Skip to content

Consist LLM Skill

Consist includes an LLM-provider-agnostic skill for agents that help users instrument downstream scientific workflows. The skill packages project-specific guidance for bootstrapping Consist, wrapping legacy file-writing steps, debugging cache and provenance behavior, inspecting runs, and querying recorded artifacts.

The canonical skill lives in the repository at skills/consist/.

When to use it

Use the skill when asking an agent to:

  • add Consist to a new project or pipeline;
  • choose between run(...), trace(...), and scenario(...);
  • wrap external tools, containers, or legacy code that writes files;
  • declare outputs, output sets, schema profiles, and cache identity;
  • inspect or analyze recorded runs and artifacts;
  • debug cache misses, hydration, mount resolution, or scenario wiring.

Install

Download or clone the Consist repository and copy the full skills/consist/ folder into your agent tool's skill directory. Preserve the folder layout; SKILL.md depends on the files under references/.

For Codex, the default install shape is:

git clone https://github.com/LBNL-UCB-STI/consist.git
mkdir -p "${CODEX_HOME:-$HOME/.codex}/skills"
cp -R consist/skills/consist "${CODEX_HOME:-$HOME/.codex}/skills/"

For active development from a local Consist checkout, use a symlink instead so skill edits stay in sync:

mkdir -p "${CODEX_HOME:-$HOME/.codex}/skills"
ln -s "$(pwd)/skills/consist" "${CODEX_HOME:-$HOME/.codex}/skills/consist"

Maintain

Keep canonical guidance in skills/consist/SKILL.md and skills/consist/references/. Provider-specific files such as skills/consist/agents/openai.yaml are optional UI metadata.

After editing the skill, validate it with your agent platform's skill validator. With Codex's skill-creator tooling, the command shape is:

python path/to/skill-creator/scripts/quick_validate.py path/to/skills/consist