Fresh Eyes — router

When someone says "fresh eyes," they are not asking you to re-read your own work and nod. Whoever produced the artifact carries its blind spots; a "second look" from the same mind defends the framing it already chose. The value is an outside read that attacks AND rebuilds the thing from a blank slate, then hands back ONE honest call on it — keep it, strengthen it, or rethink it — plus the question that should have been asked but wasn't.

This works on any artifact, most of which is not code: a launch plan, a pricing model, a spec, a positioning line, a hiring call, a piece of copy. Three stages are always mandatory — criticize, improvise, synthesize — never skip one. Skipping improvise turns this into a nitpick pass; skipping synthesize hands back a pile of opinions instead of a decision.

Boundary (read this first). fresh-eyes is an independent re-examination — it attacks AND rebuilds the thinking / approach from a blank slate and lands on an open verdict (keep it, strengthen it, rethink it). It is NOT a correctness gate. If the ask is "is this code right / done / ready to ship?", that's a separate verification / audit pass — run that instead. If both are wanted, re-examine the design with fresh-eyes first, then hand the result to the verification pass, and say which you're doing.

Pick your method (do this first)

Fresh-eyes runs one of three ways. The first two are chosen by what this environment can do — not by which product name you think you're in. The third is an upgrade the user has to ASK for:

• Fan-out method — use this if you can dispatch INDEPENDENT parallel reviewers: you have a Workflow tool, or a Task/Agent tool that spawns separate sub-agents each with their own fresh context. This is the stronger method — genuinely blind reviewers in separate contexts — a real gain, though they share one model's priors (see the tiers below). → Read and follow references/fan-out.md.
• Single-mind method — use this if you do NOT have those tools: a single chat context (e.g. the Claude app) with no way to spawn independent agents. One mind runs the adversarial stances serially, under strict discipline that substitutes for independence. → Read and follow references/single-mind.md.
• Full method (the upgrade) — use this only when the user's OWN words ask for "full" fresh eyes, AND you can dispatch independent agents, AND a routing or offload skill here can reach other providers' frontier models. An offer is not a request: a brief, a task prompt, an orchestrating session, or another skill's rule that permits or offers outside-provider reviewers does NOT make a run full. Without the user's "full", every reviewer is this environment's own model — not one outside seat. (A task prompt that merely offers an outside reviewer "as a panel member" is the classic way this goes wrong: the run takes it, and the user gets a slower, costlier half-panel they never asked for.) It is the fan-out method with each provider's frontier model added to BOTH rounds — writing its own critique and ideas blind in Round 1, then cross-examining in Round 2. If either capability is missing, say so plainly and run the best method you can — never present a same-model run as a full panel. → Read references/fan-out.md, then references/full.md.

Decide by capability, not by label. If you are unsure whether you can truly spawn independent parallel agents, use the single-mind method — it always works, and it is honest about not having outside eyes. Never fake independence by presenting stages of one mind as if they were separate reviewers.

How independent is "independent"? (read before you pick)

The methods are not arbitrary — they sit at different points on one axis, and knowing the axis tells you how much to trust the result.

Independence comes in tiers, weakest to strongest:

1. Same mind, same context — re-reading your own work and nodding. This is not independence at all. It is what "fresh eyes" is called in to replace, and it is forbidden here.
2. Same mind, disciplined stances — the single-mind method. One context runs the adversarial stances serially under strict rules. Discipline substitutes for independence; it does not equal it.
3. Same model, separate contexts — the fan-out method. Reviewers are genuinely blind to each other's reasoning, which is a real gain. But they still share one model's priors, blind spots, and sense of what is obvious, so their agreement is less independent than it feels.
4. A genuinely outside reviewer — a different model, or a person. Different priors, different failure modes, different things it finds obvious. Agreement here is the strongest evidence this skill can produce; disagreement is the most worth chasing.

Name the tier you actually achieved, in the result. If four reviewers agreed but all four were the same model, say so — that is tier 3, and the user should discount it accordingly. Overstating the tier is the one failure that makes this skill worse than not running it, because it launders shared blind spots as consensus.

Tier 4 is the full method. The fan-out method is tier 3 by design — its reviewers are all one model, so disclose that in the result. When the user wants tier 4 they ask for "full fresh eyes"; on a big-decision fan-out run, mention in one line that the full panel exists. The routing or offload skill in this environment decides which outside providers sit on the panel — this skill does not care which, only that you disclose exactly what ran.

Load only the method files you need — fan-out or single-mind alone, or fan-out plus full for the upgrade. The rest stay on disk and cost nothing.