Claude Code Digest — Jul 04–Jul 07

Agentic coding is getting more expensive to debug than to generate: Lovable burned $85K in tokens, and that’s the part enterprises keep underestimating.

70% project drift reduction

Trending Now

🔥 The 150-Instruction Budget Is the New Ceiling for CLAUDE.md
Stop stuffing every rule into one file. The 3-layer harness pattern keeps CLAUDE.md as a short constitution, pushes domain rules into skills, and delegates task execution to subagents—so you stay under the compliance budget instead of training Claude to ignore your repo. Action: split policy, skills, and execution now. 🔥 Spec-First Agentic Blueprints Cut Drift from 40% to 12%
The 6-phase Spec Kit + Superpowers + GStack workflow claims a 70% drift reduction, with Phase 1 alone dropping drift by 30%. The takeaway: front-load specs before code, or you’ll spend your time reconciling divergent agent output instead of shipping. 📈 MCP Servers Are Being Reframed as LLM UIs, Not Tool Dumps
The winning pattern is fewer tools, tighter intent, and smaller payloads: cap servers at 20 tools and use JMESPath to cut response payloads by 80–90%. If your MCP server is a kitchen sink, Claude will pay for it in tokens and confusion.

Best Practices

Use a 3-layer harness: short CLAUDE.md + skills + subagents
Before: one bloated instruction file that exceeds the 150-instruction compliance budget and gets partially ignored. After: policy stays compact, skills stay reusable, and subagents handle execution without instruction bleed. Add `toWellFormed()` normalization before parsing Claude Code JSONL
Before: malformed Unicode or broken session lines can crash your parser or silently corrupt derived data. After: tolerant parsing survives bad records and keeps session analytics stable across `~/.claude/projects/` history. Use JMESPath projections in MCP responses
Before: verbose tool payloads waste tokens and bury the signal. After: trimmed responses cut payload size by 80–90%, which makes tool calls faster and less expensive to reason over.

Tools & MCP

FastMCP — Turns type-hinted Python functions into MCP tools with `@mcp.tool()` — fastest path from script to discoverable `tools/list` and `tools/call` endpoints. PaperQuire v0.3.0 — Renders Markdown into branded PDFs from Claude Code via `paperquire mcp-server` — turns a prompt into a polished PDF without leaving the agent loop. SkillsBench Caveman Skill — A minimal skill that reportedly saves 8.5% tokens with no quality loss — a cheap win if you want lower spend without changing your workflow.

Multi-Agent Patterns

4-Skill + 2-MCP Dev Team
Install `using-superpowers`, `writing-plans`, `subagent-driven-development`, and `requesting-code-review`, then pair them with two MCP servers to simulate a parallel dev team. It’s a cleaner alternative to 132-agent chaos: fewer moving parts, more role clarity, less context thrash. 3-Layer Agent Harness
Use CLAUDE.md for immutable repo constitution, skills for reusable specialist behaviors, and subagents for bounded execution. This solves instruction sprawl and keeps each layer small enough for Claude to actually follow. Uncertainty Loop Agent
Aletheia updates beliefs instead of guessing-and-summarizing, so it can return a verdict plus explicit confidence and residual unknowns. Best for tasks where false certainty is worse than a partial answer.

Community Requests

• Native MCP benchmarking that reports token cost, latency, and payload size per tool call
• A Claude Code parser library with versioned schemas for `~/.claude/projects/` JSONL sessions
• A first-class way to enforce instruction budgets across CLAUDE.md, skills, and subagents