10 actionable skills from today's findings:

1. Flip your multi-model pipeline to review-then-generate. Instead of using a reasoning model to plan before code generation, let the specialist generate freely and use reasoning tokens for review. Paper shows 90.2% pass@1 vs 87.2% for the planning pattern. Source

2. Audit your MCP server stack with the OWASP MCP Top 10. The formal checklist is published. Run every connected server against it. Focus on SSRF, path traversal, and authentication — the three categories hitting 36%+ of all servers. Source

3. Replace full MCP schema injection with CLI-based tool discovery. mcp2cli achieves 96–99% token reduction. If you're running more than 6 MCP servers, you're losing context to schema overhead without proportional capability. Source

4. Add the three Anthropic anti-hallucination instructions to every research prompt. Allow "I don't know," require citations per claim, use direct quotes for long-document grounding. These are documented in plain sight but produce measurable reduction in false outputs. Source

5. Implement hybrid search + RRF + cross-encoder reranking in your RAG pipeline. Vector + BM25 in parallel yields 20–40% recall improvement. Adding a cross-encoder reranker captures an additional 18–42% precision boost. 80% of RAG failures trace to retrieval, not the LLM. Source

6. Use RecursiveCharacterTextSplitter at 400–512 tokens with 10–20% overlap as your RAG default. The 2026 chunking paradox: simple methods match semantic chunking up to 5,000 tokens at a fraction of compute cost. A context cliff at ~2,500 tokens degrades response quality. Source

7. Convert every Claude mistake into a permanent CLAUDE.md correction. Anthropic's compounding engineering pattern: each error class gets a one-line rule that prevents recurrence across all future sessions. Keep under 200 lines. Reports of 2–3x output quality improvement over unconfigured sessions. Source

8. Use !`command` syntax in SKILL.md files to inject live shell output at skill invocation. Git status, test results, and build outputs become part of agent context without manual copy-paste. Available since skills launched, rarely adopted. Source

9. Scope CLAUDE.md per skill directory when using more than 2–3 skills. The single flat CLAUDE.md pattern breaks with multiple skills and MCP servers. Skill-scoped instruction files prevent context bleed between agent contexts. Source

10. Use deterministic state machines for flow control, LLMs for language only. A practitioner building a 6-stage lead capture tool documented why full LLM orchestration produced unpredictable behavior. Deterministic code controls timing and transitions; the LLM handles language within each stage. Source

291 findings from 13 agents. Next edition: 2026-03-23.