text, context, confidence level, and ID.

Seed the Superwiser rule database from your past Claude Code conversations.

Instructions

Step 1: Show transcript preview

Call seed_preview MCP tool from superwiser.

Present the results to the user like this:

Found [count] transcripts spanning [oldest_date] to [newest_date].

How would you like to seed?
1. Latest N transcripts (e.g., "latest 5")
2. All after a date (e.g., "after 2025-06-01")
3. All [count] transcripts

Step 2: Wait for user choice

The user will respond with one of:

  • "latest N" (e.g., "latest 5", "latest 10")
  • "after YYYY-MM-DD" (e.g., "after 2025-06-01")
  • "all"

Step 3: Call seed_from_history with their filter

Based on the user's choice:

  • For "latest N": seed_from_history(latest_n=N)
  • For "after DATE": seed_from_history(after_date="YYYY-MM-DD")
  • For "all": seed_from_history() (no arguments)

Step 4: Report results

Tell the user:

  • How many prompts were queued
  • That processing happens in the background
  • They can use /superwiser:list to check progress

Change a Superwiser configuration setting.

Instructions

Parse the user's input to extract the setting name and value, then call set_config with those arguments.

Example inputs:

  • /superwiser:set-config extraction_model opus
  • /superwiser:set-config extraction_concurrency 4

If no arguments provided, call get_config to show available settings.

Available Settings

SettingValues
dynamic_context_enabledtrue, false
dynamic_context_interval60-900 (seconds)
extraction_modelsonnet, opus, haiku
extraction_concurrency1-10
preference_global_limit1-20
preference_contextual_limit1-20
preference_min_score0.0-1.0
default_search_limit1-50
context_max_lines100-2000
min_prompt_length5-100
semantic_weight0.0-1.0
discovery_interval1-90 (days)
discovery_timeout60-600 (seconds)
discovery_modelsonnet, opus, haiku

Star Superwiser

Star the Superwiser repository on GitHub to show your support.

Arguments

  • No argument: Star the repo
  • never: Stop showing star reminders

Instructions

If argument is "never"

Update ~/.config/superwiser/onboarding.json to set star_never_ask to true:

import json
from pathlib import Path

config_file = Path.home() / '.config' / 'superwiser' / 'onboarding.json'
config = json.loads(config_file.read_text()) if config_file.exists() else {}
config['star_never_ask'] = True
config_file.parent.mkdir(parents=True, exist_ok=True)
config_file.write_text(json.dumps(config, indent=2))

Respond: "No problem! Star reminders are now disabled."

If no argument (star the repo)

  1. Check if gh CLI is installed and authenticated:

    gh auth status
    
  2. If authenticated, check if already starred:

    gh api user/starred/mu4farooqi/superwiser
    
    • Exit code 0 = already starred
    • Exit code 1 = not starred
  3. If not starred, star the repo:

    gh api -X PUT user/starred/mu4farooqi/superwiser
    
  4. Report result to user:

Show statistics about captured rules and API token usage.

Instructions

Call the get_stats MCP tool from superwiser.

Present the results in a readable format:

  • Show total rule count
  • Display pending conflicts if any (only if non-zero)
  • Show pending/processing/failed queue counts (only if non-zero)
  • Show token usage: today and last 7 days with daily average
  • Show breakdown by operation (discovery vs extraction)

Example output format:

Rule Statistics:
- Total rules: 47
- Pending conflicts: 2
- Queue: 3 pending, 1 processing

Token Usage:
Today: 57,730 tokens (8 calls)
  Input: 38,420 | Output: 6,810

Last 7 Days: 312,450 tokens (67 calls)
  Daily avg: 44,635 tokens

By Operation (7d):
  Discovery: 125,000 tokens (12 calls)
  Extraction: 187,450 tokens (55 calls)

If there are failed items, show: "- Failed: 2 (check worker logs)" If token_usage has an error, show: "Token tracking: Not yet initialized"

Browse by Tag

Show all tags and their counts from captured rules.

Instructions

Call the list_tags MCP tool from superwiser.

Important: Do not display the raw JSON result from the tool. Present the tags in a clean table format with columns for Tag and Count.

To filter rules by a specific tag, the user can use: /superwiser:search #<tagname>

MCP servers 1

Sponsors

Details

Marketplace
mu4farooqi/superwiser
Source
plugins/superwiser
Author
Omar Ahmad
Category
Community