Anthropic流のClaude活用法がリークから判明

How to use Claude like Anthropic does (they leaked the blueprint) Anthropic accidentally published the entire source code of Claude Code this morning. All 512,000 lines of it, sitting in a single debugging file that was never supposed to leave the building. By afternoon, the code was copied across the internet. And buried inside, someone found a fully functional virtual pet system. 18 species. Shiny variants. Rarity tiers. That's who builds your AI tools. FIVE KEY LESSONS FROM THE LEAK: I. CONSTRAINT-FIRST PROMPTING When Anthropic gives Claude instructions internally, every interaction is structured around three layers: - What tools to use: "Read this file using the file reader. Do not run any other commands." - What risks to flag: "If this action would delete data, stop and confirm first." - What the output should look like: "Give the conclusion first. Then explain your reasoning." That's it. Three layers. Tool, risk, format. Loose: "Write me a LinkedIn post about our product launch." Constrained: "Write a LinkedIn post about our SaaS product launch. 150 words max. Open with a specific customer result. No hashtags. No emojis. End with a question to drive comments." Unbounded prompts produce unbounded mediocrity. II. USE AN INDEX, NOT A DUMP Claude Code memory system: each note kept under 150 characters, entire memory file stays below 200 lines. It works like a table of contents, not a journal. When Claude needs more detail, it follows a pointer to a specific topic file. They call the problem: "Context Entropy." As a conversation grows longer, the AI loses track of what matters. Their solution: compress aggressively. Three levels of compression that kick in automatically. Most aggressive level: throws away everything except most recent files, active plans, and a compressed summary. About 35,000 words of headroom. Practical takeaway: stop pasting entire documents. Give compressed summaries. Short bullet points. One line per decision. III. STOP ASKING FOR EVERYTHING IN ONE CONVERSATION "Coordinator Mode": One Claude acts as manager, breaks task into pieces, assigns each to a separate worker with focused context, reads results and synthesizes. The coordinator prompt has a line: "Do NOT say 'based on your findings.' Read the actual findings and specify exactly what to do." You can use this without technical setup: separate conversations for research, planning, execution. Copy key findings (not entire conversation) from one session to the next. IV. CLAUDE LITERALLY DREAMS Background process: "autoDream." During idle time, it reviews everything learned about your project and cleans up memory. The instruction it receives: "You are performing a dream, a reflective pass over your memory files. Synthesize what you've learned recently into durable, well-organized memories so that future sessions can orient quickly." Four phases: scan what's known → check new info → merge/fix contradictions → prune outdated. You can do this manually: at the end of any long session, type: "Summarize this session. What decisions did we make? What's still unresolved? What context would a fresh conversation need to continue where we left off?" Save that output. Paste it at the start of your next session. V. THE PROTOCOL 1. Start every session with compressed context block. Under 150 characters per point. 2. Constraint-First Prompting: specify tool + risk + format for every request. 3. One task per prompt. Anthropic built 40+ individual tools, each for exactly one thing. 4. Plan before executing: "Don't write anything yet. First, outline your approach and let me review it." 5. Run a dream pass at end of long sessions. 6. Verify everything. Anthropic's internal testing shows their latest model gets things wrong 29-30% of the time. Nearly one in three claims has an error. That's why they built verification into every layer. The gap between knowing how to use AI and actually using it well is where all the edge lives.