diff --git a/CLAUDE.md b/CLAUDE.md
new file mode 100644
index 0000000..b4c36af
--- /dev/null
+++ b/CLAUDE.md
@@ -0,0 +1,75 @@
+# CLAUDE.md
+
+This file provides guidance to Claude Code (claude.ai/code) when working with code in this repository.
+
+## Project Overview
+
+WeWrite is a WeChat public account (公众号) content generation AI skill. It automates the full workflow from trending topic discovery to draft box publishing. It works as both a Claude Code skill (via SKILL.md) and an OpenClaw-compatible skill (via `dist/openclaw/`).
+
+The core pipeline is defined in SKILL.md (Steps 1-8): environment check → topic selection → framework + material collection → writing → SEO/anti-AI verification → visual AI → formatting/publishing → wrap-up.
+
+## Commands
+
+```bash
+# Install dependencies
+pip install -r requirements.txt
+
+# Toolkit CLI
+python3 toolkit/cli.py preview article.md --theme sspai       # Preview as HTML
+python3 toolkit/cli.py publish article.md --cover cover.png --title "标题"  # Publish to WeChat
+python3 toolkit/cli.py gallery                                  # Browse all 16 themes
+python3 toolkit/cli.py themes                                   # List theme names
+python3 toolkit/cli.py image-post img1.jpg img2.jpg -t "标题"   # Image post (carousel)
+
+# Data collection scripts
+python3 scripts/fetch_hotspots.py --limit 20                   # Trending topics
+python3 scripts/seo_keywords.py --json "关键词1" "关键词2"      # SEO keyword analysis
+python3 scripts/fetch_stats.py <article_id>                     # WeChat article stats
+python3 scripts/humanness_score.py article.md --verbose         # AI detection scoring (11 checks)
+
+# Build OpenClaw-compatible skill (also runs in CI on push to main)
+python3 scripts/build_openclaw.py
+```
+
+No formal test suite exists. CI only rebuilds the OpenClaw version on push to main.
+
+## Architecture
+
+### Dual Nature: Skill + Toolkit
+
+- **As a skill** (SKILL.md): An agent-orchestrated 8-step pipeline. The LLM reads SKILL.md and executes steps, calling Python scripts as tools. Reference docs in `references/` are loaded on-demand by the agent at specific steps.
+- **As a standalone toolkit** (`toolkit/cli.py`): A Python CLI for Markdown→WeChat HTML conversion and publishing, usable independently of the skill.
+
+### Key Directories
+
+- `scripts/` — Data collection utilities (hotspots, SEO, stats) and build tools. Called by the agent during pipeline execution.
+- `toolkit/` — Markdown→WeChat HTML converter, theme engine, WeChat API client, image generation. The CLI entry point is `toolkit/cli.py`.
+- `personas/` — 5 YAML writing personality presets controlling tone, data presentation, emotional arc. Loaded in Step 4b.
+- `references/` — Agent-loaded docs (writing rules, frameworks, SEO, topic scoring). These are NOT code — they are instruction sets the LLM reads and follows.
+- `toolkit/themes/` — 16 YAML theme definitions. Parsed by `toolkit/theme.py`, applied as inline CSS by `toolkit/converter.py`.
+
+### Formatting Pipeline (toolkit)
+
+`converter.py` is the core: Markdown → HTML with inline styles + WeChat compatibility fixes (CJK spacing, bold punctuation, list→section conversion, external links→footnotes, dark mode attributes). WeChat strips `<style>` tags, so all CSS must be inlined. Themes are YAML files defining colors and base CSS; `theme.py` parses them, `converter.py` applies them.
+
+### OpenClaw Compatibility
+
+`scripts/build_openclaw.py` transforms SKILL.md for OpenClaw: replaces `{skill_dir}` with `{baseDir}`, renames tool references (WebSearch→web_search, etc.), copies referenced files. CI runs this on push to main and commits to `dist/openclaw/`.
+
+### Configuration Files
+
+- `config.yaml` (from `config.example.yaml`) — WeChat API credentials + image API key. Missing → graceful degradation (skip_publish, skip_image_gen).
+- `style.yaml` (from `style.example.yaml`) — User's writing profile (name, topics, tone, persona, theme). Auto-created via onboard flow on first run.
+- `writing-config.yaml` (from `writing-config.example.yaml`) — Writing parameters (sentence variance, idiom density, etc.). Optimized per-user via the "优化参数" auxiliary function in SKILL.md.
+
+All three are .gitignored — each user generates their own.
+
+### Graceful Degradation
+
+The pipeline never hard-fails. Missing config → skip_publish/skip_image_gen flags. Script failures → WebSearch or LLM fallback. Image gen fails → output prompts only. These flags are set in Step 1 and automatically respected by later steps.
+
+## Language & Conventions
+
+- All code is Python 3.11+. No type checking or linter configured.
+- Commit messages are in Chinese, format: `type: description` (e.g., `fix: ...`, `chore: ...`).
+- The project language (README, SKILL.md, comments, references) is Chinese.
diff --git a/README.md b/README.md
index 9f7cdc3..d5802f9 100644
--- a/README.md
+++ b/README.md
@@ -168,7 +168,7 @@ wewrite/
 ├── SKILL.md                  # 主管道（273行，Step 1-8）
 ├── config.example.yaml       # API 配置模板
 ├── style.example.yaml        # 风格配置模板
-├── writing-config.example.yaml # 写作参数模板（可用 optimize loop 调优）
+├── writing-config.example.yaml # 写作参数模板（说"优化参数"自动调优）
 ├── requirements.txt
 │
 ├── dist/openclaw/            # OpenClaw 兼容版（CI 自动构建）
@@ -179,8 +179,7 @@ wewrite/
 │   ├── fetch_stats.py          # 微信文章数据回填
 │   ├── build_playbook.py       # 从历史文章生成 Playbook
 │   ├── learn_edits.py          # 学习人工修改
-│   ├── humanness_score.py      # 文章"人味"打分器（客观 checklist + LLM 判官）
-│   ├── optimize_loop.py        # autoresearch 风格迭代优化框架
+│   ├── humanness_score.py      # 文章"人味"打分器（11 项检测 + 参数映射）
 │   └── build_openclaw.py       # SKILL.md → OpenClaw 格式转换
 │
 ├── toolkit/                  # Markdown → 微信工具链
@@ -235,19 +234,25 @@ Step 8  写入历史 → 回复用户（含编辑建议 + 飞轮提示）
 
 默认全自动。说"交互模式"可在选题/框架/配图处暂停确认。
 
-## 优化循环（实验性）
+## 写作参数优化
 
-借鉴 [autoresearch](https://github.com/karpathy/autoresearch) 的 change→score→keep/rollback 模式，WeWrite 提供写作参数自动调优框架：
+在对话中说「优化写作参数」或「优化参数」，Agent 会自动迭代调优你的 `writing-config.yaml`：
+
+1. 用当前参数写测试短文
+2. 用 `humanness_score.py` 打分（11 项检测，连续 0-1 分数）
+3. 找到最低分维度，调整对应参数
+4. 重复 N 轮（默认 3 轮）
+5. 保留得分最好的参数组合
 
 ```bash
-# 对一篇文章打分（客观 checklist + 主观 LLM 判官）
+# 独立打分（不需要 Agent）
 python3 scripts/humanness_score.py article.md --verbose
 
-# 迭代优化写作参数
-python3 scripts/optimize_loop.py --topic "AI Agent" --iterations 10
+# JSON 输出（含每项分数 + 参数映射）
+python3 scripts/humanness_score.py article.md --json
 ```
 
-框架开源，但优化后的 `writing-config.yaml` 不入 git——每个用户跑出自己的最优参数。
+优化后的 `writing-config.yaml` 不入 git——每个用户跑出自己的最优参数。
 
 ## Toolkit 独立使用
 
diff --git a/scripts/diagnose.py b/scripts/diagnose.py
index 77d48a0..86b8dd8 100644
--- a/scripts/diagnose.py
+++ b/scripts/diagnose.py
@@ -157,7 +157,7 @@ def check_enhancements():
     else:
         checks.append(make_check(
             "enhancement", "writing_config", "warn",
-            "not found → using defaults (run optimize_loop.py to tune)",
+            "not found → using defaults (say '优化参数' to tune)",
         ))
 
     # playbook.md
@@ -240,7 +240,7 @@ def compute_summary(checks):
         elif name == "playbook":
             recs.append('Edit a generated article, then say "学习我的修改" to build playbook.md')
         elif name == "writing_config":
-            recs.append('Run: python3 scripts/optimize_loop.py --topic "your topic" --iterations 10')
+            recs.append('Say "优化参数" to run the optimization loop')
         elif name == "history_articles":
             recs.append("Generate your first article to start building history")
         elif name == "dimension_variance":
diff --git a/scripts/optimize_loop.py b/scripts/optimize_loop.py
deleted file mode 100644
index e6f7600..0000000
--- a/scripts/optimize_loop.py
+++ /dev/null
@@ -1,149 +0,0 @@
-#!/usr/bin/env python3
-"""
-WeWrite Optimization Loop — autoresearch-style iterative improvement.
-
-Inspired by Karpathy's autoresearch: change → score → keep/rollback → repeat.
-But instead of optimizing ML training code, we optimize WRITING RULES to
-produce articles that pass AI detection while maintaining quality.
-
-The mutable surface: writing-config.yaml (style parameters + prompt rules)
-The fixed evaluation: humanness_score.py (objective checklist + subjective feel)
-The metric: composite_score (lower = more human, like val_bpb)
-
-Usage:
-    python3 optimize_loop.py --topic "AI Agent" --iterations 10
-    python3 optimize_loop.py --topic "AI Agent" --iterations 5 --verbose
-
-Architecture:
-    1. Load current writing-config.yaml
-    2. Generate article with current config
-    3. Score with humanness_score.py
-    4. LLM proposes a change to writing-config.yaml
-    5. Generate article with new config
-    6. Score again
-    7. If improved → keep (commit). If not → rollback.
-    8. Log to results.tsv
-    9. Repeat.
-
-Requirements:
-    - ANTHROPIC_API_KEY in environment (for article generation + LLM judge)
-    - writing-config.yaml in skill root (created on first run with defaults)
-"""
-
-import argparse
-import json
-import os
-import subprocess
-import sys
-from datetime import datetime
-from pathlib import Path
-
-import yaml
-
-SKILL_DIR = Path(__file__).parent.parent
-CONFIG_PATH = SKILL_DIR / "writing-config.yaml"
-RESULTS_PATH = SKILL_DIR / "optimization-results.tsv"
-
-DEFAULT_CONFIG = {
-    "persona": "科技媒体资深编辑，写了八年公众号，对AI行业有深度认知",
-    "sentence_variance": 0.7,
-    "broken_sentence_rate": 0.04,
-    "idiom_density": 0.15,
-    "filler_style": "mixed",  # literary / casual / mixed / minimal
-    "paragraph_rhythm": "chaotic",  # structured / chaotic / wave
-    "self_correction_rate": 0.02,
-    "tangent_frequency": "every_800_chars",  # never / every_500 / every_800 / every_1200
-    "real_data_density": "high",  # low / medium / high
-    "word_temperature_bias": "warm",  # cold / warm / hot / balanced
-    "emotional_arc": "restrained_to_burst",  # flat / gradual / restrained_to_burst / volatile
-    "opening_style": "scene",  # scene / data / question / anecdote / cold_open
-    "closing_style": "open_question",  # summary / open_question / image / abrupt
-    "structure_linearity": 0.3,  # 0=fully non-linear, 1=fully linear
-}
-
-
-def ensure_config():
-    """Create default writing-config.yaml if it doesn't exist."""
-    if not CONFIG_PATH.exists():
-        with open(CONFIG_PATH, "w", encoding="utf-8") as f:
-            yaml.dump(DEFAULT_CONFIG, f, allow_unicode=True, default_flow_style=False)
-        print(f"Created default config: {CONFIG_PATH}")
-    return yaml.safe_load(CONFIG_PATH.read_text(encoding="utf-8"))
-
-
-def score_article(article_path: str) -> dict:
-    """Run humanness_score.py on an article. Returns parsed result."""
-    result = subprocess.run(
-        ["python3", str(SKILL_DIR / "scripts" / "humanness_score.py"), article_path, "--json"],
-        capture_output=True, text=True
-    )
-    if result.returncode != 0:
-        print(f"Scoring failed: {result.stderr}", file=sys.stderr)
-        return {"composite_score": 100.0, "error": result.stderr}
-    return json.loads(result.stdout)
-
-
-def log_result(iteration: int, composite: float, config_summary: str, status: str, description: str):
-    """Append result to TSV log."""
-    header_needed = not RESULTS_PATH.exists()
-    with open(RESULTS_PATH, "a", encoding="utf-8") as f:
-        if header_needed:
-            f.write("iteration\ttimestamp\tcomposite\tstatus\tdescription\tconfig_change\n")
-        ts = datetime.now().strftime("%Y-%m-%d %H:%M:%S")
-        f.write(f"{iteration}\t{ts}\t{composite:.2f}\t{status}\t{description}\t{config_summary}\n")
-
-
-def print_banner(iteration: int, total: int):
-    print(f"\n{'='*60}")
-    print(f"  OPTIMIZATION LOOP — Iteration {iteration}/{total}")
-    print(f"{'='*60}")
-
-
-def main():
-    parser = argparse.ArgumentParser(description="WeWrite optimization loop")
-    parser.add_argument("--topic", required=True, help="Article topic for testing")
-    parser.add_argument("--iterations", type=int, default=10, help="Number of iterations")
-    parser.add_argument("--verbose", "-v", action="store_true")
-    args = parser.parse_args()
-
-    print(f"""
-╔══════════════════════════════════════════════════════╗
-║  WeWrite Optimization Loop                          ║
-║  Topic: {args.topic:<44s}║
-║  Iterations: {args.iterations:<39d}║
-║                                                      ║
-║  Pattern: change config → generate → score →         ║
-║           keep if better, rollback if worse           ║
-╚══════════════════════════════════════════════════════╝
-""")
-
-    config = ensure_config()
-
-    print("This script provides the FRAMEWORK for optimization.")
-    print("To run the full loop, you need:")
-    print("  1. An article generation function (Claude API)")
-    print("  2. A scoring function (humanness_score.py — included)")
-    print("  3. An LLM to propose config changes (Claude API)")
-    print()
-    print("Current config:")
-    print(yaml.dump(config, allow_unicode=True, default_flow_style=False))
-    print()
-    print("Run this loop via Claude Code / OpenClaw agent:")
-    print()
-    print("  Agent reads writing-config.yaml")
-    print("  → generates article with those rules")
-    print("  → scores with: python3 scripts/humanness_score.py article.md --json")
-    print("  → proposes a config change")
-    print("  → generates new article")
-    print("  → scores again")
-    print("  → if composite_score decreased → commit config change")
-    print("  → if composite_score same/worse → rollback")
-    print("  → logs to optimization-results.tsv")
-    print("  → repeats")
-    print()
-    print("To test scoring on an existing article:")
-    print(f"  python3 scripts/humanness_score.py <article.md> --verbose")
-
-
-if __name__ == "__main__":
-    main()
diff --git a/writing-config.example.yaml b/writing-config.example.yaml
index e9c1bb5..a2d7a6a 100644
--- a/writing-config.example.yaml
+++ b/writing-config.example.yaml
@@ -1,10 +1,9 @@
 # WeWrite 写作参数（可优化）
-# 复制为 writing-config.yaml，然后用 optimize loop 迭代调优
-# 或手动调整后观察朱雀检测结果
+# 复制为 writing-config.yaml，在对话中说"优化参数"让 Agent 迭代调优
+# 或手动调整后用 humanness_score.py 评估
 #
 # 这个文件是起点，不是最优解。
-# 运行: python3 scripts/optimize_loop.py --topic "你的主题" --iterations 10
-# 每次迭代会修改 writing-config.yaml 中的参数，保留得分更好的版本。
+# 在对话中说"优化参数"即可自动调优，每轮调整得分最低的参数。
 #
 # 参数分三层，对应 writing-guide.md 的反检测结构。