Skill Ablation — Prove the Scaffolding Still Earns Its Keep

memory-gardener prunes bloat inside instructions that stay. This skill tests whether an instruction should exist at all — a periodic ablation pass, not continuous pruning. A skill under the line cap can still be 100% obsolete; only removing it and watching for a real stumble proves that.

Self-improvement (do this first and last)

1. At start: read learnings.md in this skill's folder if it exists. Apply relevant lessons.
2. At end of every use: append one dated bullet — which instruction turned out obsolete, which was proven load-bearing, any restore that took more than one stumble to justify. Merge instead of duplicating; delete disproven bullets.

When to run

• The project's primary coding-agent model changed (new major version, new vendor default).
• ~6 months since the last ablation run, or never run.
• The user invokes it by name, or a memory-gardener pass finds the same instruction repeatedly untouched. Never run mid-task — this needs a clean session boundary, not a mid-edit interruption.

Procedure

1. Archive, don't delete

Move the project's session-protocol file (CLAUDE.md/AGENTS.md), skill root (.claude/skills/ or .opencode/skills/), and any hooks config into a dated archive folder (e.g. .ablation-archive/<date>/), not the trash. State the archive path before continuing — this is the undo path if the bare run goes badly.

2. Run bare, for real work only

Start the next session with none of it loaded. Do not pre-guess which instruction the model will need back — that defeats the test. Work the actual backlog (real tasks, not synthetic ones) for a defined window: a handful of real sessions, or ~1 week of normal use, whichever the user prefers.

3. Log every stumble, don't fix it inline

Each time the model does something a deleted instruction would have prevented — wrong convention, skipped step, repeated question — log it in a scratch note: what happened, which archived file would have caught it. Do not restore yet. One stumble is noise.

4. Restore only on repeated evidence

After the window, scan the stumble log. Restore an instruction only if the same stumble appears 3+ times across independent sessions/tasks. Everything else in the archive stays archived — it's confirmed dead weight, not "might need it later." State the count when restoring: "X stumbled on Y in 3 of 5 sessions → restoring."

5. Rewrite restored instructions lean

A restored instruction is a chance to fix it, not just copy it back verbatim — cut anything the stumble log shows the model didn't actually need, per the skill-writer quality bar (checkable rules, hard caps, no restated prose).

6. Report and archive the log

Summarize: kept archived (count), restored (count + evidence), any skill rewritten during restore. Fold the stumble log into this skill's learnings.md as dated bullets, then delete the scratch copy — the log itself is not worth keeping once distilled.

Relationship to other skills

• memory-gardener prunes and merges what stays; this skill decides what stays at all. Run gardener first if a learnings/handover file is over its cap — no point ablating bloat that's about to be pruned anyway.
• skill-writer's quality bar governs anything rewritten during restore (step 5).
• Tool-agnostic by design: "archive the skills folder / CLAUDE.md / hooks config" works the same in Claude Code, OpenCode, or any markdown-driven agent — no tool-specific ablation switches required.

Hard rules

• Never restore on a single stumble — the 3+ threshold is not a suggestion.
• Never skip the archive step — deleting outright removes the undo path and the evidence trail.
• Never run the bare window on synthetic/toy tasks — it must be real backlog work or the result doesn't generalize.