Claude Code Plugins

Community-maintained marketplace

Feedback

Use this skill for multi-step development tasks. Latent Chain Mode reduces token usage by 70-80% through hidden-state reasoning with 4-phase workflow: analysis → plan → impl → review.

Install Skill

1Download skill
2Enable skills in Claude

Open claude.ai/settings/capabilities and find the "Skills" section

3Upload to Claude

Click "Upload skill" and select the downloaded ZIP file

Note: Please verify skill by going through its instructions before using it.

SKILL.md

name ccg-latent
description Use this skill for multi-step development tasks. Latent Chain Mode reduces token usage by 70-80% through hidden-state reasoning with 4-phase workflow: analysis → plan → impl → review.
allowed-tools mcp__code-guardian__latent_*, mcp__code-guardian__guard_validate, mcp__code-guardian__testing_run

CCG Latent Chain Mode

Token-efficient hidden-state reasoning for complex development tasks.

When to Use

MANDATORY for tasks with 2+ steps:

  • Bug fixes requiring code analysis
  • Feature implementations
  • Refactoring across files
  • Code reviews and audits

3 Quick Flows

Flow Command Use Case
A /latent-fix Quick fix 1-2 patches
B /latent-feature Multi-file feature/refactor
C /latent-review Code review without edits

4-Phase Workflow

ANALYSIS ──► PLAN ──► IMPL ──► REVIEW
   🔍          📋        🔧        ✅

Phase 1: Analysis 🔍

  • Understand requirements
  • Identify hot spots (file:line)
  • Document constraints and risks

Phase 2: Plan 📋

  • Break into subtasks/patches
  • Order by dependencies
  • Estimate complexity

Phase 3: Implementation 🔧

  • Apply patches with latent_apply_patch
  • Run guard_validate after each patch
  • Run testing_run_affected for verification

Phase 4: Review ✅

  • Verify all changes
  • Check against constraints
  • Complete task

Core Tools

latent_context_create    - Start new task context
latent_context_get       - Retrieve current context
latent_context_update    - Merge delta (changes only!)
latent_phase_transition  - Move to next phase
latent_apply_patch       - Apply code changes (unified diff)
latent_complete_task     - Mark task complete
latent_status            - Get module status

Output Format

Human-readable format in editor:

🔍 [analysis] <title>
<1-2 sentence description>

[Hot Spots] file:line, file:line
[Decisions] D001: ..., D002: ...
[Risks] if any

---

📋 [plan] <N patches/tasks>

[Patches]
1. file:line - description
2. file:line - description

---

🔧 [impl] Patch N/M: <name>
Applied: <count> | Tests: <status>

---

✅ [review] Complete
Files: N | Patches: M | Tests: passed

Context Delta Format

CRITICAL: Only send changes, never full context!

{
  "summary": "Brief description (max 200 chars)",
  "contextDelta": {
    "codeMap": { "hotSpots": ["src/auth.ts:45"] },
    "decisions": [{ "id": "D001", "summary": "Use JWT", "rationale": "Industry standard" }],
    "risks": ["Token expiry handling"]
  },
  "actions": [
    { "type": "edit_file", "target": "src/auth.ts", "description": "Fix token validation" }
  ]
}

Strict Rules

  1. Summary max 200 chars - No essays
  2. Delta only - Never repeat full context
  3. Decision IDs - D001, D002 for tracking
  4. Always complete - Don't leave contexts hanging
  5. Guard + Test - After every patch
  6. Phase icons - 🔍 📋 🔧 ✅

Example: Bug Fix Flow

User: "Fix the login timeout bug"

Claude:
1. latent_context_create({ taskId: "fix-login-timeout" })
2. 🔍 [analysis] Investigate login timeout
   [Hot Spots] src/auth/login.ts:145
   [Decisions] D001: Root cause is missing token refresh
3. latent_phase_transition({ toPhase: "plan" })
4. 📋 [plan] 2 patches
   1. src/auth/login.ts:145 - Add token refresh logic
   2. src/auth/login.ts:160 - Handle refresh errors
5. latent_phase_transition({ toPhase: "impl" })
6. latent_apply_patch({ target: "src/auth/login.ts", patch: "..." })
7. guard_validate({ code: "...", filename: "login.ts" })
8. latent_phase_transition({ toPhase: "review" })
9. ✅ [review] Complete - Files: 1 | Patches: 2 | Tests: passed
10. latent_complete_task({ summary: "Fixed login timeout with token refresh" })

Token Savings

Traditional Latent Mode Savings
Full context each turn Delta only 70-80%
Verbose explanations Structured output 50-60%
Repeated code References 40-50%