autowriter（autowriter 搜索）

autowriter — Automated Writing System

autoresearch's core is an agent loop: modify code → run → evaluate → keep/discard → loop.

autowriter maps this paradigm to writing, embedding "de-AI" into the loop itself — not post-processing after writing, but writing, purifying, evaluating, rewriting in every single iteration.

Design Philosophy

Three core principles from autoresearch, mapped to writing:

Plus humanizer's core insight: De-AI is not post-processing polish, it's part of writing quality. The evaluation function detects AI patterns directly — rewrite if not passing, rewrite again, until clean.

Agent Loop (Core Flow)

┌─────────────────────────────────────────────────┐
│  INPUT: topic/paper/project + --depth N          │
│  (N=1 quick, N=2 standard, N=3 deep, N=4 survey) │
└──────────────────────┬──────────────────────────┘
                       ▼
              ┌────────────────┐
              │ Phase 0: Research│  Read user-provided sources
              └───────┬────────┘
                      ▼
         ┌──── Loop start (max N rounds) ───┐
         │                                  │
         │  ┌────────────────────┐          │
         │  │ Phase 1: Write     │          │
         │  │ Generate full draft│          │
         │  │ (built-in human    │          │
         │  │  feel constraints) │          │
         │  └────────┬───────────┘          │
         │           ▼                      │
         │  ┌────────────────────┐          │
         │  │ Phase 1.5: De-AI   │          │
         │  │ Scan + rewrite     │          │
         │  │ AI patterns        │          │
         │  └────────┬───────────┘          │
         │           ▼                      │
         │  ┌────────────────────┐          │
         │  │ Phase 2: Evaluate  │          │
         │  │ 6-dim quantitative │          │
         │  │ scoring            │          │
         │  └────────┬───────────┘          │
         │           ▼                      │
         │  ┌────────────────────┐          │
         │  │ Phase 3: Decide    │          │
         │  │ score>=80 → keep   │──────┐   │
         │  │ score<80 → rewrite │      │   │
         │  │ with annotations   │      │   │
         │  └────────────────────┘      │   │
         │                               │   │
         └───────────────────────────────┼───┘
                                         ▼
                                 ┌────────────────┐
                                 │ Output final   │
                                 │ article + log  │
                                 └────────────────┘

--depth Knob

One parameter controls everything. No other knobs exposed.

Phase 0: Research

Facts first, then write. autoresearch reads train.py first — same principle.

Actions (user supplies all source material — this skill does not make network requests):

1. Paper → Read the user-provided PDF/URL/clipboard text, extract core contribution, method, experimental data, limitations
2. Tech topic → Read the user-provided references, notes, or local files, extract key facts
3. Project → Read user-provided source/docs within the workspace, extract architecture, design decisions, key code

Output: research_facts.md in the current workspace directory — structured fact checklist (not an outline, not "what goes in paragraph 1")

Important: If the user has not provided source material, ask them to supply it. Do not search the web or access files outside the workspace.

Phase 1: Write / Rewrite

First round: Initial draft

• Write directly based on research_facts.md, don't overthink structure
• Write backwards: core discovery/code first, background later
• Allow bad writing — a draft is raw material for evaluation
• Built-in human feel constraints (see "Iron Rules"), but don't spend time polishing

Subsequent rounds: Targeted rewrite

• Carry self-evaluation annotations from previous round
• Only fix lowest-scoring dimensions, don't rewrite everything
• Each round must show substantial change

Phase 1.5: De-AI (Humanize Pass)

This is the key step where humanizer mechanism is embedded into the loop. Not post-processing, but a mandatory checkpoint in every iteration.

Execution

Run AI pattern scan on Phase 1 output, check and rewrite each item:

Scan checklist (fast scan, not line-by-line):

1. Filler phrases — Remove opening bromides and emphasis crutches

- Kill: "It's worth noting," "As we all know," "Obviously," "Undoubtedly," "In this era of X" - Kill: "To achieve this goal" → "To do this" - Kill: Rhetorical questions ("So the question becomes...")

1. Overemphasis — Check for exaggerated significance

- Kill: "marks," "witnesses," "crucial," "indelible" - Kill: "Not only... but also...," "This isn't just... it's..."

1. AI vocabulary blacklist — Replace with direct expression

- "Furthermore" → delete or use direct connection - "Delve into" → "analyze" / "look at" - "Demonstrates" → "shows" / delete - "Dynamic," "rich," "profound" → specific description or delete - "Ever-evolving landscape" → specific context

1. Structural patterns — Break formulas

- Rule of three → use two or four items instead - Bold heading + colon list → blend into paragraphs - Dash reveal → use direct statement - Generic positive ending → specific next step or limitation

1. Voice injection — Add human touch

- Have opinions, don't just report facts - Admit uncertainty ("I'm not sure," "Honestly") - Mix sentence lengths (Short. Then a longer one that unfolds.) - Allow tangents and half-formed thoughts

Speed control

De-AI scan must be fast. Not line-by-line proofreading, 5 minutes for a pass:

• Run blacklist keyword grep first (10 seconds)
• Then fix structural issues (2 minutes)
• Finally inject voice (2 minutes)

Don't pursue perfection. Phase 2's evaluation function catches residual AI traces — if it doesn't pass, next round will handle it.

Phase 2: Self-Evaluation

6-dimension quantitative evaluation function. Each dimension 0-100.

Human feel dimension scoring

Composite score formula

score = info_density*0.20 + code_data_ratio*0.20 + failure_showcase*0.15
      + conciseness*0.15 + actionability*0.15 + human_feel*0.15

Self-evaluation output format

## Self-Eval - Round N

| Dimension | Score | Notes |
|-----------|-------|-------|
| Information density | 75 | Paragraphs 3-4 too much setup |
| Code/data ratio | 60 | "Significant improvement" has no data |
| Failure showcase | 40 | Missing failed experiments |
| Conciseness | 70 | First two paragraphs removable |
| Actionability | 85 | Code examples clear |
| Human feel | 55 | "Furthermore" x3, rule-of-three x2, cliched ending |

**Composite: 64/100**
**Decision: REWRITE**
**Focus areas:**
1. Add failed experiments (failure 40 -> 70+)
2. Replace "significantly improved" with data (code/data 60 -> 75+)
3. Rewrite lowest human-feel paragraphs: kill "furthermore", change rule-of-three, new ending (human 55 -> 75+)

Phase 3: Decision

score >= 80  →  KEEP, proceed to output
score < 80   →  DISCARD, enter next round with annotations

Early termination

• Two consecutive rounds with score difference < 5 → stop, take the higher-scoring version
• Max iterations reached → stop, take the highest-scoring version

Draft Log

Append after each evaluation, equivalent to autoresearch's results.tsv:

| Round | Score | Human | Decision | Main Issues | Fix Actions |
|-------|-------|-------|----------|-------------|-------------|
| 1 | 52 | 40 | DISCARD | Heavy AI, no data | De-AI, add experiments |
| 2 | 71 | 68 | DISCARD | Rule-of-three remnant, no failures | Restructure, add detours |
| 3 | 83 | 80 | KEEP | - | - |

Draft log stays at the end of the article or as an attachment. Fully transparent, no secret recipe.

Storage limit: Only retain the current draft and the final draft_log summary. Discarded intermediate drafts are NOT saved to disk — only their scores and fix actions are recorded in the log table. This prevents accumulation of potentially sensitive content.

Iron Rules

Enforced on every write/rewrite. These rules fuse Karpathy style with humanizer principles:

Five Iron Rules

1. Show Don't Tell — Put code/data, not prose descriptions of effects
2. One thing per paragraph — Delete any paragraph and information is lost
3. Experiments first — No claims without code/data/search results backing them
4. Record failures — Every article must include at least one "what didn't work"
5. Zero filler — Kill all filler phrases, rhetorical questions, universal summary sentences

Language rules (built-in de-AI)

Use: First person, specific numbers, code snippets, colloquial tech language, admitting ignorance, mixed sentence lengths, opinionated reactions Don't use:

• "This article will introduce," "As we all know," "It's worth noting," "In this era of X"
• "Furthermore," "Delve into," "Demonstrates," "Dynamic," "Ever-evolving landscape"
• Rhetorical questions ("Why does this matter?" → just say why)
• Negative parallelism ("Not only... but also...")
• Rule-of-three lists (use two or four items)
• Bold heading + colon lists (blend into paragraphs)
• Generic positive endings ("What's exciting is..." → specific next step)

Voice injection

• Have opinions. "Honestly I think this direction is flawed" > "This direction has certain limitations"
• Admit complexity. "I tried three approaches, first two bombed" > "After multiple experimental validations"
• Allow tangents. Real thinking isn't linear.
• Mixed rhythm. Short sentences. Then a longer one that unfolds slowly, with a turn, and lands.

Article Structure Selection

Automatically chosen based on --depth and content type, not forced into templates:

depth 1-2 (concise output)

• Opening: one-sentence conclusion (result first)
• Core: code/data + the single most important finding
• Closing: limitation + one-sentence summary

depth 3-4 (deep output)

• Opening: one-sentence conclusion
• Background: why this matters (<=3 sentences)
• Body: minimal runnable example → expand step by step → experimental data
• Failure: what didn't work + why
• Closing: code links + limitations

Structure is a result, not a constraint.

Skill Integration

• agent-browser: If the user has already gathered research material via agent-browser, autowriter reads those results (workspace files only)
• WeChat article style guide: For WeChat publishing format requirements

No humanizer-zh post-processing needed. De-AI is built-in. This skill does not initiate network requests. All source material must be user-provided.

Further Reading

• autoresearch design philosophy → references/autoresearch-philosophy.md
• Karpathy code style → writing style mapping → references/karpathy-code-style.md