Consilium — Development skill for Claude Code

consilium

**English** · [Українська](README.uk.md) · [Русский](README.ru.md)

Let one AI ask a different AI when it isn't sure.

What this is, in one breath

There's a kind of AI that lives in your terminal and writes code *with* you — it reads your files, edits them, runs commands. Claude Code, Codex, Gemini's `agy`, and OpenCode are all this kind of tool. People call them coding assistants, or just "agents."

They're good. But like anyone working alone, an agent can't always tell when it's wrong. It will talk itself into a risky change, miss a bug it wrote two minutes ago, and keep retrying the same broken fix long after it should have stopped.

**consilium adds one command, `consult`, that lets your agent phone a _different_ AI for a second opinion.** Your agent hands the same code or question to a model from another company and asks: *what do you think?* The other AI reads it and answers in plain text. It advises; it doesn't act for you.

consult codex -- "Is it safe to run this migration twice?"

You — or your agent — type that. A different AI reads the question, looks at the code in your current folder, and prints back what it thinks. You decide what to do with the answer.

Still early, but I reach for it most days.

Why ask a *different* AI, and not just ask the same one again

Think of it like checking your own homework. Re-read your own answer and you see what you *meant* to write, not what's actually on the page. Hand it to someone from a different class — who studied from different notes and has no reason to agree with you — and they spot the mistake in five seconds.

An AI has the same blind spot. Ask the same model to "double-check," and it mostly nods along with itself. Ask a *rival* model from another company — trained differently, with no stake in the first answer — and it actually pushes back. That pushback is the whole point. It's where the useful catches come from.

So the rule is short: **ask someone other than the AI you're