Source of truth for how Claude works with Samuel Aure across every project, and the record of how those rules evolved.
Everything under ~/.claude/ is a copy. Edit here, commit, then copy out. Never edit the live
location directly — the change would be unrecorded and lost on the next sync.
global/CLAUDE.md → ~/.claude/CLAUDE.md always loaded, every session
skills/<name>/SKILL.md → ~/.claude/skills/<name>/SKILL.md loaded on demand
templates/ scaffolding copied into projects
archive/ superseded protocols, kept for history
| Layer | Lives in | Cost | For |
|---|---|---|---|
| Global | ~/.claude/CLAUDE.md |
Every session | Who Samuel is, how to work with him, incident memory, repo index |
| Per-repo | <repo>/CLAUDE.md |
That repo only | Architecture, naming, traps specific to one codebase |
| Skill | ~/.claude/skills/<name>/ |
Zero until invoked | Deep operational knowledge for one area |
| Project memory | ~/.claude/projects/<key>/memory/ |
That project only | Facts learned during sessions |
Project memory is keyed by working directory, so a memory written in nau does not load in
another repo. Anything that must cross projects belongs in the global layer or a skill.
cp global/CLAUDE.md ~/.claude/CLAUDE.md
cp -r skills/. ~/.claude/skills/Verify with diff -q afterwards. Both directions should be identical; if they differ, the live copy
was edited directly and the difference needs reconciling into this repo first.
A rule earns its place only if it encodes something non-obvious — a real incident, a domain constraint, a decision that had valid alternatives. Rules that restate general engineering practice consume context in every session and replace judgment with compliance, which makes the output worse, not safer.
When a file grows, cut rather than append.
archive/GEMINI.md and archive/workflows/ are the previous generation: a 450-line constitution
plus 15 workflow actors, written for Gemini. They were never wired into Claude Code, which is why
the rules had to be re-explained by hand each session.
They are kept because parts of them are genuinely good — the security rules are memory of real incidents — but they are not authoritative. Known problems with them at the time of archiving:
- Rule S5 mandated
NAU_SERVICE_KEYfor inter-service auth. The platform migrated to signed service JWTs;nau/CLAUDE.mdsaid the opposite. Two constitutions disagreeing is worse than one missing, because which one applies is arbitrary. - The server profile described a 4 GB / 40 GB CX23 that had already been upgraded to 7.6 GB / 75 GB, so every derived resource limit was wrong.
- The "Anti-Data-Loss" section covered Docker pruning and said nothing about database backups — while the nightly backup had been failing silently for weeks.
- The Intercept Gate routed every request through a fixed six-branch tree, which substitutes procedure for judgment on exactly the decisions where judgment is worth most.
The workflow actors still need review: some are naŭ-specific, some general, and the step sequence itself has not been evaluated against how the work actually goes.