ad-level-up

name	ad-level-up
description	Run this skill when the user explicitly invokes `/ad-level-up` or names it, or asks to evolve the project's rule-set — "add a convention", "update the rules", "new rule", "merge these rules", "we keep hitting X, make it a rule", "retire this rule" — or hands over a rule gap surfaced by `ad-audit`. Companion to `ad-audit`: where ad-audit audits against the rules, this evolves them, leanly. This skill is user-invoked (`allow_implicit_invocation: false`) because it can write to the rule-set — invoke it explicitly with `/ad-level-up` or by naming it, not from inferred conversation. Mechanical shape: ONE pass in the current session. Every candidate clears four anti-overfitting gates plus an effectiveness pass or is rejected out loud; the drafted candidate then passes an adversarial multi-lens review; and NOTHING is written without the user's explicit approval, one item at a time. It presents a proposal with a plain-language rationale and applies only on your OK. Writes land in the curated rule-set — the ADR-0035 machine store for you-everywhere conventions, or the ADR-0043 project layer (`.agentic/rules/`) for this-project conventions.
summary	Human-gated rule-set curation, companion to ad-audit. Four anti-overfitting gates + effectiveness pass + adversarial multi-lens review per candidate, then a HARD human gate — never writes unprompted. Targets the ADR-0035 machine store or ADR-0043 project layer. Ships a rule-candidate-reviewer subagent.

Codex skills run inline in the current session. This skill can write (via edits), so its safety rests entirely on the human gate below, not on tool restriction. The default pass runs the gates, the effectiveness pass, and the adversarial multi-lens review inline; the optional escalation runs the lens reviews as user-spawned subagents for true isolation.

Mechanical shape:

THIS SESSION:
  1. State the candidate + its citation (or stop — not grounded).
  2. Trace it to a root cause.
  3. Run the four anti-overfitting gates (all must pass) + the effectiveness pass.
  4. Place it deterministically; draft the minimal edit (do NOT write yet).
  5. Adversarial multi-lens review of the candidate (already-covered / coherence / placement).
  6. PRESENT the proposal + rationale. Apply ONLY on explicit approval, one item at a time.

HARD human gate: NEVER write to the rule-set without explicit human approval. Always present the proposal plus a plain-language rationale first; apply only on the user's OK, one item at a time. No batch writes, no "while I'm here" edits. This gate is enforced by this contract — hold to it even though the session can write. - Do NOT write, edit, or create any rule file before the user explicitly approves that specific item. - Do NOT add a rule that an existing rule (or a repo binding doc) already covers — merge or sharpen instead. - Do NOT add a rule for a lone, uncited, cheaply-caught slip — that is a note, not a rule (each extra rule lowers adherence to all the others). - Do NOT invent a rule-set location — machine store resolves `$AGENTIC_RULES_DIR` / `~/.agentic/rules/`; project layer is `.agentic/rules/` at the repo root; if the target layer is absent, offer to create it on approval. - Do NOT put the project layer in `.gitignore` when the user picks machine-local visibility — the exclusion goes in `.git/info/exclude` (per-clone, never committed). - Do NOT route an architectural decision or an engineering standard here — send it to `/ad-adr` or `/ad-guidelines`. - Do NOT begin any file I/O before printing the Step 0 announce line. Step 0 — announce. Print the shape before any work: - Re-run `/ad-audit` after a rule lands to confirm it catches what it was written for. - For a decision bigger than a rule line: `/ad-adr` or `/ad-guidelines`. - `ad-audit`'s Step 8 handoff is this skill's Step 1 candidate.