Skip to content

racmac57/dotclaude

Repository files navigation

dotclaude

Personal Claude Code dotfiles for Windows: global skills under skills/, shared settings.json, and per-machine CLAUDE.md. Documented skill index and how-tos live in docs/skills/ (SKILLS_INDEX.md + how_to/). (The ai_enhancement repo is a downstream mirror, regenerated post-OneDrive-recovery.)

Skill staging (skills/_inbox/): skills generated by /reflect land in skills/_inbox/<slug>/SKILL.md with a status: draft-inbox frontmatter marker — a review-staging area, not active. Promotion to an active skills/<slug>/ directory is a manual, QA-gated step (run qa-skill-hardening, then drop the marker). Only promoted skills appear in docs/skills/SKILLS_INDEX.md.

Not in this repo: machine-local caches, session history, plugins marketplace trees, and optional per-folder tooling — see .gitignore.

Multi-Machine CLAUDE.md

The repo contains three variants of the global Claude instructions:

File Purpose
CLAUDE.desktop.md Desktop (carucci_r is the real profile, no junction)
CLAUDE.laptop.md Laptop (RobertCarucci is real; carucci_r is a junction)
CLAUDE.shared.md Reference copy of the identical shared content
CLAUDE.md Current deployed version (kept in sync with the above)

To deploy on any machine, run from the repo root:

.\Deploy-ClaudeConfig.ps1

The script auto-detects which machine it's on (junction check) and copies the correct variant to ~/.claude/CLAUDE.md, stamping it with a <!-- generated by Deploy-ClaudeConfig.ps1 ... --> marker.

Duplication-scan exclusion: CLAUDE.md is a generated build artifact kept byte-in-sync with its machine variant — it is expected to match CLAUDE.desktop.md/CLAUDE.laptop.md. Any duplication / dedup check must ignore the glob CLAUDE.md and flag only when it diverges from its expected variant (that's deploy drift, not a true duplicate). CLAUDE.shared.md is the canonical content source; edit it, then regenerate the variants.

CLAUDE.md Conventions

Behavior baked into the deployed CLAUDE.md:

  • Next-Prompt Continuity — after every substantive response, Claude Code appends a ready-to-paste follow-up under a --- separator using a <next>...</next> block. One imperative sentence, no preamble. Omitted when there's no natural continuation (e.g., simple lookups). When a Process Health flag is raised, the <next> block reflects the proposed rework path, not the flagged tactical fix.
  • Process Health — a proactive rework gate. When session context shows the same workaround was already applied in a prior session, or a known issue has persisted across two or more handoffs without permanent resolution, Claude stops and surfaces the pattern: names the recurrence, proposes one alternative approach with tradeoffs, and asks permission to pivot rather than silently re-applying the workaround. Only fires when prior-session evidence is explicit in context (no cross-session memory is inferred).

When changing behavior, update all four files (CLAUDE.md + the three variants) so the next Deploy-ClaudeConfig.ps1 run doesn't revert the change.

About

indows user-level Claude Code configuration: global skills (~/.claude/skills), commands, and related dotfiles—synced across machines; ai_enhancement docs live in a separate repo.

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

 
 
 

Contributors