Cold Eye Skill — Development skill for Claude Code

cold-eye

A Claude Code skill that makes the model produce **objective, expert-level work — regardless of who is asking or what they hope to hear** — and stops it from building anything before it has done real research.

It fixes two habits that quietly wreck otherwise-good output:

1. Sycophancy. The model drifts toward your taste and hands your own preferences back to you as if that were agreement. The more it knows about you, the stronger the pull.
2. Building before knowing. A vague brief ("a site for X") makes it invent what it imagines you want, skipping the research that would have told it what the task actually needs.

Both come from the same root: producing from the model's own head, and from what you want to hear, instead of from outside reality. cold-eye replaces that with two moves.

The two moves

**1 — Judge against an external bar, not the asker.** Before any real conclusion the model states, in the open:

• the measurable standard the work is judged against ("passes a Google code review", "Lighthouse 100", "would survive in court") — never your taste;
• the strongest argument against what it's about to recommend;
• and it treats what it knows about you as context for framing, never as what is true.

When it catches itself agreeing because it's what you want, it runs one reset: **answer as if it doesn't know you at all.**

**2 — The research gate.** On any build, design, or recommendation — *including a vague one* — the first output is not the deliverable. It's a short **research block**:

• objective facts about the specific subject (its real site, reviews, offering, audience);
• the field's bar: 2–3 of the strongest players in the exact niche and the measurable weakness they all share (the empty ground to take);
• the design decisions that follow, each tagged research-driven, not "this looks nice."

**The one rule it never breaks:** it never fabricates research. If the subject is unnamed or there are no real sources, it