Want to remove the 'machine' feel from AI-written text? Many people only swap words.

Image Analysis

1. Overall interface and visible phenomena (facts)

- Top page title: sepia

- Top language switch links: English | 繁體中文

- The page shows a white content panel on a colorful desktop background (looks like macOS wallpaper).

- The content language is Traditional Chinese (article text in Traditional Chinese).

- There are title, subtitle, paragraphs and multiple tables (clear columns and rows).

- Three main paragraph headings are discernible:

- “為什麼還需要另一個 humanizer” (Why another humanizer is needed)

- Table column headers include “Pass”, “層次”, “例子”

- The text contains English words or proper nouns (e.g., Agent Skill, Claude Code, Codex, Grok Build, Antigravity, SKILL.md, write, review, refactor, recreate, StoryScope, research, Claude, GPT, Gemini, DeepSeek, Kimi, file:line).

- At least two tables appear: the first with columns “Pass / 層次 / 例子”, the second with “領域 / 要點”.

- The page lists 3 Pass levels (marked 1,2,3).

- The text cites data and references such as “Russell et. al., 2026: 61,608 stories”, “93.2% macro-F1”, “95.5%”, “93.9%”.

- The page mentions “another 30 diagnostic features (rubric)” and model fingerprint corrections (Claude, GPT, Gemini, DeepSeek, Kimi).

- The “領域 / 要點” table lists domains: release notes/announcements, PR/issue replies, incident reviews, tickets, technical articles (these domain labels are clearly readable).

- In the incident review row the note readable: “do not blame individuals, root system issues; attach timeline, record steps, each action has an owner” (approximate phrasing).

- There is a concluding paragraph near the page bottom (partly readable) with fragments like “sepia 每 ... 3–5 (or 3–5 techniques)”.

2. Line-by-line / block readable excerpts (facts, as precise as possible)

- Title: sepia

- Links: English | 繁體中文

- Opening paragraph (readable fragments):

- “Work from the layer that really makes AI ‘wash brains’. For fiction, fix narrative structure first then break down wording; professional documents (release notes, PR replies, incident reviews, tickets, technical articles) each use a venue-specific set of rules.”

- Next paragraph:

- “This is a portable Agent Skill, supporting Claude Code, Codex, Grok Build and Antigravity. A single canonical SKILL.md is shared across platforms, no separate platform branches.”

- “Four operations: write, review (diagnose only), refactor (minimal edits), recreate (full rewrite).”

- Heading: Why another humanizer is still needed

- Paragraph (readable fragments):

- “Common humanizers change diction and syntax. StoryScope (Russell et. al., 2026: 61,608 stories, covering humans and 5 top LLMs) shows that a classifier using only narrative-structure features can detect AI fiction with 93.2% macro-F1; removing sentence/word style drops classifier performance only from 95.5% to 93.9%. Remaining flaws are in the narrative layer: narrator stating theme directly, single-threaded and overly tidy causal plots, emotion shown only via bodily sensations, lack of real-world anchors, absent reader, linear time, and chapters ending in growth/acceptance.”

- Following paragraph (partly readable):

- “sepia turns these distance measures, together with a related paper gathered in research/, into three pass processes for fiction writing and revision:”

- Table header (readable): Pass / 層次 / 例子

- Table rows (readable points, row by row):

- Pass 1 level: Narrative structure (fiction)

- Example (readable fragments): “Stop explaining the theme, loosen the causal key, put scene later, mix emotion presentation modes, sparse character networks, name real-world objects” (some words may vary)

- Pass 2 level: Section/chapter progression

- Example (readable fragments): “Break paragraph—problem-sequence templates, fix sagging middle of story, vary pacing and placement” (some fragments uncertain)

- Pass 3 level: Sentence & diction style

- Example (readable fragments): “What all humanizers edit: pronouns..., syntactic templates, word choice, register” (some parts unreadable)

- Continued text (readable fragment):

- “Another 30 diagnostic features (rubric), and per-model fingerprint corrections (Claude, GPT, Gemini, DeepSeek, Kimi).”

- Next paragraph (readable fragment):

- “Professional documents follow different patterns. Research notes common problems: lack of persuasiveness; hedged tone where judgement is required; chatbot residue in voice; disregard for venue register; and layout that looks like the same template. Each document type has a checklist and a tailored modification suggestion.”

- Second table (columns readable): domain / key points

- Domains listed (readable): release notes / announcements, PR / issue replies, incident reviews, tickets, technical articles

- Incident review key points readable: “do not blame people, root system issues; attach timeline, record steps, each action item has an owner.”

- Other domain notes are visible but contain many unclear words (see unreadable section below).

- Bottom of page (concluding fragment, readable): mentions “sepia every ... 3–5 techniques, remainder omitted” (specific numbers/phrasing partly blurred).

3. Unreadable or uncertain words/fragments (explicitly listed)

Below are words or fragments I could not clearly read or are uncertain; locations are marked for review:

- In the opening sentence the phrase “從真正會讓 AI 洗腦的層次下手” includes the word “洗腦” which is somewhat uncertain; overall meaning is clear though minor wording may differ.

- The phrase “sepia 把這些差測距，連同 research/ 裡整理過的一篇相關研究，轉成...” contains “差測距” which is uncertain—could be “差別測度”, “差異測距” or similar.

- Specific wording inside table examples (e.g., Pass1’s “點名真實事物”, Pass2’s “拆掉段落—問題序列的模板”) contain unclear or compressed wording; sentence structure lost in places.

- Pass 3 row example’s opening fragment for sentence-style edits is incomplete (some words missing), preventing exact transcription.

- In the second table (domain/key points) the release notes / announcement row has several unclear words (e.g., “users should understand prerequisites, attach evidence for each important item, no marketing words” is approximate and partly guessed).

- The PR / issue replies row has line breaks and blurs; “cite file:line” is readable but the full sentence is not.

- Small-print lines at the bottom are blurred and cannot be fully extracted.

4. Inferences (reasonable context-based guesses, labeled as inference)

- Inference A: The page documents a tool or workflow named sepia that aims to improve “humanizer” methods for AI text by focusing on narrative structure and venue-specific rules for professional documents.

- Inference B: The three Passes in the table map to large-scale structure (narrative), mid-level chapter/section progression (pace/paragraphing), and surface-level sentence/style edits, each with actionable example fixes.

- Inference C: The cited numbers (61,608 stories; 93.2% macro-F1; 95.5%; 93.9%) are presented to support the claim that humanizers focusing only on diction/syntax miss structural cues detectable by narrative-structure classifiers.

- Inference D: The second table lists different professional document types (release notes, PR replies, incident reviews, tickets, tech docs) with domain-specific checklists; incident reviews emphasize chronology, process recording and assigning action owners (this point is relatively readable and thus higher-confidence inference).

- Inference E: The “another 30 diagnostic features (rubric)” likely refers to a set of finer-grained checks used by the system to detect or advise on structural issues in fiction and documents.

5. Other visible UI elements (facts)

- A scrollbar on the right indicates a long page.

- Narrow-column layout with side margins.

- Standard web fonts and table borders/shading between paragraphs.

If you need a full verbatim transcription (including uncertain fragments marked in brackets) or a structured plain-text version of the tables, I can produce a more fine-grained transcription with all uncertain words explicitly annotated based on this image evidence.

SkillPI Creation Analysis

Purpose: The material describes an open-source writing Skill named sepia that aims to humanize AI-generated text by changing underlying narrative structure rather than only surface wording. It targets both fiction and venue-specific professional documents (release notes, PR/issue replies, incident reviews, tickets, technical articles) and produces diagnostic output and model-specific correction lists.

Image evidence summary: The page header is sepia and offers language switch (English | 繁體中文). The UI shows multiple tables including a 3-pass workflow (Pass 1-3), mentions a research citation (Russell et al., 2026) with 61,608 stories and a 93.2% macro-F1 figure, references “another 30 diagnostic features (rubric)”, and lists per-model fingerprint corrections for Claude, GPT, Gemini, DeepSeek and Kimi. The page states four operation modes: write, review, refactor, recreate, and that it can be used as a portable Agent Skill across Claude Code, Codex, Grok Build and Antigravity with a single canonical SKILL.md.

Steps to use the Skill (based on visible workflow and terms; presented as an actionable checklist to prepare for trial):

1) Obtain sepia Skill/ repository (confirm SKILL.md and installation instructions) and collect target text samples (fiction or specific professional documents).

2) Run a review (diagnostic) using sepia to get the 30-item rubric output and identify structural issues; the tool should recommend which Passes to apply.

3) Choose an operation: review only to diagnose; refactor for minimal edits; recreate for full structural rewrite; or write to generate new text.

4) Apply Passes in order: Pass 1 (narrative structure) → Pass 2 (section/chapter progression) → Pass 3 (sentence & diction), following the rubric and model-specific corrections.

5) Re-run diagnostic checks and iterate until the desired reduction in detectable AI-structure signatures is achieved.

Reproduction prerequisites and verification items:

- Access to sepia source or Skill definition and installation instructions (SKILL.md).

- An Agent runtime or adapter for the supported platforms referenced (Claude Code, Codex, Grok Build, Antigravity) if you plan to run it as an Agent Skill.

- Sample texts for the intended domain and a test procedure to measure outcomes.

- To verify research claims (e.g., 93.2% macro-F1 from StoryScope/Russell et al.), the original dataset, evaluation splits and metric definitions must be obtained.

Items that need confirmation (explicitly flagged):

- The exact location and license of the sepia repository and how to install it.

- The full list and decision rules of the “30 diagnostic features (rubric).”

- The concrete per-model correction strategies and rewrite examples for Claude, GPT, Gemini, DeepSeek, Kimi.

- Reproducibility details for the cited research results and any comparative benchmarks versus other humanizer tools.

Caveat: The page itself includes a disclaimer that sepia is experimental and performance varies by text type. The above steps and interpretation are derived from readable page content and reasonable inference; they do not constitute a tested confirmation of sepia’s effectiveness.