Agent skill · laurigates
meta-local-notes
Audit a CLAUDE.local.md — verify each claim live, fix drifted environment facts, cut findings duplicating versioned docs. Use when local notes are stale, or before appending a finding.
What it needs
About 6k tokens when loaded.
What this skill does
Machine-Local Notes: Verify, Then Trim A CLAUDE.local.md is the only always-loaded context file with no reviewer. It is git-ignored, so it never appears in a diff, never gets a PR comment, and no CI check ever reads it. Every other doc in the repo is defended by review; this one is defended by nothing but somebody happening to look. So it rots in two directions at once, and both are invisible: 1. Findings creep — session conclusions get written where they are convenient rather than where they belong, duplicating docs/adrs/, docs/roadmap.md, or the issue tracker. 2. Silent staleness — environment facts that were true when written. Nothing re-checks them, and the file reads exactly as authoritative on the day it goes wrong as on the day it was right. A stale claim here is worse than no claim: it is loaded on every turn, it is specific, and it is trusted. The failure mode is a session confidently acting on a fact that stopped being true weeks ago. When to Use This Skill Use this skill when... Use a different skill when... --- --- A machine-local notes file has gone stale, bloated, or contradicts the ADRs/roadmap You want to move always-loaded content to an on-demand skill — use meta-context-diet (load cost, not truth) You are about to append a session finding to a CLAUDE.local.md and need the keep test You want to turn a session's learnings into new rules — use session-distill You need environment claims verified against the live host before trusting them You want to move rules/skills between scopes — use meta-promote The keep test One question per claim: does this describe the machine, or the project? Keep — environment facts that change how you develop or test here Cut — project findings, however they got here --- --- Host and access: ssh target, hardware, driver/toolkit versions Measurements, benchmark numbers, VRAM/throughput tables Where things live: checkout, caches, datasets, staged model files Root-cause narratives and experiment logs How to run this project's …
How to use it
Reference it in AdaL, Claude Code, Cursor or any coding agent — nothing to install:
@skills laurigates/meta-local-notes