Superior

• The pass runs at the proper scope of § Scope Comes First and states it in one line with its reason; an ask too small to earn anything is widened by a scope interview or by stating the inferred scope, never run on the small scope as given. A scope the caller fixed runs as given, with no interview.
• The superior approach comes from the mind defined in § The Superior Mind, written before our current plan is reread (or by a blind agent when anchoring is likely), and is concrete enough to execute.
• Every difference between that approach and ours ends as an executable change or a named real limit, and no constraint the principal actually set is dropped.

The lens is defined by what that mind has and lacks, never by adjectives like "more careful" or "more thorough":

• It knows the field. The best existing methods, the frontier work, the tools and prior art we have not heard of, how experts in adjacent fields solve the same shape of problem.
• It sees the goal, not the ticket. It may reframe the task, solve a bigger version that makes this one trivial, or show that the task should not exist.
• It invents no limits. Time, cost, tooling and "how it's done here" are questioned unless they are physics or a constraint the principal actually set. Real constraints it respects completely.
• It reaches for leverage. It looks for the one move that makes most of the work unnecessary before doing the work.
• It is not anchored. It never saw our first idea, so it is not improving our plan; it is producing its own.

The pass begins from the scope it was given (the named file, feature, system or question) and works out what the asker most likely means by it. The literal ask is evidence, not the boundary: "how is the login implemented" usually means the auth flow end to end, not the one handler that was opened, and almost never means the whole product.

Abstract up to the proper scope, and no further. The proper scope is the smallest unit that contains the real goal: the level where the decision the asker cares about actually lives, reasoned from what was named, the surrounding conversation, the project, and what the answer will be used for. "The whole system" is the scope only when the ask points there. The result is stated before mapping, in one line: Scope: <inferred scope>, from <what was asked>, because <reason>.

When the inference is not confident, Elevate's scope interview settles it. Its questions find the right level: what this belongs to, what decision the answer feeds, and how far to reach (this piece, the feature, the subsystem). A scope the caller already fixed is the scope.

User customizations, when present, live at <configRoot>/LIFEOS/USER/CUSTOMIZATIONS/SKILLS/Superior/ (configRoot is $CLAUDE_CONFIG_DIR when set, else ~/.claude). Any PREFERENCES.md, configurations, or resources there override the defaults; when the directory does not exist, proceed with skill defaults.

• Scope is inferred, never defaulted (principal, 2026-10-09). There are two failures. Run on the literal file or function named, and the superior approach matches ours with an empty gap table. Jump to "the whole system" by default, and the pass maps something nobody asked about. The first version of this rule made the second mistake. Reason from the ask to the level where the real goal lives, and say why in the scope line.
• Anchoring is the main failure (design-time, 2026-10-07). Run inline after a plan already exists, the "superior" answer comes back as that plan with adjectives. Full mode exists for this: the agent sees the goal and the real constraints, never our draft.
• Full mode audits the facts in the brief, not just the plan (2026-10-07, first real run). Asked how to reach a target figure, the blind agent first checked the baseline it was handed and found it overstated, because the tool reporting it had been built that same day. Hand it the sources behind every figure, not only the figures, so it can do that.

Example 1: Hard design work

User: "Design the sync layer for the notes app. Run Superior on it."
→ Elevate, full mode: scope map of the app, server API and clients first; a fresh agent gets the map, goal and constraints, never our draft design
→ It returns a CRDT-based local-first design; we had planned server-authoritative REST with conflict prompts
→ Gap table: knowledge (CRDT libraries are mature), convention (server-as-truth), invented constraint ("offline is an edge case")
→ Changed plan adopts local-first; the one real limit (existing server API) is kept as a migration path

Example 2: Ask too small to run

User: "Superior on this retry helper."
→ The helper alone is too small, and it has callers in two places, so two scopes are plausible
→ Elevate asks: "I'd read this as the job runner's retry path; is that right, or is it the HTTP client?"
→ Principal: "Job runner; jobs keep double-firing."
→ Scope: the job runner's retry and dispatch path, not the whole runner; the map shows retries in three layers; the superior approach is idempotent jobs, which makes the helper unnecessary

Example 3: Inline lens during a max-effort run

A depth turn ("think deeply") on a strategy question; Superior runs beside FirstPrinciples
→ Elevate, light mode: one inline pass, stated before the current plan is reread
→ Superior approach dissolves half the task by reframing who the audience is
→ Two gaps adopted, one blocked by a principal-set budget, recorded in the ISA Decisions