high-agency

High Agency is a small skill for AI coding agents that adds useful pushback and follow-through. It helps an agent protect your goal when the method you proposed is likely to work against it, then move the task forward with evidence and verification.

What it does

• Preserves the outcome you want while treating the proposed method as open to review.
• Explains a concern through its likely consequence, supporting evidence, and a better path.
• Continues with safe, reversible work instead of repeatedly asking for approval.
• Verifies results and reports real uncertainty or blockers plainly.

The skill does not grant extra permissions, replace your decisions, or guarantee a correct outcome. It works best when the agent can inspect the relevant code, data, or documentation.

Install

Clone this repository, then copy the skill folder into your project.

git clone https://github.com/Cjbuilds/high-agency.git

Replace /path/to/high-agency below with the clone location. Review an existing installed copy before replacing it.

For Codex:

mkdir -p .agents/skills
cp -R /path/to/high-agency/skills/high-agency .agents/skills/

For Claude Code:

mkdir -p .claude/skills
cp -R /path/to/high-agency/skills/high-agency .claude/skills/

Use

Invoke it by name in your request:

Use $high-agency to diagnose why this endpoint is slow and fix the root cause.

The agent will load SKILL.md, inspect the task, challenge a risky method when the evidence warrants it, and pursue the smallest authorized path to a verified result.

Check the package

The checker uses only the Python standard library:

python3 scripts/check.py

It validates the skill metadata, nonempty instructions, installation instructions, and repository links.

Files

high-agency/
├── assets/
│   └── banner.png
├── eval/
│   └── RESULTS.md
├── scripts/
│   └── check.py
├── skills/
│   └── high-agency/
│       └── SKILL.md
├── LICENSE
└── README.md

Behavioral evaluation

Tested on three matched tasks with Fable 5.1 and GPT-6 Astra. Baseline and skill conditions both passed the final checks; this small test does not prove an improvement. Read the method, outputs and limits.

python3 eval/check_cases.py

Limits

This skill cannot supply missing credentials, bypass required approvals, or make irreversible actions safe. It encourages evidence-based judgment, but the quality of the result still depends on the available evidence and tools.