Skip to content

PRD.md

Latest commit

 

History

History
94 lines (65 loc) · 5.72 KB

PRD.md

File metadata and controls

94 lines (65 loc) · 5.72 KB

PRD: AgentReverse

Vision

A "surgical" integration tool that prevents agent/plugin bloat by reverse-engineering existing tools and extracting only the necessary "DNA" (code, prompts, or logic) for a streamlined, high-performance workflow.

Core Philosophy

Every token counts.

The fundamental goal of AgentReverse is to maximize the utility of every byte in the context window. As users install more skills, plugins, and MCPs, their agent's context becomes cluttered with overlapping, redundant, or underutilized code. This directly degrades agent performance.

AgentReverse combats this by:

  1. Extracting only what's needed — no full repo clones, just the essential logic.
  2. Synthesizing duplicates — merging overlapping capabilities into a single, optimized version.
  3. Centralizing knowledge — one manifest, one source of truth, minimal file sprawl.
  4. Invisible integration — extracted features are documented so the agent just knows when to use them. Users don't need to remember special prompts or tool names.
  5. Workflow gap detection — analyze user's current setup to identify missing behaviors, friction points, and optimization opportunities. When a gap is discovered (e.g., "session resume wasn't automatic"), AgentReverse captures this as a learnable pattern and either extracts a solution from known repos or suggests adding it to the user's config.

The result: a leaner, faster, smarter agent workflow.

The Problem

  • Bloat: Users install whole plugins for a single feature.
  • Context Shrinkage: Large skill folders and complex MCPs consume token budget.
  • Complexity: Integrating parts of different tools requires deep technical knowledge.
  • Portability: Users lack a way to "sync" their custom-built agent capabilities across different environments or IDEs.
  • Fragmentation: Skills, plugins, and MCPs are scattered across many GitHub repos with no standard format.

Core Features (MVP)

  1. Source Analysis: Given a GitHub URL, provide a "Bill of Materials" of its capabilities (skills, tools, prompts).
  2. Feature Extraction: Identify and isolate specific routines, MCP tools, or prompt-based skills within a codebase.
  3. Surgical Implementation: Generate a streamlined implementation and write usage instructions to the appropriate agent config file (claude.md, agent.md).
  4. Agent Requirements Manifest: A requirements.txt-style file (agent-reverse.json) that tracks all installed features, their source repos, versioning, and target agent system.
  5. Portability & Recovery: One-command re-installation (sync) of all integrated features for a new environment.
  6. Conflict Resolution & Synthesis: When a new capability overlaps with an existing one, analyze differences and synthesize an improved, combined version.
  7. Dependency Resolution: Automatically install dependent capabilities if Capability A requires Capability B.
  8. Version Management: Check for updates against source repo HEAD; support rollback/uninstall.
  9. Audit & Slim: Scan an existing setup (local folders or list of repo URLs) for all installed skills/plugins/MCPs, identify duplicates and bloat, and propose a consolidation plan.
  10. Research Interpreter: Ability to read a technical document/article/URL and synthesize a code-based "Skill" or "Tool" based on the abstract concepts described.
  11. Workflow Gap Detection: Analyze user's current workflow to identify friction points and missing behaviors, then either extract solutions from known repos or generate config additions.

Workflow Gap Detection (Detail)

AgentReverse doesn't just extract features from external repos — it learns from user interactions to identify what's missing.

How it works:

  1. Observe friction: When a user manually does something repeatedly (e.g., "check git state on session start"), log it as a pattern.
  2. Match to solutions: Search known repos/skills for existing solutions to the observed friction.
  3. Generate if needed: If no external solution exists, generate the minimal config change (e.g., add to CLAUDE.md).
  4. Track improvements: Record what was added and why, so the manifest becomes a history of workflow optimizations.

Example workflow gaps:

  • Session resume not automatic → Add git state check to CLAUDE.md
  • Repeated web fetches for same docs → Add doc caching behavior
  • Manual worktree creation → Extract/generate worktree skill
  • Forgetting to run tests → Add pre-commit reminder

Output: Each detected gap becomes either:

  • An extracted capability (from external repo)
  • A generated config snippet (for CLAUDE.md/agent.md)
  • A logged "known friction" for future resolution

Target Audience

  • Software Engineers: Wanting to modularize their toolchain.
  • "Vibe Coders": Non-technical users who want to "remix" agent capabilities and take them wherever they go.

Target Agent Systems

AgentReverse is multi-agent aware. The manifest tracks which agent system a capability was installed for, and the sync tool adapts installation steps accordingly.

  • Claude Code
  • Antigravity
  • Cursor (Agent mode)
  • Custom/Other

Target System

  • Architecture: MCP (Model Context Protocol) Server.
  • Primary Language: TypeScript/Node.js.
  • LLM Dependency: None. The host agent (e.g., Claude Code) performs analysis using AgentReverse's tools.

Success Metrics

  • Reduction in total files in skills/ or mcp/ directories while maintaining the same utility.
  • Successful extraction and integration of a "Tool" from a repo into the user's local setup in under 60 seconds.
  • Successful "Restore" of a workflow onto a clean environment using only the manifest file.
  • Successful synthesis of two overlapping capabilities into a single, improved version.