Claude Code Digest — May 31–Jun 03

Claude Code is quietly becoming an operating system: teams are replacing brittle UI layers with deterministic tools, while per-project rules and skills finally make the agent behave like it belongs in the repo.

350+ tools in YC’s finance ecosystem

Trending Now

🔥 CLAUDE.md as the New Source of Truth for Repo Behavior
The big shift is that project rules now need to live in-repo, not in a user-level grab bag. If your instructions are getting ignored, the fix is to move them into `CLAUDE.md`, split them by domain, and make the important ones explicit enough that Claude can’t hand-wave them away. This matters because it turns prompt hygiene into versioned code review. 📈 Per-Project Skill Manifests End Global Skill Chaos
Global skill directories are becoming a liability: they drift, conflict, and make teams non-reproducible. `skills.toml` plus `skillkit sync` gives each repo a declared skill set, so onboarding becomes a sync step instead of a scavenger hunt. The practical move: treat skills like dependencies, not personal preferences. 🔥 YC’s 350+ Tool Model Beats UI-Building for Ops
The contrarian lesson is that many internal workflows should not get a polished interface at all. YC’s model shows that if you expose deterministic tools to an agent registry, non-coders can run finance/ops safely without a bespoke app layer. Build small tools, not dashboards; the agent becomes the interface.

Best Practices

Put repo rules in `CLAUDE.md` instead of relying on legacy prompt files
Before: rules in scattered or user-local config get skipped or forgotten. After: the repo owns behavior, so Claude sees the same constraints every time and team output becomes consistent. Split `CLAUDE.md` by domain with `alwaysApply: true` for the rules that must never drift
Before: one giant instruction blob gets diluted and ignored. After: domain-specific rules stay targeted, and the always-on ones survive context churn. Run `skillkit sync` from a `skills.toml` manifest to make skills reproducible per project
Before: every developer accumulates a different global skill set and Claude behavior varies by machine. After: the repo declares skills once, and everyone gets the same toolchain on sync.

Tools & MCP

CLAUDE.md — Repo-local instruction layer that makes Claude obey project rules instead of drifting across machines and sessions. skillkit — Per-project skill manager that syncs declared skills from `skills.toml` — kills global skill sprawl and makes setup deterministic. Agent registry tool layer — YC-style deterministic tool surface that lets non-coders execute ops without building a full UI — fewer interfaces, faster execution.

Multi-Agent Patterns

Agent Registry Over UI Pattern
Expose a registry of small, deterministic tools instead of a monolithic interface. The agent routes requests to the right tool, which keeps operations auditable and lets non-coders execute workflows safely. Domain-Split Rules Pattern
Break project instructions into separate rule files by domain so the agent only loads what matters for the task. This reduces instruction dilution and makes behavior easier to debug. Manifest-Declared Skills Pattern
Keep skills in `skills.toml` and sync them into the project so every agent instance sees the same capabilities. This solves the “works on my machine” problem for Claude skills.

Community Requests

• Native support for repo-scoped skills without needing a separate sync tool
• A first-class agent registry format with permissions, audit logs, and tool versioning
• Better debugging for why a specific `CLAUDE.md` rule was ignored