antidote

Stop AI agents from patching symptoms. Fix the root cause instead of masking the problem.

When code breaks, AI agents default to band-aids: null checks, swallowed errors, and 30-line workarounds. They hide bugs under layers of guards and sanitizers that rot over time.

antidote forces the agent to fix the actual cause:

• Validate once at the boundary instead of checking everywhere downstream.
• Use the right type or schema so invalid data can't reach internal code.
• Delete the broken path instead of writing glue code around it.

One clean fix replaces a dozen defensive guards.

The skill prompt and its rules live in skills/antidote/SKILL.md. This README covers why you'd install it, the benchmark results, and how to install it.

Benchmark

The skill was tested with the skill-eval-harness across five models. Each test gives the model bugged code and asks it to fix the root cause; a script oracle compiles and runs the output to check that the fix is structural, not a patch. Four cases, 2 runs per arm. Pass rate = share of runs that passed the oracle.

Four of the five models passed zero runs without the skill — they patch symptoms by default. With the skill, they find the structural fix instead. Full methodology and reproduction commands: docs/benchmark.md.

Install

This skill follows the open Agent Skills standard. It works in Claude Code, Codex, Cursor, Devin, Cline, GitHub Copilot, and other popular AI coding platforms.

Universal (recommended)

npx skills add Avtr99/antidote

This command detects supported agents on your machine and installs the skill into each agent's native skills directory.

npx skills add Avtr99/antidote --global          # all projects, not just this one
npx skills add Avtr99/antidote --skill antidote  # install only this skill
npx skills add Avtr99/antidote --list            # preview without installing

Claude Code

claude plugin marketplace add Avtr99/antidote
claude plugin install antidote@avtr-antidote

Codex / ChatGPT

codex plugin marketplace add Avtr99/antidote
codex plugin install antidote@avtr-antidote

Then install from the Plugins directory in ChatGPT Work or the Codex desktop app.

GitHub Copilot CLI

copilot plugin marketplace add Avtr99/antidote
copilot plugin install antidote@avtr-antidote

Cursor

cursor plugin install Avtr99/antidote

Devin

devin plugins install Avtr99/antidote

Verify with devin plugins list. Plugins load at session start, so start a new Devin session after installing.

Grok Build CLI

grok plugin marketplace add Avtr99/antidote
grok plugin install antidote@avtr-antidote

Or install directly: grok plugin install Avtr99/antidote.

Pi

Copy .pi/extensions/antidote.ts into your Pi workspace's .pi/extensions/ directory and restart Pi.

OpenCode

Add antidote to the plugin array in your opencode.json and restart OpenCode. See .opencode/INSTALL.md for full instructions.

{
  "plugin": ["antidote@git+https://github.com/Avtr99/antidote.git"]
}

Manual

Copy skills/antidote/ into the skills directory of your agent:

Usage

The skill activates on its own when the agent writes or reviews code. You can also invoke it by name:

• Slash command (Claude Code, Codex, Cursor, Devin): /antidote
• Natural language: "use antidote", "audit for overcomplication", "find the root cause"

It has two modes — fix (one bug) and audit (a whole codebase) — described in SKILL.md.

What is inside

Compatibility

antidote is a pure-prompt skill. It has no scripts, no network calls, and no dependencies. It works in any agent that reads the SKILL.md format. The adapter manifests and plugins in this repo (.claude-plugin/, .codex-plugin/, .cursor-plugin/, .devin-plugin/, .grok-plugin/, .github/plugin/, .pi/, .opencode/) add native install support on top.

License

MIT — see LICENSE.