{"id":8547,"date":"2026-07-30T17:30:00","date_gmt":"2026-07-30T09:30:00","guid":{"rendered":"https:\/\/meta-quantum.today\/?p=8547"},"modified":"2026-07-30T13:44:47","modified_gmt":"2026-07-30T05:44:47","slug":"claude-5-changes-how-we-use-ai-harness-claude-code","status":"publish","type":"post","link":"https:\/\/meta-quantum.today\/?p=8547","title":{"rendered":"Claude 5 Changes How We Use AI &#038; Harness Claude Code"},"content":{"rendered":"\n<h3 class=\"wp-block-heading\">Claude 5 \u5982\u4f55\u6539\u53d8\u6211\u4eec\u4f7f\u7528 AI \u4e0e Claude Code \u7684\u65b9\u5f0f<\/h3>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<h2 class=\"wp-block-heading\">Introduction \/ \u5f15\u8a00<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">This review dives into two documents released almost simultaneously by Anthropic: the official &#8220;context engineering for Claude 5 generation models&#8221; guide (published July 24, 2026) and a &#8220;field guide&#8221; for Fable 5 published two weeks earlier. The presenter \u2014 a self-described theoretical physics and mathematics power-user of Claude Code \u2014 reads both papers side by side and extracts what genuinely changes for coders and heavy AI harness users, rather than repeating example-level content anyone can read themselves.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u8fd9\u671f\u8bc4\u8bba\u89e3\u8bfb\u4e86 Anthropic \u51e0\u4e4e\u540c\u65f6\u53d1\u5e03\u7684\u4e24\u4efd\u6587\u4ef6:\u4e00\u4efd\u662f 2026 \u5e74 7 \u6708 24 \u65e5\u53d1\u5e03\u7684\u5b98\u65b9&#8221;Claude 5 \u4e16\u4ee3\u6a21\u578b\u7684\u4e0a\u4e0b\u6587\u5de5\u7a0b&#8221;\u6307\u5357,\u53e6\u4e00\u4efd\u662f\u4e24\u5468\u524d\u53d1\u5e03\u7684 Fable 5 &#8220;\u73b0\u573a\u6307\u5357&#8221;(field guide)\u3002\u89c6\u9891\u4f5c\u8005\u662f\u4e00\u4f4d\u957f\u671f\u4f7f\u7528 Claude Code \u5904\u7406\u7406\u8bba\u7269\u7406\u4e0e\u6570\u5b66\u4efb\u52a1\u7684\u91cd\u5ea6\u7528\u6237,\u4ed6\u5c06\u4e24\u4efd\u6587\u4ef6\u5bf9\u7167\u9605\u8bfb,\u63d0\u70bc\u51fa\u5bf9\u7a0b\u5e8f\u5458\u548c\u91cd\u5ea6 AI harness \u7528\u6237\u771f\u6b63\u6709\u610f\u4e49\u7684\u53d8\u5316,\u800c\u4e0d\u662f\u590d\u8ff0\u4eba\u4eba\u90fd\u80fd\u81ea\u884c\u9605\u8bfb\u7684\u793a\u4f8b\u5185\u5bb9\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">The core thesis: Claude Opus 5 and the broader Claude 5 generation (Fable 5, Mythos 5) shift the bottleneck away from the model&#8217;s raw capability and toward the <em>human&#8217;s specification<\/em>. Anthropic removed roughly 80% of the Claude Code system prompt, arguing that heavily engineered, rule-dense harness prompts now constrain a more capable model rather than help it.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u6838\u5fc3\u8bba\u70b9\u662f:Claude Opus 5 \u4ee5\u53ca\u66f4\u5e7f\u6cdb\u7684 Claude 5 \u4e16\u4ee3(Fable 5\u3001Mythos 5)\u628a&#8221;\u74f6\u9888&#8221;\u4ece\u6a21\u578b\u672c\u8eab\u7684\u80fd\u529b,\u8f6c\u79fb\u5230\u4e86&#8221;\u4eba\u7c7b\u7ed9\u51fa\u7684\u4efb\u52a1\u89c4\u683c\u8bf4\u660e&#8221;\u4e0a\u3002Anthropic \u79fb\u9664\u4e86 Claude Code \u7cfb\u7edf\u63d0\u793a\u8bcd\u4e2d\u5927\u7ea6 80% \u7684\u5185\u5bb9,\u7406\u7531\u662f\u8fc7\u5ea6\u5de5\u7a0b\u5316\u3001\u89c4\u5219\u5806\u780c\u7684 harness \u63d0\u793a\u8bcd\u53cd\u800c\u4f1a\u9650\u5236\u4e00\u4e2a\u80fd\u529b\u66f4\u5f3a\u7684\u6a21\u578b,\u800c\u4e0d\u662f\u5e2e\u52a9\u5b83\u3002<\/p>\n\n\n\n<figure class=\"wp-block-video\"><video height=\"720\" style=\"aspect-ratio: 1280 \/ 720;\" width=\"1280\" autoplay controls loop muted src=\"https:\/\/meta-quantum.today\/wp-content\/uploads\/2026\/07\/Claude-5-Context-Engineering.mp4\" playsinline><\/video><\/figure>\n\n\n\n<h2 class=\"wp-block-heading\">Features and Concepts: What&#8217;s New in This Version \/ \u65b0\u7248\u672c\u7684\u529f\u80fd\u4e0e\u6838\u5fc3\u6982\u5ff5<\/h2>\n\n\n\n<h3 class=\"wp-block-heading\">1. From Rule-Heavy Prompts to Intent-Based Specification \/ \u4ece\u89c4\u5219\u5806\u780c\u5230&#8221;\u610f\u56fe\u5bfc\u5411&#8221;\u7684\u89c4\u683c\u8bf4\u660e<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">The old paradigm compensated for a weaker, less reliable model with long system prompts, exhaustive examples, rigid step-by-step workflows, and explicit rules for every edge case. The new paradigm asks the human to define only: intent, environment\/reality, constraints, and evidence \u2014 then lets the model discover the execution path itself.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u65e7\u8303\u5f0f\u7528\u5197\u957f\u7684\u7cfb\u7edf\u63d0\u793a\u8bcd\u3001\u6d77\u91cf\u793a\u4f8b\u3001\u50f5\u5316\u7684\u5206\u6b65\u6d41\u7a0b\u4ee5\u53ca\u9488\u5bf9\u6bcf\u4e2a\u8fb9\u7f18\u60c5\u51b5\u7684\u660e\u786e\u89c4\u5219,\u6765\u5f25\u8865\u6a21\u578b\u80fd\u529b\u8f83\u5f31\u3001\u4e0d\u591f\u53ef\u9760\u7684\u95ee\u9898\u3002\u65b0\u8303\u5f0f\u5219\u8981\u6c42\u4eba\u7c7b\u53ea\u9700\u5b9a\u4e49:<strong>\u610f\u56fe(intent)\u3001\u73af\u5883\/\u73b0\u5b9e(reality)\u3001\u7ea6\u675f\u6761\u4ef6(constraints)\u3001\u4ee5\u53ca\u9a8c\u8bc1\u8bc1\u636e(evidence)<\/strong>\u2014\u2014\u7136\u540e\u8ba9\u6a21\u578b\u81ea\u884c\u53d1\u73b0\u6267\u884c\u8def\u5f84\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">Anthropic frames the remaining human responsibility using a four-quadrant model of knowledge:<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">Anthropic \u7528\u4e00\u4e2a\u56db\u8c61\u9650\u7684\u77e5\u8bc6\u6846\u67b6\u6765\u63cf\u8ff0\u4eba\u7c7b\u6b64\u540e\u4ecd\u9700\u627f\u62c5\u7684\u8d23\u4efb:<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>Known knowns<\/strong> \u2014 facts you explicitly provide \/ \u4f60\u660e\u786e\u63d0\u4f9b\u7684\u5df2\u77e5\u4e8b\u5b9e<\/li>\n\n\n\n<li><strong>Known unknowns<\/strong> \u2014 open questions you&#8217;re aware of \/ \u4f60\u610f\u8bc6\u5230\u4f46\u5c1a\u672a\u89e3\u51b3\u7684\u95ee\u9898<\/li>\n\n\n\n<li><strong>Unknown knowns<\/strong> \u2014 tacit preferences you only recognize when you see the wrong alternative \/ \u53ea\u6709\u770b\u5230&#8221;\u4e0d\u5bf9\u7684\u65b9\u6848&#8221;\u65f6\u624d\u4f1a\u610f\u8bc6\u5230\u7684\u9690\u6027\u504f\u597d<\/li>\n\n\n\n<li><strong>Unknown unknowns<\/strong> \u2014 issues neither human nor model has considered yet \/ \u4eba\u7c7b\u4e0e\u6a21\u578b\u90fd\u5c1a\u672a\u60f3\u5230\u7684\u95ee\u9898<\/li>\n<\/ul>\n\n\n\n<h3 class=\"wp-block-heading\">2. Bi-Directional Prompting \/ \u53cc\u5411\u63d0\u793a(Bi-Directional Prompting)<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">Perhaps the most practically significant shift: prompting is no longer one-directional. The agent is now expected to interview the human back \u2014 asking prioritized, one-at-a-time questions, focused only on ambiguities that would materially change architecture, interpretation, cost, or final output \u2014 rather than guessing or burning tokens on the wrong path.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u4e5f\u8bb8\u662f\u6700\u5177\u5b9e\u9645\u610f\u4e49\u7684\u53d8\u5316:\u63d0\u793a(prompting)\u4e0d\u518d\u662f\u5355\u5411\u7684\u3002\u73b0\u5728\u671f\u671b\u667a\u80fd\u4f53(agent)\u53cd\u8fc7\u6765\u5411\u4eba\u7c7b\u63d0\u95ee\u2014\u2014\u4e00\u6b21\u53ea\u95ee\u4e00\u4e2a\u3001\u6309\u4f18\u5148\u7ea7\u6392\u5e8f\u7684\u95ee\u9898,\u5e76\u4e14\u53ea\u9488\u5bf9\u90a3\u4e9b\u4f1a\u5b9e\u8d28\u6027\u6539\u53d8\u67b6\u6784\u3001\u89e3\u91ca\u65b9\u5f0f\u3001\u6210\u672c\u6216\u6700\u7ec8\u8f93\u51fa\u7684\u6a21\u7cca\u70b9\u63d0\u95ee,\u800c\u4e0d\u662f\u9760\u731c\u6d4b\u6216\u5728\u9519\u8bef\u65b9\u5411\u4e0a\u6d88\u8017\u5927\u91cf token\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">3. Progressive Disclosure Replaces the Monolithic System Prompt \/ \u6e10\u8fdb\u5f0f\u62ab\u9732\u53d6\u4ee3&#8221;\u5de8\u77f3\u578b&#8221;\u7cfb\u7edf\u63d0\u793a\u8bcd<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">Instead of stuffing coding rules, review procedures, documentation standards, deployment\/testing instructions, and every organizational policy into one giant system prompt, Claude Code now starts with a small, high-level context and lets the agent retrieve specialized instructions \u2014 via selectively loaded <strong>Skills<\/strong> \u2014 only when it decides they&#8217;re relevant. Some tool definitions are similarly deferred until the model actively searches for them.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">Claude Code \u4e0d\u518d\u628a\u7f16\u7801\u89c4\u5219\u3001\u4ee3\u7801\u5ba1\u67e5\u6d41\u7a0b\u3001\u6587\u6863\u6807\u51c6\u3001\u90e8\u7f72\u4e0e\u6d4b\u8bd5\u8bf4\u660e\u4ee5\u53ca\u5404\u79cd\u7ec4\u7ec7\u653f\u7b56\u5168\u90e8\u585e\u8fdb\u4e00\u4e2a\u5de8\u5927\u7684\u7cfb\u7edf\u63d0\u793a\u8bcd,\u800c\u662f\u4ece\u4e00\u4e2a\u7cbe\u7b80\u7684\u3001\u9ad8\u5c42\u7ea7\u7684\u4e0a\u4e0b\u6587\u5f00\u59cb,\u8ba9\u667a\u80fd\u4f53\u5728\u81ea\u884c\u5224\u65ad&#8221;\u786e\u5b9e\u9700\u8981&#8221;\u65f6,\u518d\u901a\u8fc7\u6309\u9700\u52a0\u8f7d\u7684 <strong>Skills(\u6280\u80fd\u6a21\u5757)<\/strong> \u53bb\u68c0\u7d22\u4e13\u95e8\u7684\u6307\u4ee4\u3002\u90e8\u5206\u5de5\u5177\u5b9a\u4e49(tool definitions)\u540c\u6837\u88ab\u5ef6\u540e\u52a0\u8f7d,\u76f4\u5230\u6a21\u578b\u4e3b\u52a8\u641c\u7d22\u5b83\u4eec\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">This mirrors a broader industry pattern documented in community write-ups on progressive-disclosure skill architectures: a lightweight YAML\/markdown frontmatter (roughly ~100 tokens) determines whether a skill activates, while the full skill body (kept under roughly 5k tokens) only loads when triggered \u2014 keeping the resting context small while expert-level capability stays one retrieval away.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u8fd9\u4e0e\u793e\u533a\u6587\u6863\u4e2d\u8bb0\u5f55\u7684\u66f4\u5e7f\u6cdb\u7684&#8221;\u6e10\u8fdb\u5f0f\u62ab\u9732\u6280\u80fd\u67b6\u6784&#8221;\u6a21\u5f0f\u76f8\u547c\u5e94:\u4e00\u4e2a\u8f7b\u91cf\u7684 YAML\/markdown \u524d\u7f6e\u5143\u6570\u636e(\u7ea6 100 tokens)\u51b3\u5b9a\u67d0\u4e2a\u6280\u80fd\u662f\u5426\u88ab\u6fc0\u6d3b,\u800c\u5b8c\u6574\u7684\u6280\u80fd\u6b63\u6587(\u63a7\u5236\u5728\u7ea6 5000 tokens \u4ee5\u5185)\u53ea\u6709\u5728\u88ab\u89e6\u53d1\u65f6\u624d\u4f1a\u52a0\u8f7d\u2014\u2014\u8fd9\u6837\u65e2\u80fd\u4fdd\u6301\u9759\u606f\u72b6\u6001\u4e0b\u4e0a\u4e0b\u6587\u7cbe\u7b80,\u53c8\u80fd\u8ba9\u4e13\u5bb6\u7ea7\u80fd\u529b&#8221;\u4e00\u89e6\u5373\u8fbe&#8221;\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">4. References Instead of Exhaustive Examples \/ \u7528&#8221;\u53c2\u8003\u8303\u4f8b&#8221;\u53d6\u4ee3\u7a77\u4e3e\u5f0f\u793a\u4f8b<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">Rather than listing every desired attribute of an output (font size, background, layout \u2014 and inevitably forgetting one or two), the new approach hands the model a structural reference \u2014 a prior output the human was happy with \u2014 and specifies what to preserve (typography, hierarchy, ideas-per-slide, progressive reveal pattern) and, just as importantly, what <em>not<\/em> to preserve.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u65b0\u65b9\u6cd5\u4e0d\u518d\u662f\u9010\u6761\u5217\u51fa\u8f93\u51fa\u7684\u6240\u6709\u671f\u671b\u5c5e\u6027(\u5b57\u4f53\u5927\u5c0f\u3001\u80cc\u666f\u3001\u6392\u7248\u2014\u2014\u800c\u4e14\u96be\u514d\u4f1a\u6f0f\u6389\u4e00\u4e24\u6761),\u800c\u662f\u7ed9\u6a21\u578b\u4e00\u4e2a<strong>\u7ed3\u6784\u6027\u53c2\u8003\u8303\u4f8b<\/strong>(\u4eba\u7c7b\u4e4b\u524d\u6ee1\u610f\u7684\u67d0\u4e2a\u8f93\u51fa),\u5e76\u660e\u786e\u6307\u51fa\u54ea\u4e9b\u5143\u7d20\u9700\u8981\u4fdd\u7559(\u6392\u7248\u98ce\u683c\u3001\u5c42\u7ea7\u7ed3\u6784\u3001\u6bcf\u9875\u4fe1\u606f\u5bc6\u5ea6\u3001\u6e10\u8fdb\u5c55\u793a\u7684\u8282\u594f),\u540c\u6837\u91cd\u8981\u7684\u662f,\u4e5f\u8981\u8bf4\u660e\u54ea\u4e9b\u5143\u7d20<strong>\u4e0d\u9700\u8981<\/strong>\u4fdd\u7559\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">5. A New Planning Loop with Self-Learning \/ \u5177\u5907\u81ea\u6211\u5b66\u4e60\u80fd\u529b\u7684\u65b0\u89c4\u5212\u5faa\u73af<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">The workflow expands from plan \u2192 implement \u2192 result into: discover unknowns \u2192 brainstorm alternatives \u2192 build a prototype \u2192 interview the human \u2192 study references \u2192 produce a reviewable plan \u2192 implement \u2192 record deviations \u2192 explain and verify. Anthropic recommends maintaining an <strong>implementation notes file<\/strong> that records deviations from the plan, discovered edge cases, assumptions made, and lessons for future attempts \u2014 turning execution into a source of context for self-learning, not just a pass\/fail result.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u5de5\u4f5c\u6d41\u4ece&#8221;\u89c4\u5212 \u2192 \u5b9e\u73b0 \u2192 \u7ed3\u679c&#8221;\u6269\u5c55\u4e3a:\u53d1\u73b0\u672a\u77e5\u9879 \u2192 \u5934\u8111\u98ce\u66b4\u5907\u9009\u65b9\u6848 \u2192 \u6784\u5efa\u539f\u578b \u2192 \u5411\u4eba\u7c7b\u63d0\u95ee\u786e\u8ba4 \u2192 \u7814\u7a76\u53c2\u8003\u8303\u4f8b \u2192 \u751f\u6210\u53ef\u5ba1\u9605\u7684\u8ba1\u5212 \u2192 \u5b9e\u65bd \u2192 \u8bb0\u5f55\u504f\u5dee \u2192 \u89e3\u91ca\u5e76\u9a8c\u8bc1\u3002Anthropic \u5efa\u8bae\u7ef4\u62a4\u4e00\u4efd<strong>\u5b9e\u65bd\u7b14\u8bb0\u6587\u4ef6(implementation notes file)<\/strong>,\u8bb0\u5f55\u4e0e\u539f\u8ba1\u5212\u7684\u504f\u5dee\u3001\u53d1\u73b0\u7684\u8fb9\u7f18\u60c5\u51b5\u3001\u6240\u505a\u7684\u5047\u8bbe\u4ee5\u53ca\u4e3a\u540e\u7eed\u5c1d\u8bd5\u603b\u7ed3\u7684\u7ecf\u9a8c\u6559\u8bad\u2014\u2014\u8ba9&#8221;\u6267\u884c\u8fc7\u7a0b&#8221;\u672c\u8eab\u6210\u4e3a\u81ea\u6211\u5b66\u4e60\u7684\u4e0a\u4e0b\u6587\u6765\u6e90,\u800c\u4e0d\u4ec5\u4ec5\u662f\u4e00\u4e2a\u901a\u8fc7\/\u5931\u8d25\u7684\u7ed3\u679c\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">6. A Layered Memory Model \/ \u5206\u5c42\u7684\u8bb0\u5fc6\u6a21\u578b<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">Claude Code has moved from manually stuffing everything into <code>CLAUDE.md<\/code> to an automatic memory function \u2014 but this doesn&#8217;t eliminate the need for a deliberate memory structure. Anthropic recommends keeping <code>CLAUDE.md<\/code> lightweight, reserved mainly for unusual repository-specific gotchas, while a sensible layered structure looks like:<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">Claude Code \u5df2\u7ecf\u4ece\u624b\u52a8\u628a\u6240\u6709\u5185\u5bb9\u585e\u8fdb <code>CLAUDE.md<\/code>,\u5347\u7ea7\u4e3a\u4e00\u4e2a\u81ea\u52a8\u8bb0\u5fc6\u529f\u80fd\u2014\u2014\u4f46\u8fd9\u5e76\u4e0d\u610f\u5473\u7740\u4e0d\u518d\u9700\u8981\u6709\u610f\u8bc6\u5730\u8bbe\u8ba1\u8bb0\u5fc6\u7ed3\u6784\u3002Anthropic \u5efa\u8bae\u8ba9 <code>CLAUDE.md<\/code> \u4fdd\u6301\u8f7b\u91cf,\u4e3b\u8981\u7528\u4e8e\u8bb0\u5f55\u8be5\u4ee3\u7801\u4ed3\u5e93\u7279\u6709\u7684&#8221;\u53cd\u5e38\u89c4&#8221;\u6ce8\u610f\u4e8b\u9879\u3002\u4e00\u4e2a\u5408\u7406\u7684\u5206\u5c42\u8bb0\u5fc6\u7ed3\u6784\u5927\u81f4\u5982\u4e0b:<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>Project instructions<\/strong> \u2014 the project, its rules, usual\/unusual limitations \/ \u9879\u76ee\u8bf4\u660e\u2014\u2014\u9879\u76ee\u672c\u8eab\u3001\u89c4\u5219\u3001\u5e38\u89c4\u4e0e\u7279\u6b8a\u9650\u5236<\/li>\n\n\n\n<li><strong>Task artifacts<\/strong> \u2014 current spec, markup, test plans \/ \u4efb\u52a1\u4ea7\u7269\u2014\u2014\u5f53\u524d\u89c4\u683c\u8bf4\u660e\u3001\u6807\u8bb0\u3001\u6d4b\u8bd5\u8ba1\u5212<\/li>\n\n\n\n<li><strong>Long-term memory<\/strong> \u2014 validated lessons affecting future runs \/ \u957f\u671f\u8bb0\u5fc6\u2014\u2014\u7ecf\u8fc7\u9a8c\u8bc1\u3001\u5f71\u54cd\u672a\u6765\u6267\u884c\u7684\u7ecf\u9a8c\u6559\u8bad<\/li>\n\n\n\n<li><strong>Current execution state<\/strong> \u2014 what&#8217;s been completed this run, and why prior steps succeeded or failed \/ \u5f53\u524d\u6267\u884c\u72b6\u6001\u2014\u2014\u672c\u6b21\u8fd0\u884c\u5df2\u5b8c\u6210\u7684\u5185\u5bb9,\u4ee5\u53ca\u4e4b\u524d\u6b65\u9aa4\u6210\u529f\u6216\u5931\u8d25\u7684\u539f\u56e0<\/li>\n<\/ul>\n\n\n\n<h3 class=\"wp-block-heading\">7. Claude Code as a Harness Around the Model, Not the Model Itself \/ Claude Code \u662f\u56f4\u7ed5\u6a21\u578b\u7684&#8221;\u5916\u58f3&#8221;,\u800c\u975e\u6a21\u578b\u672c\u4f53<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">A key conceptual clarification: Claude Code is a <em>product<\/em> \u2014 a harness (system prompt, project context, <code>CLAUDE.md<\/code>, skills, file tools, shell execution, web access, permissions, sub-agent spawning, agentic looping, API access) wrapped around a core LLM (Opus 5, Fable 5, or Mythos 5). The harness constructs context, retrieves memory and skills, manages tools, executes actions, and runs repeated model calls \u2014 a layer distinct from the core model&#8217;s own reasoning.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u4e00\u4e2a\u5173\u952e\u7684\u6982\u5ff5\u6f84\u6e05:Claude Code \u662f\u4e00\u4e2a<strong>\u4ea7\u54c1<\/strong>\u2014\u2014\u662f\u4e00\u4e2a\u56f4\u7ed5\u6838\u5fc3 LLM(Opus 5\u3001Fable 5 \u6216 Mythos 5)\u6784\u5efa\u7684&#8221;\u5916\u58f3&#8221;(harness),\u5305\u542b\u7cfb\u7edf\u63d0\u793a\u8bcd\u3001\u9879\u76ee\u4e0a\u4e0b\u6587\u3001<code>CLAUDE.md<\/code>\u3001\u6280\u80fd\u6a21\u5757\u3001\u6587\u4ef6\u5de5\u5177\u3001shell \u6267\u884c\u3001\u7f51\u7edc\u8bbf\u95ee\u3001\u6743\u9650\u7ba1\u7406\u3001\u5b50\u667a\u80fd\u4f53\u751f\u6210\u3001\u667a\u80fd\u4f53\u5faa\u73af(agentic looping)\u4ee5\u53ca API \u8bbf\u95ee\u3002\u8fd9\u4e2a harness \u8d1f\u8d23\u6784\u5efa\u4e0a\u4e0b\u6587\u3001\u68c0\u7d22\u8bb0\u5fc6\u4e0e\u6280\u80fd\u3001\u7ba1\u7406\u5de5\u5177\u3001\u6267\u884c\u52a8\u4f5c\u3001\u5e76\u53cd\u590d\u8c03\u7528\u6a21\u578b\u2014\u2014\u5b83\u662f\u4e0e\u6838\u5fc3\u6a21\u578b\u81ea\u8eab\u63a8\u7406\u80fd\u529b\u76f8\u72ec\u7acb\u7684\u4e00\u5c42\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">This reframing also explains why some long-running, autonomous-agent workflows in the wild \u2014 where a coding agent is given a single prompt and left to run for hours without interference \u2014 succeed or fail based as much on harness design as on the underlying model&#8217;s intelligence, a pattern echoed in community case studies of long-running Claude Code agents building projects end-to-end without human interruption.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u8fd9\u4e00\u91cd\u65b0\u5b9a\u4f4d\u4e5f\u89e3\u91ca\u4e86\u4e3a\u4ec0\u4e48\u73b0\u5b9e\u4e2d\u4e00\u4e9b&#8221;\u957f\u65f6\u95f4\u81ea\u4e3b\u8fd0\u884c&#8221;\u7684\u667a\u80fd\u4f53\u5de5\u4f5c\u6d41\u2014\u2014\u5373\u7ed9\u7f16\u7801\u667a\u80fd\u4f53\u4e00\u4e2a\u63d0\u793a\u8bcd,\u7136\u540e\u8ba9\u5b83\u8fde\u7eed\u8fd0\u884c\u6570\u5c0f\u65f6\u800c\u4e0d\u52a0\u5e72\u9884\u2014\u2014\u5176\u6210\u8d25\u5f80\u5f80\u540c\u6837\u53d6\u51b3\u4e8e harness \u7684\u8bbe\u8ba1,\u800c\u4e0d\u4ec5\u4ec5\u662f\u5e95\u5c42\u6a21\u578b\u7684\u667a\u80fd\u6c34\u5e73\u3002\u8fd9\u4e00\u70b9\u5728\u793e\u533a\u4e2d\u8bb0\u5f55\u7684\u3001Claude Code \u667a\u80fd\u4f53\u5728\u65e0\u4eba\u5e72\u9884\u4e0b\u7aef\u5230\u7aef\u6784\u5efa\u9879\u76ee\u7684\u6848\u4f8b\u7814\u7a76\u4e2d\u4e5f\u5f97\u5230\u4e86\u5370\u8bc1\u3002<\/p>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<h2 class=\"wp-block-heading\">Video about Claude 5 \/ \u5173\u4e8e Clause 5 \u7684\u89c6\u9891<\/h2>\n\n\n\n<figure class=\"wp-block-embed is-type-video is-provider-youtube wp-block-embed-youtube wp-embed-aspect-16-9 wp-has-aspect-ratio\"><div class=\"wp-block-embed__wrapper\">\n<iframe loading=\"lazy\" title=\"Opus 5 Changes How We Use AI &amp; Harness Claude Code\" width=\"500\" height=\"281\" src=\"https:\/\/www.youtube.com\/embed\/zmELzKKgdIo?feature=oembed\" frameborder=\"0\" allow=\"accelerometer; autoplay; clipboard-write; encrypted-media; gyroscope; picture-in-picture; web-share\" referrerpolicy=\"strict-origin-when-cross-origin\" allowfullscreen><\/iframe>\n<\/div><\/figure>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<h2 class=\"wp-block-heading\">Setup, Configuration, and Practical Examples \/ \u8bbe\u7f6e\u3001\u914d\u7f6e\u4e0e\u5b9e\u8df5\u793a\u4f8b<\/h2>\n\n\n\n<h3 class=\"wp-block-heading\">Try the <code>\/doctor<\/code> command \/ \u5c1d\u8bd5 <code>\/doctor<\/code> \u547d\u4ee4<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">Anthropic has folded its context-engineering best practices into a built-in command: running <code>\/doctor<\/code> in Claude Code is intended to help &#8220;right-size&#8221; your skills and <code>CLAUDE.md<\/code> files, flagging bloat or misconfiguration rather than requiring manual audits.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">Anthropic \u5df2\u7ecf\u628a\u5176\u4e0a\u4e0b\u6587\u5de5\u7a0b\u6700\u4f73\u5b9e\u8df5\u6574\u5408\u8fdb\u4e86\u4e00\u4e2a\u5185\u7f6e\u547d\u4ee4:\u5728 Claude Code \u4e2d\u8fd0\u884c <code>\/doctor<\/code> \u65e8\u5728\u5e2e\u52a9\u4f60&#8221;\u5408\u7406\u8c03\u6574&#8221;\u6280\u80fd\u6a21\u5757\u4e0e <code>CLAUDE.md<\/code> \u6587\u4ef6\u7684\u89c4\u6a21,\u81ea\u52a8\u6807\u8bb0\u5197\u4f59\u6216\u914d\u7f6e\u4e0d\u5f53\u4e4b\u5904,\u800c\u4e0d\u9700\u8981\u4eba\u5de5\u9010\u4e00\u6392\u67e5\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">Reworking your <code>CLAUDE.md<\/code> \/ \u91cd\u65b0\u8c03\u6574\u4f60\u7684 <code>CLAUDE.md<\/code><\/h3>\n\n\n\n<ul class=\"wp-block-list\">\n<li>Strip it down to unusual, repository-specific gotchas only \/ \u53ea\u4fdd\u7559\u8be5\u9879\u76ee\u7279\u6709\u7684&#8221;\u53cd\u5e38\u89c4&#8221;\u6ce8\u610f\u4e8b\u9879<\/li>\n\n\n\n<li>Move coding rules, review procedures, and deployment\/testing instructions into <strong>selectively loaded skills<\/strong> instead of the root file \/ \u5c06\u7f16\u7801\u89c4\u5219\u3001\u5ba1\u67e5\u6d41\u7a0b\u3001\u90e8\u7f72\u4e0e\u6d4b\u8bd5\u8bf4\u660e\u8fc1\u79fb\u5230<strong>\u6309\u9700\u52a0\u8f7d\u7684\u6280\u80fd\u6a21\u5757<\/strong>\u4e2d,\u800c\u4e0d\u662f\u585e\u8fdb\u6839\u76ee\u5f55\u6587\u4ef6<\/li>\n\n\n\n<li>Reserve the automatic memory feature for durable, cross-session lessons \u2014 don&#8217;t try to replicate it manually \/ \u628a\u81ea\u52a8\u8bb0\u5fc6\u529f\u80fd\u7559\u7ed9\u53ef\u6301\u7eed\u3001\u8de8\u4f1a\u8bdd\u7684\u7ecf\u9a8c\u6559\u8bad,\u4e0d\u8981\u624b\u52a8\u91cd\u590d\u5b9e\u73b0\u5b83<\/li>\n<\/ul>\n\n\n\n<h3 class=\"wp-block-heading\">Rewriting prompts: rule-based \u2192 objective-based \/ \u63d0\u793a\u8bcd\u91cd\u5199:\u4ece&#8221;\u89c4\u5219\u5bfc\u5411&#8221;\u5230&#8221;\u76ee\u6807\u5bfc\u5411&#8221;<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>Old style example (rule-based):<\/strong><br>&#8220;Never write multi-line comments. Never create documentation unless I ask. Use one-line comments only. Do not explain obvious code.&#8221;<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>\u65e7\u5f0f\u5199\u6cd5\u793a\u4f8b(\u89c4\u5219\u5bfc\u5411):<\/strong><br>&#8220;\u6c38\u8fdc\u4e0d\u8981\u5199\u591a\u884c\u6ce8\u91ca\u3002\u9664\u975e\u6211\u8981\u6c42,\u5426\u5219\u4e0d\u8981\u521b\u5efa\u6587\u6863\u3002\u53ea\u7528\u5355\u884c\u6ce8\u91ca\u3002\u4e0d\u8981\u89e3\u91ca\u663e\u800c\u6613\u89c1\u7684\u4ee3\u7801\u3002&#8221;<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>New style example (objective-based):<\/strong><br>&#8220;Match the conventions of the surrounding code. Preserve existing public interfaces. Do not introduce unrelated changes.&#8221;<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>\u65b0\u5f0f\u5199\u6cd5\u793a\u4f8b(\u76ee\u6807\u5bfc\u5411):<\/strong><br>&#8220;\u4e0e\u5468\u56f4\u4ee3\u7801\u7684\u98ce\u683c\u4fdd\u6301\u4e00\u81f4\u3002\u4fdd\u7559\u73b0\u6709\u7684\u516c\u5171\u63a5\u53e3(public interfaces)\u3002\u4e0d\u8981\u5f15\u5165\u65e0\u5173\u7684\u6539\u52a8\u3002&#8221;<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">Example: research\/paper-summarization task, restructured \/ \u793a\u4f8b:\u91cd\u65b0\u8bbe\u8ba1\u4e00\u4e2a&#8221;\u8bba\u6587\u603b\u7ed3&#8221;\u4efb\u52a1<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>Old style:<\/strong> &#8220;Read this paper, summarize the abstract, explain every section, explain every mathematical term, analyze every equation, discuss every table, use exactly 10 sections.&#8221;<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>\u65e7\u5f0f\u5199\u6cd5:<\/strong> &#8220;\u9605\u8bfb\u8fd9\u7bc7\u8bba\u6587,\u603b\u7ed3\u6458\u8981,\u89e3\u91ca\u6bcf\u4e00\u8282,\u89e3\u91ca\u6bcf\u4e00\u4e2a\u6570\u5b66\u672f\u8bed,\u5206\u6790\u6bcf\u4e00\u4e2a\u65b9\u7a0b,\u8ba8\u8bba\u6bcf\u4e00\u5f20\u8868\u683c,\u5e76\u4f7f\u7528\u6070\u597d 10 \u4e2a\u5c0f\u8282\u5448\u73b0\u3002&#8221;<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>New style:<\/strong> Define the objective (&#8220;identify the paper&#8217;s genuine scientific delta, not a summary&#8221;), define the audience (&#8220;technically sophisticated but not necessarily domain-expert&#8221;), then request a structured analytical pass: a blind-spot pass (what background concepts must be understood, which claims are genuinely new vs. reframed known facts, what evidence would falsify the central interpretation), followed by an output that covers the prior bottleneck, the new mechanism, why it matters, the strongest evidence, and the limitations.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>\u65b0\u5f0f\u5199\u6cd5:<\/strong> \u5148\u5b9a\u4e49\u76ee\u6807(&#8220;\u627e\u51fa\u8bba\u6587\u771f\u6b63\u7684\u79d1\u5b66\u589e\u91cf,\u800c\u4e0d\u662f\u505a\u6458\u8981&#8221;),\u518d\u5b9a\u4e49\u53d7\u4f17(&#8220;\u6280\u672f\u4e0a\u8f83\u4e3a\u6210\u719f,\u4f46\u672a\u5fc5\u662f\u8be5\u7ec6\u5206\u9886\u57df\u7684\u4e13\u5bb6&#8221;),\u7136\u540e\u8981\u6c42\u8fdb\u884c\u7ed3\u6784\u5316\u7684\u5206\u6790:\u5148\u505a\u4e00\u6b21&#8221;\u76f2\u70b9\u6392\u67e5&#8221;(\u9700\u8981\u7406\u89e3\u54ea\u4e9b\u80cc\u666f\u6982\u5ff5\u3001\u54ea\u4e9b\u4e3b\u5f20\u662f\u771f\u6b63\u65b0\u9896\u7684\u3001\u54ea\u4e9b\u53ea\u662f\u5df2\u77e5\u4e8b\u5b9e\u7684\u91cd\u65b0\u5305\u88c5\u3001\u4ec0\u4e48\u6837\u7684\u8bc1\u636e\u80fd\u8bc1\u4f2a\u5176\u6838\u5fc3\u8bba\u70b9),\u63a5\u7740\u8f93\u51fa\u5185\u5bb9\u5e94\u6db5\u76d6:\u6b64\u524d\u7684\u79d1\u5b66\u74f6\u9888\u3001\u65b0\u673a\u5236\u3001\u5176\u91cd\u8981\u6027\u3001\u6700\u6709\u529b\u7684\u8bc1\u636e,\u4ee5\u53ca\u5c40\u9650\u6027\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">Configuring tool interfaces for the new model generation \/ \u4e3a\u65b0\u4e00\u4ee3\u6a21\u578b\u914d\u7f6e\u5de5\u5177\u63a5\u53e3<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">Rather than relying on in-context example-heavy tool-call demonstrations, Anthropic recommends designing:<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u4e0e\u5176\u4f9d\u8d56\u5927\u91cf\u793a\u4f8b\u6765\u6f14\u793a\u5de5\u5177\u8c03\u7528,Anthropic \u5efa\u8bae\u5728\u8bbe\u8ba1\u5de5\u5177\u63a5\u53e3\u65f6\u505a\u5230:<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>Meaningful, self-explanatory function\/parameter names \/ \u6709\u610f\u4e49\u3001\u81ea\u89e3\u91ca\u7684\u51fd\u6570\u4e0e\u53c2\u6570\u547d\u540d<\/li>\n\n\n\n<li>Clear constraint enumeration and explicit state transitions \/ \u6e05\u6670\u7684\u7ea6\u675f\u6761\u4ef6\u679a\u4e3e\u4e0e\u660e\u786e\u7684\u72b6\u6001\u8f6c\u6362<\/li>\n\n\n\n<li>Informative tool results \/ \u4fe1\u606f\u91cf\u5145\u5206\u7684\u5de5\u5177\u8fd4\u56de\u7ed3\u679c<\/li>\n\n\n\n<li>Tools that expose the right degree of freedom \u2014 not too rigid, not too open-ended \/ \u63d0\u4f9b\u6070\u5f53\u81ea\u7531\u5ea6\u7684\u5de5\u5177\u2014\u2014\u65e2\u4e0d\u8fc7\u4e8e\u6b7b\u677f,\u4e5f\u4e0d\u8fc7\u4e8e\u5f00\u653e<\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\">Examples are still useful, but now primarily to communicate taste and output form rather than to teach step-by-step imitation.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u793a\u4f8b\u4f9d\u7136\u6709\u7528,\u4f46\u73b0\u5728\u4e3b\u8981\u7528\u4e8e\u4f20\u8fbe&#8221;\u54c1\u5473\u504f\u597d&#8221;\u4e0e\u8f93\u51fa\u5f62\u5f0f,\u800c\u4e0d\u662f\u7528\u6765\u624b\u628a\u624b\u5730\u6559\u6a21\u578b\u9010\u6b65\u6a21\u4eff\u3002<\/p>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<h2 class=\"wp-block-heading\">Conclusion and Key Takeaways \/ \u7ed3\u8bba\u4e0e\u6838\u5fc3\u8981\u70b9<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">The presenter is candid that this is his own interpretation of two Anthropic documents, not a settled consensus, and that he remains skeptical of fully trusting a black-box agent \u2014 especially for specialized domains like theoretical physics or finance where hallucination is unacceptable. Several open questions are flagged as undisclosed by Anthropic: whether the serving stack uses hidden planner models, neural context compressors, model routers, separate verifier models, or prompt-rewriting layers.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u89c6\u9891\u4f5c\u8005\u5766\u8a00,\u8fd9\u53ea\u662f\u4ed6\u4e2a\u4eba\u5bf9 Anthropic \u4e24\u4efd\u6587\u4ef6\u7684\u89e3\u8bfb,\u5e76\u975e\u5df2\u6210\u5b9a\u8bba\u7684\u5171\u8bc6,\u4ed6\u672c\u4eba\u4e5f\u5bf9&#8221;\u5b8c\u5168\u4fe1\u4efb\u4e00\u4e2a\u9ed1\u7bb1\u667a\u80fd\u4f53&#8221;\u6301\u4fdd\u7559\u6001\u5ea6\u2014\u2014\u5c24\u5176\u662f\u5728\u7406\u8bba\u7269\u7406\u3001\u91d1\u878d\u7b49\u5bf9&#8221;\u5e7b\u89c9&#8221;\u96f6\u5bb9\u5fcd\u7684\u4e13\u4e1a\u9886\u57df\u3002\u4ed6\u4e5f\u6307\u51fa\u4e86 Anthropic \u5c1a\u672a\u516c\u5f00\u7684\u51e0\u4e2a\u95ee\u9898:\u670d\u52a1\u7aef(serving stack)\u662f\u5426\u4f7f\u7528\u4e86\u9690\u85cf\u7684\u89c4\u5212\u6a21\u578b(planner models)\u3001\u795e\u7ecf\u4e0a\u4e0b\u6587\u538b\u7f29\u5668\u3001\u6a21\u578b\u8def\u7531\u5668\u3001\u72ec\u7acb\u7684\u9a8c\u8bc1\u6a21\u578b,\u6216\u63d0\u793a\u8bcd\u91cd\u5199\u5c42\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">Key takeaways \/ \u6838\u5fc3\u8981\u70b9:<\/h3>\n\n\n\n<ol class=\"wp-block-list\">\n<li><strong>Specification, not the model, is now the bottleneck.<\/strong> Write intent, environment, constraints, and evidence \u2014 not exhaustive rule lists.<br><strong>\u74f6\u9888\u5df2\u4ece\u6a21\u578b\u8f6c\u79fb\u5230\u89c4\u683c\u8bf4\u660e\u3002<\/strong> \u5e94\u5199\u6e05\u695a\u610f\u56fe\u3001\u73af\u5883\u3001\u7ea6\u675f\u6761\u4ef6\u4e0e\u9a8c\u8bc1\u8bc1\u636e,\u800c\u4e0d\u662f\u7a77\u4e3e\u5f0f\u7684\u89c4\u5219\u6e05\u5355\u3002<\/li>\n\n\n\n<li><strong>Prompting is bi-directional now.<\/strong> Let the agent ask prioritized clarifying questions instead of guessing.<br><strong>\u63d0\u793a\u73b0\u5728\u662f\u53cc\u5411\u7684\u3002<\/strong> \u8ba9\u667a\u80fd\u4f53\u6309\u4f18\u5148\u7ea7\u63d0\u51fa\u6f84\u6e05\u95ee\u9898,\u800c\u4e0d\u662f\u9760\u731c\u6d4b\u3002<\/li>\n\n\n\n<li><strong>Adopt progressive disclosure.<\/strong> Keep the system prompt\/<code>CLAUDE.md<\/code> lean; push detail into selectively loaded skills.<br><strong>\u91c7\u7528\u6e10\u8fdb\u5f0f\u62ab\u9732\u3002<\/strong> \u4fdd\u6301\u7cfb\u7edf\u63d0\u793a\u8bcd\u4e0e <code>CLAUDE.md<\/code> \u7cbe\u7b80,\u628a\u7ec6\u8282\u4e0b\u653e\u5230\u6309\u9700\u52a0\u8f7d\u7684\u6280\u80fd\u6a21\u5757\u4e2d\u3002<\/li>\n\n\n\n<li><strong>Use structural references, not exhaustive spec lists<\/strong>, when the desired output is stylistic or hard to fully enumerate.<br><strong>\u5728\u8f93\u51fa\u504f\u5411\u98ce\u683c\u5316\u3001\u96be\u4ee5\u5b8c\u5168\u679a\u4e3e\u7684\u573a\u666f\u4e0b,\u4f7f\u7528\u7ed3\u6784\u6027\u53c2\u8003\u8303\u4f8b,\u800c\u4e0d\u662f\u7a77\u4e3e\u5f0f\u7684\u89c4\u683c\u6e05\u5355\u3002<\/strong><\/li>\n\n\n\n<li><strong>Treat execution as a learning loop<\/strong> \u2014 maintain an implementation-notes file that captures deviations, edge cases, and lessons.<br><strong>\u628a&#8221;\u6267\u884c\u8fc7\u7a0b&#8221;\u5f53\u4f5c\u4e00\u4e2a\u5b66\u4e60\u5faa\u73af<\/strong> \u2014\u2014\u7ef4\u62a4\u4e00\u4efd\u5b9e\u65bd\u7b14\u8bb0\u6587\u4ef6,\u8bb0\u5f55\u504f\u5dee\u3001\u8fb9\u7f18\u60c5\u51b5\u4e0e\u7ecf\u9a8c\u6559\u8bad\u3002<\/li>\n\n\n\n<li><strong>Separate your memory layers<\/strong>: project instructions, task artifacts, long-term validated lessons, and current execution state.<br><strong>\u533a\u5206\u4e0d\u540c\u7684\u8bb0\u5fc6\u5c42\u7ea7:<\/strong> \u9879\u76ee\u8bf4\u660e\u3001\u4efb\u52a1\u4ea7\u7269\u3001\u957f\u671f\u9a8c\u8bc1\u8fc7\u7684\u7ecf\u9a8c\u6559\u8bad,\u4ee5\u53ca\u5f53\u524d\u6267\u884c\u72b6\u6001\u3002<\/li>\n\n\n\n<li><strong>Remember Claude Code is a harness, not the model.<\/strong> Optimizing your workflow means optimizing the harness (context, memory, skills, tools) around whichever core model \u2014 Opus 5, Fable 5, or Mythos 5 \u2014 you&#8217;re using.<br><strong>\u7262\u8bb0 Claude Code \u662f\u4e00\u4e2a harness,\u800c\u975e\u6a21\u578b\u672c\u8eab\u3002<\/strong> \u4f18\u5316\u4f60\u7684\u5de5\u4f5c\u6d41,\u610f\u5473\u7740\u56f4\u7ed5\u4f60\u6240\u4f7f\u7528\u7684\u6838\u5fc3\u6a21\u578b(Opus 5\u3001Fable 5 \u6216 Mythos 5)\u53bb\u4f18\u5316 harness \u672c\u8eab(\u4e0a\u4e0b\u6587\u3001\u8bb0\u5fc6\u3001\u6280\u80fd\u3001\u5de5\u5177)\u3002<\/li>\n\n\n\n<li><strong>Validate independently, especially in specialized domains.<\/strong> These are general recommendations from Anthropic \u2014 not universal law for every field.<br><strong>\u5728\u4e13\u4e1a\u9886\u57df\u4e2d\u8981\u72ec\u7acb\u9a8c\u8bc1\u3002<\/strong> \u8fd9\u4e9b\u662f Anthropic \u7ed9\u51fa\u7684\u901a\u7528\u5efa\u8bae,\u5e76\u975e\u9002\u7528\u4e8e\u6bcf\u4e2a\u9886\u57df\u7684\u666e\u9002\u6cd5\u5219\u3002<\/li>\n<\/ol>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<h2 class=\"wp-block-heading\">Related References \/ \u76f8\u5173\u53c2\u8003\u8d44\u6599<\/h2>\n\n\n\n<ul class=\"wp-block-list\">\n<li>Anthropic Engineering, <em>Claude Code: Best practices for agentic coding<\/em> \u2014 foundational best-practices post on CLAUDE.md, context gathering, and harness customization: <a href=\"https:\/\/www.anthropic.com\/engineering\/claude-code-best-practices\">https:\/\/www.anthropic.com\/engineering\/claude-code-best-practices<\/a><\/li>\n\n\n\n<li>Community research on progressive-disclosure skill architecture for Claude Code (<code>SKILL.md<\/code> format, ~100-token frontmatter, on-demand loading): via Glasp-indexed source \u2014 <a href=\"https:\/\/glama.ai\/mcp\/servers\/@bobmatnyc\/mcp-skills\/blob\/888755c2756ea666a44b4301a6410eb66943fd91\/docs\/research\/progressive-skills-format-2025-11-29.md\">https:\/\/glama.ai\/mcp\/servers\/@bobmatnyc\/mcp-skills\/blob\/888755c2756ea666a44b4301a6410eb66943fd91\/docs\/research\/progressive-skills-format-2025-11-29.md<\/a><\/li>\n\n\n\n<li><em>Stop Using Claude Code the Normal Way<\/em> \u2014 case study on long-running, largely unsupervised Claude Code agents building full projects from a single prompt, illustrating the harness-design stakes discussed in this video: <a href=\"https:\/\/glasp.co\/youtube\/lGWFlpffWk4\">https:\/\/glasp.co\/youtube\/lGWFlpffWk4<\/a><\/li>\n\n\n\n<li><em>Claude Skills Explained: 4 Skills to 10x Your Coding Workflow<\/em> (referenced via &#8220;How I Use 10 AI Agents in Claude Code to Build Features Faster&#8221;) \u2014 practical walkthrough of skill-based workflow design: <a href=\"https:\/\/glasp.co\/youtube\/c68Rqn8PriE\">https:\/\/glasp.co\/youtube\/c68Rqn8PriE<\/a><\/li>\n<\/ul>\n","protected":false},"excerpt":{"rendered":"<p>Claude 5 shifts the bottleneck from model capability to human specification. Anthropic cut 80% of Claude Code&#8217;s system prompt, embracing intent-based prompting, bi-directional clarification, progressive disclosure via skills, and layered memory \u2014 turning execution into a self-learning loop rather than rigid rule-following.<\/p>\n<p>Claude 5 \u5c06\u74f6\u9888\u4ece\u6a21\u578b\u80fd\u529b\u8f6c\u79fb\u5230\u4eba\u7c7b\u7684\u89c4\u683c\u8bf4\u660e\u3002Anthropic \u524a\u51cf\u4e86 Claude Code \u7cfb\u7edf\u63d0\u793a\u8bcd\u7684 80%,\u8f6c\u5411\u610f\u56fe\u5bfc\u5411\u7684\u63d0\u793a\u65b9\u5f0f\u3001\u53cc\u5411\u6f84\u6e05\u5f0f\u63d0\u95ee\u3001\u57fa\u4e8e\u6280\u80fd\u7684\u6e10\u8fdb\u5f0f\u62ab\u9732\u4ee5\u53ca\u5206\u5c42\u8bb0\u5fc6\u673a\u5236\u2014\u2014\u8ba9\u6267\u884c\u8fc7\u7a0b\u6210\u4e3a\u81ea\u6211\u5b66\u4e60\u7684\u5faa\u73af,\u800c\u975e\u50f5\u5316\u7684\u89c4\u5219\u9075\u5faa\u3002<\/p>\n","protected":false},"author":1,"featured_media":8548,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[15,31,13,7],"tags":[],"class_list":["post-8547","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-ai","category-chinese-","category-quantum-and-u","category-quantum-mindset-programme"],"aioseo_notices":[],"featured_image_src":"https:\/\/meta-quantum.today\/wp-content\/uploads\/2026\/07\/Claude-5-Context-Engineering-00-scaled.jpg","featured_image_src_square":"https:\/\/meta-quantum.today\/wp-content\/uploads\/2026\/07\/Claude-5-Context-Engineering-00-scaled.jpg","author_info":{"display_name":"coffee","author_link":"https:\/\/meta-quantum.today\/?author=1"},"rbea_author_info":{"display_name":"coffee","author_link":"https:\/\/meta-quantum.today\/?author=1"},"rbea_excerpt_info":"Claude 5 shifts the bottleneck from model capability to human specification. Anthropic cut 80% of Claude Code's system prompt, embracing intent-based prompting, bi-directional clarification, progressive disclosure via skills, and layered memory \u2014 turning execution into a self-learning loop rather than rigid rule-following.\n\nClaude 5 \u5c06\u74f6\u9888\u4ece\u6a21\u578b\u80fd\u529b\u8f6c\u79fb\u5230\u4eba\u7c7b\u7684\u89c4\u683c\u8bf4\u660e\u3002Anthropic \u524a\u51cf\u4e86 Claude Code \u7cfb\u7edf\u63d0\u793a\u8bcd\u7684 80%,\u8f6c\u5411\u610f\u56fe\u5bfc\u5411\u7684\u63d0\u793a\u65b9\u5f0f\u3001\u53cc\u5411\u6f84\u6e05\u5f0f\u63d0\u95ee\u3001\u57fa\u4e8e\u6280\u80fd\u7684\u6e10\u8fdb\u5f0f\u62ab\u9732\u4ee5\u53ca\u5206\u5c42\u8bb0\u5fc6\u673a\u5236\u2014\u2014\u8ba9\u6267\u884c\u8fc7\u7a0b\u6210\u4e3a\u81ea\u6211\u5b66\u4e60\u7684\u5faa\u73af,\u800c\u975e\u50f5\u5316\u7684\u89c4\u5219\u9075\u5faa\u3002","category_list":"<a href=\"https:\/\/meta-quantum.today\/?cat=15\" rel=\"category\">AI<\/a>, <a href=\"https:\/\/meta-quantum.today\/?cat=31\" rel=\"category\">Chinese\/\u4e2d\u6587<\/a>, <a href=\"https:\/\/meta-quantum.today\/?cat=13\" rel=\"category\">Quantum and U<\/a>, <a href=\"https:\/\/meta-quantum.today\/?cat=7\" rel=\"category\">Quantum Mindset Programme<\/a>","comments_num":"0 comments","_links":{"self":[{"href":"https:\/\/meta-quantum.today\/index.php?rest_route=\/wp\/v2\/posts\/8547","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/meta-quantum.today\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/meta-quantum.today\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/meta-quantum.today\/index.php?rest_route=\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/meta-quantum.today\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=8547"}],"version-history":[{"count":4,"href":"https:\/\/meta-quantum.today\/index.php?rest_route=\/wp\/v2\/posts\/8547\/revisions"}],"predecessor-version":[{"id":8553,"href":"https:\/\/meta-quantum.today\/index.php?rest_route=\/wp\/v2\/posts\/8547\/revisions\/8553"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/meta-quantum.today\/index.php?rest_route=\/wp\/v2\/media\/8548"}],"wp:attachment":[{"href":"https:\/\/meta-quantum.today\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=8547"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/meta-quantum.today\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=8547"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/meta-quantum.today\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=8547"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}