AgentSkillsCN

search-interceptor

⛔ 截止触发：在读取3个以上文件或使用广谱模式Glob之前自动调用。验证是否应将批量文件操作替换为语义搜索。防止连续读取文件造成令牌浪费。

SKILL.md
--- frontmatter
name: search-interceptor
description: "⛔ INTERCEPT TRIGGER: Automatically invoked BEFORE Read 3+ files OR Glob with broad patterns. Validates whether bulk file operations should be replaced with semantic search. Prevents token waste from sequential file reads."
allowed-tools: Bash, AskUserQuestion

Search Interceptor

code
╔══════════════════════════════════════════════════════════════════════════════╗
║                                                                              ║
║   ⛔ INTERCEPT TRIGGERS:                                                     ║
║                                                                              ║
║   • About to Read 3+ files in same directory                                ║
║   • About to Glob with **/*.ts, **/*.py, or similar broad pattern           ║
║   • Planning sequential file reads to "understand" something                 ║
║   • Rationalizing "let me read while agents work"                           ║
║                                                                              ║
║   WHEN TRIGGERED: Validate if claudemem search is better                    ║
║                                                                              ║
╚══════════════════════════════════════════════════════════════════════════════╝

Purpose

This skill intercepts bulk file operations before they execute, validating whether semantic search would be more efficient.

When This Skill Triggers

Trigger 1: Multiple File Reads Planned

code
YOU ARE ABOUT TO:
  Read file1.ts
  Read file2.ts
  Read file3.ts
  Read file4.ts
  ...

STOP. Ask: Can this be ONE claudemem query?

Trigger 2: Broad Glob Pattern

code
YOU ARE ABOUT TO:
  Glob("src/services/**/*.ts")
  Then read all N matches

STOP. Ask: What am I looking for SEMANTICALLY?

Trigger 3: Parallelization Rationalization

code
YOU ARE THINKING:
  "Let me read these files while the agent works..."

STOP. This is tool familiarity bias.

Trigger 4: File Paths in Prompt

code
PROMPT MENTIONS:
  src/services/prime/internal_api/client.ts
  src/services/prime/api.ts
  ...

YOUR INSTINCT: Read them directly
STOP. Search semantically first for context.

Interception Protocol

Step 1: Pause Before Execution

When you're about to execute bulk file operations, STOP and run:

bash
claudemem status

Step 2: Evaluate

If claudemem is indexed:

Your PlanBetter Alternative
Read 5 auth filesclaudemem search "authentication login session"
Glob all servicesclaudemem search "service layer business logic"
Read mentioned pathsclaudemem search "[concept from those paths]"

If claudemem is NOT indexed:

bash
claudemem index -y

Then proceed with semantic search.

Step 3: Execute Better Alternative

bash
# Instead of reading N files, run ONE semantic query
claudemem search "concept describing what you need" -n 15

# ONLY THEN read specific lines from results

Interception Decision Matrix

SituationIntercept?Action
Read 1-2 specific filesNoProceed with Read
Read 3+ files in investigationYESConvert to claudemem search
Glob for exact filenameNoProceed with Glob
Glob for pattern discoveryYESConvert to claudemem search
Grep for exact stringNoProceed with Grep
Grep for semantic conceptYESConvert to claudemem search
Files mentioned in promptYESSearch semantically first

Examples of Interception

Example 1: Auth Investigation

❌ Original plan:

code
I see the task mentions auth, let me read:
- src/services/auth/login.ts
- src/services/auth/session.ts
- src/services/auth/jwt.ts
- src/services/auth/middleware.ts
- src/services/auth/utils.ts

✅ After interception:

bash
claudemem status  # Check if indexed
claudemem search "authentication login session JWT token validation" -n 15
# Now I have ranked, relevant chunks instead of 5 full files

Example 2: API Integration Audit

❌ Original plan:

code
Audit mentions Prime API files:
- src/services/prime/internal_api/client.ts
- src/services/prime/api.ts
Let me just Read these directly...

✅ After interception:

bash
claudemem search "Prime API integration endpoints HTTP client" -n 20
# This finds ALL Prime-related code, ranked by relevance
# Not just the 2 files mentioned

Example 3: Pattern Discovery

❌ Original plan:

code
Glob("src/**/*.controller.ts")
Then read all 15 controllers to understand routing

✅ After interception:

bash
claudemem search "HTTP controller endpoint route handler" -n 20
# Gets the most relevant routing code, not all controllers

The Psychology of Tool Familiarity Bias

Why You Default to Read/Glob

  1. •Predictability: Read always works, output is deterministic
  2. •No skill overhead: Don't need to invoke a skill first
  3. •Instant gratification: See file contents immediately
  4. •Habit: These are your "native" tools

Why This Is Wrong for Investigation

  1. •No ranking: File #5 might be more relevant than File #1
  2. •No context: You see code but not relationships
  3. •Token waste: Reading 5 files costs ~5000 tokens; claudemem search costs ~500
  4. •Missing code: You only see what you explicitly request

Breaking the Habit

code
BEFORE: "I need to understand X, let me Read files..."
AFTER:  "I need to understand X, let me claudemem search for X concepts..."

Integration with Other Skills

This skill works with:

SkillRelationship
code-search-selectorSelector determines WHAT tool; Interceptor validates BEFORE execution
claudemem-searchInterceptor redirects to claudemem; this skill shows HOW to search
deep-analysisInterceptor prevents bad patterns; deep-analysis uses good patterns
Detective skillsInterceptor prevents duplicate work by trusting detective agents

Quick Reference

code
┌─────────────────────────────────────────────────────────────────┐
│                    INTERCEPTION QUICK CHECK                      │
├─────────────────────────────────────────────────────────────────┤
│                                                                  │
│  BEFORE bulk Read/Glob, ask:                                    │
│                                                                  │
│  1. Is claudemem indexed?     → claudemem status                │
│  2. Can this be ONE query?    → Usually YES                     │
│  3. Am I rationalizing?       → "While agents work" = BAD       │
│  4. Files in prompt?          → Search first, not Read          │
│                                                                  │
│  DEFAULT: Use claudemem search. EXCEPTION: Exact string match.  │
│                                                                  │
└─────────────────────────────────────────────────────────────────┘

Maintained by: tianzecn Plugin: code-analysis v2.2.0 Purpose: Intercept and redirect bulk file operations to semantic search