unhobble

Purpose

As models improve, instruction surfaces written for older models become the ceiling: the model reads every standing line every session, and lines that correct mistakes it no longer makes cost context and constrain behavior. Official doctrine says cut any line whose removal would not cause mistakes (best-practices); the strongest form of that test is empirical: delete, run, watch. This skill operationalizes the experiment its sibling audit-instructions can only reason about: instead of judging instruction text against doctrine, it measures the model against the repo with the instructions gone, and lets observed stumbles, not guesses, decide what returns.

Rebuild rule (the whole contract in one line): an instruction returns only after the bare model repeatedly stumbles on the same thing, and the re-added line cites the evidence.

When to run

• A frontier model generation ships (the canonical trigger, since instructions written for the previous generation are now suspect).
• The repo's instruction surface has grown past the point anyone can say which lines still earn their cost.
• On a cadence the operator chooses (see Cadence wiring below). The talk-circuit heuristic is "every six months", but the model release is the real event.

Scope and safety rails

• Project scope by default. The experiment strips the project's surfaces: project CLAUDE.md / CLAUDE.local.md, .claude/rules/, .claude/skills/, .claude/agents/, project-settings hooks, and project-enabled plugins. User-global surfaces (~/.claude/**) are included only when the operator explicitly opts in per phase-1 prompt, never by default.
• Managed settings are never touched. Org-managed policy is not the operator's to ablate.
• Reversible by construction. Tracked-file changes happen on a dedicated experiment branch; untracked/settings changes are backed up to plugin state before modification and restored from that manifest. Nothing is destroyed: git history and the snapshot manifest are the safety net.
• Human-gated. Every mutating step (strip, restore, re-add) presents its exact change set and waits for operator confirmation. Bare invocation of a phase never mutates silently.
• Security posture is out of scope. Hooks that enforce policy (secrets gates, PR-body contracts, permission guards) are classified policy at snapshot time and are NOT stripped by default: the experiment measures model capability, and policy gates are not model-era workarounds. The operator may force-include one explicitly; the manifest records that choice.