| name | code-search-selector |
| description | โก AUTO-INVOKE when user asks: 'audit', 'investigate', 'how does X work', 'find all', 'where is', 'trace', 'understand', 'map the codebase', 'comprehensive'. MUST run BEFORE Read/Glob when planning to read 3+ files. Prevents tool familiarity bias toward native tools. |
| allowed-tools | Bash, Read, AskUserQuestion |
โ MANDATORY CODE SEARCH GATE โ
โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
โ โ
โ โก THIS SKILL AUTO-TRIGGERS ON THESE KEYWORDS: โ
โ โ
โ "audit" | "investigate" | "how does X work" | "find all" | "where is" โ
โ "trace" | "understand" | "map the codebase" | "comprehensive" โ
โ "all integration points" | "find implementations" | "architecture" โ
โ โ
โ ๐ซ INTERCEPTION: Triggers when about to Read 3+ files OR Glob broadly โ
โ โ
โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
Why This Gate Exists
The Tool Familiarity Bias Problem:
You have "native" tools (Read, Glob, Grep) that are always available with predictable output. These feel safe. But they produce INFERIOR results for semantic queries.
The "Known File Path" Trap:
When a prompt mentions specific file paths, your instinct is to Read directly. RESIST THIS. Semantic search provides CONTEXT around those files that direct reads miss.
The Parallelization Excuse:
"Let me Read files while agents work" is inefficient. Claudemem's indexed data is FASTER and provides better context.
This skill ensures you use the RIGHT tool for code search tasks. Using Grep when claudemem is indexed is a critical mistake that produces inferior results.
The Problem This Solves
โ WRONG: User asks "How does authentication work?"
โ You use: grep -r "auth" src/
โ Result: 500 lines of noise, no understanding
โ
RIGHT: User asks "How does authentication work?"
โ You check: claudemem status
โ You use: claudemem search "authentication login flow JWT"
โ Result: Top 10 semantically relevant code chunks
MANDATORY Decision Tree
Step 1: Classify the Task
โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
โ WHAT IS THE USER ASKING? โ
โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโค
โ โ
โ "Find all X" โ SEMANTIC (go to Step 2) โ
โ "How does X work" โ SEMANTIC (go to Step 2) โ
โ "Audit X integration" โ SEMANTIC (go to Step 2) โ
โ "Map the data flow" โ SEMANTIC (go to Step 2) โ
โ "Understand architecture" โ SEMANTIC (go to Step 2) โ
โ "Trace X through code" โ SEMANTIC (go to Step 2) โ
โ "Find implementations" โ SEMANTIC (go to Step 2) โ
โ "What patterns are used" โ SEMANTIC (go to Step 2) โ
โ โ
โ "Find exact string 'foo'" โ EXACT MATCH (use Grep, skip tree) โ
โ "Count occurrences of X" โ EXACT MATCH (use Grep, skip tree) โ
โ "Find symbol UserService" โ EXACT MATCH (use Grep, skip tree) โ
โ โ
โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
Step 2: Check claudemem Status (MANDATORY for Semantic)
claudemem status
Interpret the output:
| Status | What It Means | Next Action |
|---|
| Shows chunk count (e.g., "938 chunks") | โ
Indexed | USE CLAUDEMEM (Step 3) |
| "No index found" | โ Not indexed | Offer to index (Step 2b) |
| "command not found" | โ Not installed | Fall back to Detective agent |
Step 2b: If Not Indexed, Offer to Index
AskUserQuestion({
questions: [{
question: "Claudemem is not indexed. Index now for better semantic search results?",
header: "Index?",
multiSelect: false,
options: [
{ label: "Yes, index now (Recommended)", description: "Takes 1-2 minutes, enables semantic search" },
{ label: "No, use grep instead", description: "Faster but less accurate for semantic queries" }
]
}]
})
If user says yes:
claudemem index -y
Step 3: Execute the Search
IF CLAUDEMEM IS INDEXED (from Step 2):
claudemem ai developer
claudemem search "authentication login JWT token validation" -n 15
IF CLAUDEMEM IS NOT AVAILABLE:
Use the detective agent:
Task({
subagent_type: "code-analysis:detective",
description: "Investigate [topic]",
prompt: "Use semantic search to find..."
})
Step 4: NEVER Do This
โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
โ โ FORBIDDEN when claudemem is indexed: โ
โ โ
โ grep -r "pattern" src/ # Use claudemem search instead โ
โ Grep tool for semantic queries # Use claudemem search instead โ
โ Glob to find implementations # Use claudemem search instead โ
โ find . -name "*.ts" | xargs... # Use claudemem search instead โ
โ โ
โ These tools are for EXACT MATCHES only, not semantic search. โ
โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
Task-to-Tool Mapping Reference
| User Request | โ DON'T Use | โ
DO Use |
|---|
| "Audit all API endpoints" | grep -r "router|endpoint" | claudemem search "API endpoint route handler" |
| "How does auth work?" | grep -r "auth|login" | claudemem search "authentication login flow" |
| "Find all database queries" | grep -r "prisma|query" | claudemem search "database query SQL prisma" |
| "Map the data flow" | grep -r "transform|map" | claudemem search "data transformation pipeline" |
| "What's the architecture?" | ls -la src/ | claudemem search "architecture layer service" |
| "Find error handling" | grep -r "catch|error" | claudemem search "error handling exception" |
| "Trace user creation" | grep -r "createUser" | claudemem search "user creation registration" |
When Grep IS Appropriate
โ
Use Grep for:
- Finding exact string:
grep -r "DEPRECATED_FLAG" src/
- Counting occurrences:
grep -c "import React" src/**/*.tsx
- Finding specific symbol:
grep -r "class UserService" src/
- Regex patterns:
grep -r "TODO:\|FIXME:" src/
โ Never use Grep for:
- Understanding how something works
- Finding implementations by concept
- Architecture analysis
- Tracing data flow
- Auditing integrations
Integration with Detective Skills
After using this skill's decision tree, invoke the appropriate detective:
| Investigation Type | Detective Skill |
|---|
| Architecture patterns | code-analysis:architect-detective |
| Implementation details | code-analysis:developer-detective |
| Test coverage | code-analysis:tester-detective |
| Bug root cause | code-analysis:debugger-detective |
| Comprehensive audit | code-analysis:ultrathink-detective |
Quick Reference Card
โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
โ CODE SEARCH QUICK REFERENCE โ
โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโค
โ โ
โ 1. ALWAYS check first: claudemem status โ
โ โ
โ 2. If indexed: claudemem search "semantic query" โ
โ โ
โ 3. For exact matches: Grep tool (only this case!) โ
โ โ
โ 4. For deep analysis: Task(code-analysis:detective) โ
โ โ
โ โ ๏ธ GREP IS FOR EXACT MATCHES, NOT SEMANTIC UNDERSTANDING โ
โ โ
โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
Pre-Investigation Checklist
Before ANY code investigation task, verify:
๐ซ MULTI-FILE READ INTERCEPTION
โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
โ STOP BEFORE BULK FILE OPERATIONS โ
โ โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโฃ
โ โ
โ INTERCEPT TRIGGER: Before executing any of these: โ
โ โ
โ โข Read 3+ files in same directory โ
โ โข Glob with broad patterns (**/*.ts, **/*.py) โ
โ โข Sequential reads to "understand" a feature โ
โ โข "Let me read files while agents work" โ
โ โ
โ ASK YOURSELF: โ
โ 1. Is claudemem indexed? (claudemem status) โ
โ 2. Can this be ONE semantic query instead of N file reads? โ
โ 3. Am I falling into tool familiarity bias? โ
โ โ
โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
Interception Examples
โ About to do:
Read src/services/auth/login.ts
Read src/services/auth/session.ts
Read src/services/auth/jwt.ts
Read src/services/auth/middleware.ts
Read src/services/auth/types.ts
Read src/services/auth/utils.ts
โ
Do instead:
claudemem search "authentication login session JWT middleware" -n 15
โ About to do:
Glob pattern: src/services/prime/**/*.ts
Then read all 12 matches sequentially
โ
Do instead:
claudemem search "Prime API integration service endpoints" -n 20
โ Parallelization trap:
"Let me Read these 5 files while the detective agent works..."
โ
Do instead:
Trust the detective agent to use claudemem.
Don't duplicate work with inferior Read/Glob.
๐ด ANTI-PATTERNS TO AVOID
| Anti-Pattern | Why It's Wrong | Correct Alternative |
|---|
| Reading 5+ files sequentially | Token waste, no ranking | claudemem search once |
| Glob โ Read all matches | No semantic understanding | claudemem search with concept |
| "Files mentioned, let me Read" | Misses context around files | Search semantically first |
| Grep for "how does X work" | Text match โ meaning | claudemem search |
| Read while agents work | Duplicate inferior work | Trust agent's claudemem usage |
โ
CORRECT WORKFLOW
โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
โ CORRECT INVESTIGATION FLOW โ
โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโค
โ โ
โ 1. TASK ARRIVES with keywords: โ
โ "audit", "investigate", "how does", "find all", etc. โ
โ โ
โ 2. AUTO-TRIGGER this skill (code-search-selector) โ
โ โ
โ 3. CHECK: claudemem status โ
โ โข If indexed โ Use claudemem search โ
โ โข If not โ Index first OR launch detective agent โ
โ โ
โ 4. SEARCH SEMANTICALLY: โ
โ claudemem search "concept query" -n 15 โ
โ โ
โ 5. ONLY THEN Read specific files/lines from results โ
โ โ
โ โ ๏ธ NEVER start with Read/Glob for semantic tasks โ
โ โ
โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
Maintained by: tianzecn
Plugin: code-analysis v2.2.0
Purpose: Prevent tool familiarity bias, intercept multi-file reads, enforce semantic search
