Yo claude settings 2026 07 10

yo-claude-settings - Behavioral & Model Baseline

This is the user's Claude Code baseline. Treat it as operating procedure, not a preference list.

Hard rules

1. Use the good brain by default

Default to the strongest available Claude model, fixed high thinking, and a normal working context. Do not use adaptive/auto thinking as the default for serious engineering work.

Practical behavior:

• Think before coding, even when the change looks small.
• Do not skim files that define the behavior being changed.
• Prefer a fresh session with a tight handoff over pushing a degraded long thread.
• Increase context only when the task genuinely needs breadth; do not make huge context the default.
• If the active Claude Code model/thinking/context setup looks wrong, say so briefly and offer to fix it.

2. Apply the Karpathy baseline on every change

Use the four engineering rules from references/karpathy-baseline.md:

• Think before coding.
• Simplicity first.
• Surgical changes.
• Goal-driven execution with concrete verification.

This means no speculative features, no unrelated refactors, no hidden confusion, and no completion claim without evidence.

3. Read project instructions first

On the first substantive task in a project, locate and read project instructions before editing:

find . -maxdepth 3 \( -name CLAUDE.md -o -name AGENTS.md \) -type f 2>/dev/null | head -20

Get-ChildItem -Recurse -Depth 3 -Include CLAUDE.md,AGENTS.md -File | Select-Object -First 20

Priority:

• CLAUDE.md is the Claude Code project instruction file.
• AGENTS.md is also relevant in repos shared with Codex.
• Nested instruction files can override broader ones for their subtree.

If project instructions contradict a generic best practice, the project instructions win unless they are unsafe or impossible.

4. Keep process state clean

Before starting a dev server or watcher, check whether one is already running. Do not stack processes that hold the same port or serve stale assets.

Unix-like quick checks:

pgrep -fl "dotnet run|dotnet watch|mdbook serve|vite|npm run dev" | head
lsof -i :3000 -i :5000 -i :5173 -i :8000 | grep LISTEN

Windows PowerShell quick checks:

Get-Process | Where-Object { $_.ProcessName -match 'node|dotnet|vite' }
Get-NetTCPConnection -State Listen | Where-Object { $_.LocalPort -in 3000,5000,5173,8000 }

Kill only the process you can identify as stale or conflicting.

Session ramp-up

1. Read this skill.
2. Read relevant CLAUDE.md / AGENTS.md files.
3. Check whether model/thinking/context look aligned with the user's baseline.
4. Use memory only when it is likely to help the repo/task.
5. Route the actual work through the right Yo skill.

Skills to pair with

• yo-engineering - senior risk gate for broad, high-risk, or "best engineer" work.
• yo-flow - route ambiguous work through the Yo pack.
• yo-verify-premise - verify claims before fixing or investigating.
• yo-feature - build a new feature from spec to verified change.
• yo-fix-loop - clear multiple findings with fix/verify discipline.
• yo-audit - independent read-only verification before trusting a change.
• yo-branch - branch and PR readiness conventions.
• yo-gh-write - GitHub issues, PRs, and review comments in the user's voice.

Anti-patterns

• Treating this as soft preference.
• Skipping project instructions because memory "probably knows."
• Silently accepting adaptive thinking on serious work.
• Expanding scope while claiming to be surgical.
• Starting another dev server because checking processes feels slower.

Sources

See references/sources.md and references/karpathy-baseline.md.