[{"data":1,"prerenderedAt":45},["ShallowReactive",2],{"$f2d4LJ2hbCZW9FCIGUA4vJmGnC2-ry0ElSdR7Na0lOHk":3},{"date":4,"generated_at":5,"picks":6,"candidates_scanned":44,"candidates_scored":17},"2026-04-15","2026-04-15T06:00:00.000000+00:00",[7,21,34],{"rank":8,"title":9,"source":10,"url":11,"category":12,"tldr":13,"score":14,"scores":15,"why":20},1,"v2.1.108","Claude Code Releases","https://github.com/anthropics/claude-code/releases/tag/v2.1.108","Release","- 1-hour prompt cache TTL is now opt-in with `ENABLE_PROMPT_CACHING_1H` — previously API key users (and anyone who set `DISABLE_TELEMETRY`) were silently capped at the 5-minute TTL, burning extra cost on long sessions\n- New `/recap` command summarizes what happened since you stepped away so you can resume context instantly without re-reading the full transcript\n- Claude can now discover and self-invoke built-in slash commands like `/init`, `/review`, and `/security-review` via the Skill tool — opens the door to agent workflows that trigger standard operations without manual prompting\n- Error messages are now meaningfully differentiated: rate limits vs plan limits are separate, 5xx/529 errors link directly to status.claude.com, and unknown slash commands suggest the nearest valid match\n- `/resume` defaults to sessions in your current directory; Ctrl+A expands to all projects",73,{"direct_claude_relevance":16,"practical_utility":17,"novelty":18,"source_credibility":19},33,20,7,13,"The 1-hour prompt cache TTL fix is the standout: users with `DISABLE_TELEMETRY` set were silently downgraded to 5-minute caching and paying the full price every 5 minutes on warm sessions — this restores the expected behavior. The `/recap` feature is a clean addition to long-session workflows where you walk away mid-task. Novelty is penalized because this is the sixth consecutive day of Claude Code release coverage, but the cache TTL fix and session recap are distinct from anything in v2.1.105 or earlier.",{"rank":22,"title":23,"source":24,"url":25,"category":26,"tldr":27,"score":28,"scores":29,"why":33},2,"YAML vs Markdown vs JSON vs TOON: Which Format Is Most Efficient for the Claude API","Dev.to Claude","https://dev.to/webramos/yaml-vs-markdown-vs-json-vs-toon-which-format-is-most-efficient-for-the-claude-api-4l94","Guide","- The same 200-product catalog costs 15,879 tokens in JSON vs 6,088 in TOON — a 62% difference that hits your API bill every call\n- This is the first public benchmark run on Claude specifically (Haiku 4.5, Sonnet 4.6, Opus 4.6): 450 API calls, 120 data files, 8 real-world scenarios, Anthropic's production tokenizer\n- Practical rule: prefer Markdown or TOON over JSON whenever you're passing structured data as context — not just in prompts, but in tool responses and retrieved documents",64,{"direct_claude_relevance":30,"practical_utility":30,"novelty":31,"source_credibility":32},22,14,6,"Token format efficiency is one of those cost levers that most API users ignore because the existing benchmarks all target GPT or Gemini. This is the first Claude-specific benchmark with real methodology: Anthropic's tokenizer, multiple models, multiple scenarios. The 62-73% savings gap between JSON and TOON/Markdown is large enough to meaningfully change how you structure tool outputs and context documents. Dev.to source credibility is modest, but the methodology is reproducible and the numbers come from a production tokenizer run.",{"rank":35,"title":36,"source":24,"url":37,"category":26,"tldr":38,"score":39,"scores":40,"why":43},3,"Context budget optimization: how to design MCP tools that don't waste tokens","https://dev.to/vdalhambra/context-budget-optimization-how-to-design-mcp-tools-that-dont-waste-tokens-3jcg","- Most MCP tools return raw API JSON (4,000+ tokens) when Claude only needs a pre-digested summary — a 10-100x token waste per call\n- Design MCP responses to include what changed, what it means, and what to do next — not raw data dumps",53,{"direct_claude_relevance":41,"practical_utility":17,"novelty":42,"source_credibility":32},16,11,"A complementary pair with today's Silver pick: if the format benchmark tells you which format to use, this tells you what to put in it. The FinanceKit case study is concrete — raw SDMX-XML → pre-summarized flat response — and the pattern generalizes to any MCP tool that proxies an external API. Different angle from yesterday's MCP server release (how to design tool responses vs announcing a new server).",42,1776402242933]