Skip to content

test_case.md

Latest commit

 

History

History
95 lines (64 loc) · 4.15 KB

test_case.md

File metadata and controls

95 lines (64 loc) · 4.15 KB

Test Cases & Usage Examples: AgentReverse

This document outlines practical scenarios where AgentReverse can be used to streamline workflows, reduce bloat, and enhance agent capabilities.

Scenario 1: Surgical Feature Extraction (GitHub)

User Intent: "I like the way this repo handles context window management, but I don't want the rest of the bloat." Input: https://github.com/example/complex-agent-framework Agent Command: /agent-reverse analyze https://github.com/example/complex-agent-framework

Interaction:

  1. AgentReverse clones the repo (shallow) and parses it.
  2. It returns a "Bill of Materials" (BOM):
    • Skill: context-winnowing (proactive token reduction)
    • Tool: git-history-search
    • Prompt: structured-task-breakdown
  3. User: "Extract context-winnowing and install it for my Claude Code setup."
  4. AgentReverse: Extracts logic, converts it to a .claude/skills/context-winnowing.md file, adds a usage hint to CLAUDE.md, and updates agent-reverse.json.

Scenario 2: Proactive Capability Setup

User Intent: "I found a cool repo, set it up for me." Input: https://github.com/vibe-coder/cool-plugins Agent Command: Use AgentReverse to see what's useful in this repo and set it up: https://github.com/vibe-coder/cool-plugins

Interaction:

  1. AgentReverse identifies a set of MCP tools and specialized prompts.
  2. It asks: "This repo has a really good 'Image-to-React' component extractor. Should I install it as an MCP tool and add the workflow rules to your agent.md?"
  3. User: "Yes."
  4. AgentReverse configures the MCP server, adds the rule, and verifies the setup.

Scenario 3: Legacy Setup Audit & Optimization

User Intent: "My environment is slow. I have too many skills and I don't know what's overlapping." Agent Command: /agent-reverse audit

Interaction:

  1. AgentReverse scans .claude/skills/, skills/, and all connected MCP servers.
  2. Analysis Report:
    • RED: 3 overlapping file-search tools found (Skill A, Skill B, MCP C).
    • YELLOW: web-search-old is outdated (Source HEAD is 10 commits ahead).
    • GREEN: 5 unique capabilities.
  3. Optimization Plan: "I can merge the 3 file-search tools into a single 'Ultimate Search' tool that uses the best logic from each. Shall I proceed?"

Scenario 4: Workflow Hardening (claude.md / agent.md)

User Intent: "My current search workflow is buggy. Optimize it." Agent Command: AgentReverse, analyze my search-related skills and plugins, then write better rules in my CLAUDE.md.

Interaction:

  1. AgentReverse looks at how current search features are implemented.
  2. It realizes the user often forgets to specify a directory.
  3. It updates CLAUDE.md with a new "Invisible Integration" rule:

    "When searching, if no path is provided, default to current workspace root and use the google-search-lite capability to verify external context first."

Scenario 5: Research-to-Code Implementation (Deep Feature)

User Intent: "Look at this blog post on Reasoning with Language Models (RLM) and implement the core idea for me." Input: https://alexzhang13.github.io/blog/2025/rlm/ Agent Command: /agent-reverse interpret https://alexzhang13.github.io/blog/2025/rlm/ and implement as a skill.

Interaction:

  1. AgentReverse reads the article text.
  2. It identifies the "Core Idea": Recursive refinement of internal search queries before external retrieval.
  3. Synthesis: It creates a new skill rlm-search.md that implements this recursive logic using the user's existing search tools.
  4. Portability: It adds this "Custom Synthesis" to agent-reverse.json so it can be recreated elsewhere.

Scenario 6: Zero-Friction Migration (Sync)

User Intent: "I just switched to a new laptop. Get my agent back to how it was." Agent Command: /agent-reverse sync

Interaction:

  1. AgentReverse reads agent-reverse.json.
  2. It sees 12 pinned capabilities from various repos.
  3. It fetches each repo at the pinned commit, extracts the features, and re-implants them into the new environment.
  4. Result: Identical workflow restored in minutes without manual cloning.