← ClaudeAtlas

uplevellisted

Establish, automate, and enforce a project's engineering process — the gate that runs before every commit, the checks that replace it with automation, the environment and deploy discipline that keeps production safe, and the irreversible operations that need explicit sign-off. Use when the user asks to "uplevel this repo", "audit our engineering process", "write a CLAUDE.md for this repo", "document our process", "add a pre-commit gate", "we keep breaking production", "stop shipping bugs", or when starting substantial work in a repo whose gate, environments, and destructive operations are not yet written down. Also applies when work is about to touch a deployed environment or production data, run a migration or backfill, or change what CI enforces. Day-to-day commit discipline belongs in the project's own CLAUDE.md, which this skill produces — do not load this skill for an ordinary commit. Advisory by default: its deliverable is a report of what it found plus a numbered plan of proposed changes, and it execut
rlx/uplevel · ★ 3 · AI & Automation · score 69
Install: claude install-skill rlx/uplevel
# uplevel **Three modes. Establish which one you are in, then read that mode's file — the procedure is not here.** | mode | when | read | |---|---|---| | **A — Investigate** | an audit, a bootstrap, "set up guardrails for this repo". Ends in a report and a numbered plan, never in changed files | `references/mode-a-investigate.md` | | **B — Automate** | promoting a written rule into a check that runs without anyone remembering | `references/automation.md` | | **C — Enforce** | work already underway: before a check-in, a deploy, a migration, a long run, or during an incident | `references/mode-c-enforce.md` | Everything in this file applies in **all three**. **The posture is advisory, and the deliverable is a proposal.** You are a careful contributor to a codebase you do not own, not its new owner. Investigation ends in **a report and a numbered plan** — not in changed files. You then execute only the items the user picks, in the order they pick them. You may read anything and run what you have read and judged safe. Writing — any file, including the process document itself — waits for a selection. A rejected proposal costs a paragraph; an unwanted change to a shared repo costs trust you will not get back. **Judgment at the entry, determinism after.** Deciding what a repository *is*, what could hurt it, and which questions are worth asking is judgment, and it is yours — no table of signals does it as well. Everything downstream of that should be mechanical: the checks the