first commit

文档润色流和知识库构建流/claude-scholar/skills/planning-with-files/reference.md

@@ -0,0 +1,110 @@
+# Reference: Manus Context Engineering Principles
+
+This skill is based on the context engineering principles from Manus, the AI agent company acquired by Meta for $2 billion in December 2025.
+
+## The 6 Manus Principles
+
+### 1. Filesystem as External Memory
+
+> "Markdown is my 'working memory' on disk."
+
+**Problem:** Context windows have limits. Stuffing everything in context degrades performance and increases costs.
+
+**Solution:** Treat the filesystem as unlimited memory:
+- Store large content in files
+- Keep only paths in context
+- Agent can "look up" information when needed
+- Compression must be REVERSIBLE
+
+### 2. Attention Manipulation Through Repetition
+
+**Problem:** After ~50 tool calls, models forget original goals ("lost in the middle" effect).
+
+**Solution:** Keep a `task_plan.md` file that gets RE-READ throughout execution:
+```
+Start of context: [Original goal - far away, forgotten]
+...many tool calls...
+End of context: [Recently read task_plan.md - gets ATTENTION!]
+```
+
+By reading the plan file before each decision, goals appear in the attention window.
+
+### 3. Keep Failure Traces
+
+> "Error recovery is one of the clearest signals of TRUE agentic behavior."
+
+**Problem:** Instinct says hide errors, retry silently. This wastes tokens and loses learning.
+
+**Solution:** KEEP failed actions in the plan file:
+```markdown
+## Errors Encountered
+- [2025-01-03] FileNotFoundError: config.json not found → Created default config
+- [2025-01-03] API timeout → Retried with exponential backoff, succeeded
+```
+
+The model updates its internal understanding when seeing failures.
+
+### 4. Avoid Few-Shot Overfitting
+
+> "Uniformity breeds fragility."
+
+**Problem:** Repetitive action-observation pairs cause drift and hallucination.
+
+**Solution:** Introduce controlled variation:
+- Vary phrasings slightly
+- Don't copy-paste patterns blindly
+- Recalibrate on repetitive tasks
+
+### 5. Stable Prefixes for Cache Optimization
+
+**Problem:** Agents are input-heavy (100:1 ratio). Every token costs money.
+
+**Solution:** Structure for cache hits:
+- Put static content FIRST
+- Append-only context (never modify history)
+- Consistent serialization
+
+### 6. Append-Only Context
+
+**Problem:** Modifying previous messages invalidates KV-cache.
+
+**Solution:** NEVER modify previous messages. Always append new information.
+
+## The Agent Loop
+
+Manus operates in a continuous loop:
+
+```
+1. Analyze → 2. Think → 3. Select Tool → 4. Execute → 5. Observe → 6. Iterate → 7. Deliver
+```
+
+### File Operations in the Loop:
+
+| Operation | When to Use |
+|-----------|-------------|
+| `write` | New files or complete rewrites |
+| `append` | Adding sections incrementally |
+| `edit` | Updating specific parts (checkboxes, status) |
+| `read` | Reviewing before decisions |
+
+## Manus Statistics
+
+| Metric | Value |
+|--------|-------|
+| Average tool calls per task | ~50 |
+| Input-to-output ratio | 100:1 |
+| Acquisition price | $2 billion |
+| Time to $100M revenue | 8 months |
+
+## Key Quotes
+
+> "If the model improvement is the rising tide, we want Manus to be the boat, not the piling stuck on the seafloor."
+
+> "For complex tasks, I save notes, code, and findings to files so I can reference them as I work."
+
+> "I used file.edit to update checkboxes in my plan as I progressed, rather than rewriting the whole file."
+
+## Source
+
+Based on Manus's official context engineering documentation:
+https://manus.im/de/blog/Context-Engineering-for-AI-Agents-Lessons-from-Building-Manus