autoforge

What This Skill Does

AutoForge is a production-grade autonomous optimization framework designed to move AI agents beyond subjective, unreliable self-reflection. Instead of "vibing" on whether a prompt or piece of code is good, AutoForge implements a mathematically rigorous convergence loop. It enforces a standard where every iteration is measured, logged in a TSV file, and validated against objective criteria. By supporting four distinct modes—prompt, code, audit, and project—the skill allows agents to methodically refine everything from technical documentation and prompt templates to complex CLI tools and multi-file repositories. AutoForge provides the architecture to track improvement over time, using cross-validation to ensure that the code or text being produced actually meets the intended performance targets rather than hallucinating quality.

Installation

To integrate AutoForge into your OpenClaw environment, execute the following command in your terminal:

clawhub install openclaw/skills/skills/akrimm702/autoforge

Ensure that you have appropriate write permissions in your current working directory, as the framework generates audit logs, TSV state files, and reporting artifacts during the optimization loop.

Use Cases

AutoForge excels in scenarios where reliability and consistency are paramount:

• Prompt Optimization: Use the 'prompt' mode to iteratively refine system prompts or SKILL.md templates by simulating scenarios and evaluating pass/fail rates.
• Code Robustness: Utilize 'code' mode to run sandboxed test executions, measuring stdout/stderr against defined criteria to ensure script reliability.
• Documentation Integrity: Employ 'audit' mode to verify that CLI behavior matches your documentation, preventing documentation-code drift.
• Repository Maintenance: Leverage 'project' mode for cross-file consistency analysis, ensuring that updates to one part of a repository do not break dependencies in another.

Example Prompts

1. "AutoForge, please optimize the Python script in my current directory; use code mode and ensure it passes all existing unit tests."
2. "Run autoforge on this repository in project mode to check for any inconsistencies between the README and the actual CLI command outputs."
3. "I need to improve my agent's persona prompt. Please use autoforge in prompt mode to simulate 10 user interactions and optimize the instructions until the response pass-rate exceeds 90%."

Tips & Limitations

• Multi-Model Advantage: For high-stakes tasks, consider splitting the Optimizer and Validator roles across different models (e.g., Opus to optimize, GPT-4/Gemini to validate). This removes the bias inherent in a single model grading its own output.
• Convergence over Speed: Do not expect instantaneous results. AutoForge is designed for iterative, systematic improvement. The time taken to reach convergence is proportional to the complexity of the evaluation metrics.
• Environment Variables: For automated reporting, configure AF_CHANNEL, AF_CHAT_ID, and AF_TOPIC_ID. If these are not provided, the framework will default to standard output, which is suitable for local debugging but less ideal for long-running autonomous processes.