CLAUDE.md git:20261014.d220237B

Immutable. This exact content never changes and is served at /api/v1/blob/8d3f037c86aaeda0.

# daftAI Skills — Project Guidelines

This file provides guidance to AI agents (Amp, Claude Code) when working with code in this repository.

## Project Overview

AI agent skills for video/subtitle processing and content utilities. All skills use `daftai-` prefix to avoid conflicts when users import this plugin.

## Architecture

Cloudflare project audit lives in `developer-tools` at `skills/daftai-cloudflare-billing-audit/`. Its scanner runs offline and reads only the explicit target; it never executes the target project. Optional Agent-led evidence collection may use existing official connections only after explicit, bounded user authorization, as defined in the skill; it does not log in, expand permissions, change account configuration, or create persistent subscriptions. `.github/workflows/cloudflare-billing-audit.yml` validates the native-platform and installed-skill contracts. `.gitattributes` keeps this skill LF-normalized for matching installed-artifact hashes across platforms.

The Cloudflare audit's optional `EXTEND.json` controls report language and the Cloudflare-check mode; its legacy writing field cannot disable the required report stage. Before review, an explicit current mode or saved mode skips the first-use card; otherwise ask once whether to inspect the project only or also query Cloudflare through MCP, including current-only and remembered choices. Only an explicit remembered choice saves preferences outside the audited target. Missing or unanswered preferences do not block project-only review. A saved mode is not account authorization. It never auto-reads preferences from the audited target. All reports require final writing review after evidence and facts are frozen. Chinese uses `daftai-chinese-technical-writing`; English uses the bundled ASD-STE100-inspired report-writing reference; other languages use the common structure and factual checks in their natural conventions. Guidance never selects the output language. Resolve the existing dependency internally or obtain only that skill from this public repository under host permissions. A missing dependency preserves audit evidence but blocks claiming completed final writing. Optional Cloudflare MCP setup is separate, user-authorized client configuration and user-completed official OAuth. A requested Cloudflare check actively offers missing connection/login setup and waits for completion or explicit skip; it does not silently finish project-only after discovery fails. Evidence queries do not authenticate themselves or widen permissions. Connected evidence bounds requests, pages and rows, with no default duration deadline. Explicit user duration limits and actual host/tool timeouts remain separate.

Skills are organized into plugin categories in `marketplace.json`:

```
skills/
├── [video-skills]                      # Video and subtitle processing
│   ├── daftai-video-compositor/            # Embed video into a screenshot aperture
│   ├── daftai-video-subtitler/             # Burn subtitles into videos (FFmpeg)
│   ├── daftai-subtitle-translator/         # Translate subtitle files between languages
│   └── daftai-extracting-video-subtitles/  # Extract subtitles from video via Whisper
│
├── [utility-skills]                    # Content utilities
    ├── daftai-url-to-markdown/             # Fetch URL and convert to markdown with images
    ├── daftai-chinese-copywriting/         # Normalize Chinese copywriting
    ├── daftai-chinese-technical-writing/   # Chinese technical, product, and UI writing
    ├── daftai-image-to-svg/                # Convert transparent or solid-background images into path-only SVG
    └── daftai-gpt-image-2-5-prompt-optimizer/ # Optimize image prompts using live official references
│
└── [developer-tools]
    └── daftai-cloudflare-billing-audit/     # Offline Cloudflare project billing diagnostics
```

**Plugin Categories**:

The GPT Image 2.5 prompt optimizer returns text only: agents compare a published layout under `references/gpt-image-2-5/` and a Bun script fetches the official 2.5 section only when that layout is missing. No image generation or source article is bundled.

| Category | Description |
|----------|-------------|
| `video-skills` | Skills that process video or subtitle files (burning, translating, extracting) |
| `utility-skills` | Helper tools for content processing (URL conversion, etc.) |
| `developer-tools` | Read-only Cloudflare project billing diagnostics |

Each skill contains:
- `SKILL.md` — YAML front matter (name, description) + workflow documentation
- `scripts/` — TypeScript implementations
- `references/` — Glossaries, config schemas, workflow details (optional)
- `agents/` — Product-specific skill metadata (optional)

`daftai-chinese-technical-writing` bundles six verbatim public-domain sections from ruanyf/document-style-guide. Product and UI extensions stay separate from those sources. It has no CLI or persistent preferences; requested formatting checks delegate to `daftai-chinese-copywriting` in read-only `review` mode unless the user asks for fixes.

## Running Scripts

All scripts are TypeScript, executed via `bun` (no build step).

The Cloudflare billing audit overrides the missing-runtime stop below: automatically obtain missing Bun and its required writer at user scope through its `references/tool-setup.md`, verify them, and continue the same audit. No separate setup question is needed. Explicit user restrictions and host permissions still apply; no target mutation or administrator escalation is allowed.

### Runtime Detection

Before running any script, the agent MUST detect the runtime **once per session**:

```bash
# Detect runtime (run once, reuse result)
if command -v bun &>/dev/null; then
  RUNTIME="bun"
else
  echo "Error: bun not found. Install Bun: https://bun.sh/"
  exit 1
fi
```

| Priority | Condition | Runtime | Notes |
|----------|-----------|---------|-------|
| 1 | `bun` available | `bun` | Standard execution via Bun |
| 2 | Not found | Error + install guide | Suggest installing Bun |

### Script Execution

```bash
bun skills/<skill>/scripts/<script>.ts [options]
```

Scripts use only Node.js built-in modules (`fs`, `path`, `child_process`, `os`). No npm packages required.

**CRITICAL**: When running scripts from a skill, set the working directory (`cwd`) to the skill's base directory (where SKILL.md is located). Do NOT run from the user's project directory.

## Key Dependencies

- **Bun**: TypeScript runtime for all scripts
- **FFmpeg**: Required by `daftai-video-subtitler` for subtitle burning and by `daftai-video-compositor` for screenshot composition
- **Potrace + librsvg/Inkscape**: Required by `daftai-image-to-svg` for path tracing and candidate measurement
- **Whisper**: Required by `daftai-extracting-video-subtitles` for speech-to-text
- **Chrome**: Required by `daftai-url-to-markdown` for CDP page rendering
- **No npm packages**: Self-contained TypeScript, no external dependencies

## Authentication

`daftai-url-to-markdown` uses Chrome CDP for browser automation:
- First run opens Chrome for page rendering
- Some pages may require user login before capture
- Supports "wait for user signal" mode for login-required pages

## Plugin Configuration

`.claude-plugin/marketplace.json` defines plugin metadata and skill paths. Version follows semver.

## Skill Loading Rules

**IMPORTANT**: When working in this project, follow these rules:

| Rule | Description |
|------|-------------|
| **Load project skills first** | Discover project skill metadata first; load only skills relevant to the current task and only at their applicable stage. Project skills take priority over system/user-level skills with the same name. |

**Loading Priority** (highest → lowest):
1. Current project `skills/` directory
2. User-level skills (`$HOME/.daftAI-skills/`)
3. System-level skills

## Release Process

**IMPORTANT**: When user requests release/发布/push, ALWAYS follow this checklist.

**Never skip**:
- `CHANGELOG.md` + `CHANGELOG.zh.md` — Both must be updated
- `.claude-plugin/marketplace.json` version bump
- `README.md` + `README.zh.md` if applicable
- All files committed together before tag

## Adding New Skills

**IMPORTANT**: All skills MUST use `daftai-` prefix to avoid conflicts when users import this plugin.

### Key Requirements

| Requirement | Details |
|-------------|---------|
| **Concise is key** | The agent is smart — only add context it doesn't have. Challenge each token. |
| **name field** | Max 64 chars, lowercase letters/numbers/hyphens only |
| **description field** | Max 1024 chars, non-empty, MUST be third person, include what + when to use |
| **SKILL.md body** | Keep under 500 lines; use separate files for additional content |
| **Naming convention** | `daftai-<name>` prefix required |
| **References** | Keep one level deep from SKILL.md; avoid nested references |
| **No time-sensitive info** | Avoid dates/versions that become outdated |

### Steps

1. Create `skills/daftai-<name>/SKILL.md` with YAML front matter
   - Directory name: `daftai-<name>`
   - SKILL.md `name` field: `daftai-<name>`
2. Add TypeScript scripts in `skills/daftai-<name>/scripts/`
3. Add reference files in `skills/daftai-<name>/references/` if needed
4. **Choose the appropriate category** and register in `marketplace.json`:
   - `video-skills`: For video/subtitle processing
   - `utility-skills`: For helper tools (conversion, etc.)
   - If none fit, create a new category with descriptive name
5. **Add Script Directory section** to SKILL.md (see template below)

### Choosing a Category

| If your skill... | Use category |
|------------------|--------------|
| Processes video or subtitle files | `video-skills` |
| Converts or processes content | `utility-skills` |
| Doesn't fit existing categories | Create a new category with descriptive name |

**Creating a new category**: Add a new plugin object to `marketplace.json` with:
- `name`: Descriptive kebab-case name (e.g., `audio-skills`)
- `description`: Brief description of the category
- `skills`: Array with the skill path

### Writing Effective Descriptions

**MUST write in third person** (not "I can help you" or "You can use this"):

```
# Good
description: Translates subtitle files between languages with auto-detection. Use when user asks to "translate subtitles", "翻译字幕", or needs subtitles converted.

# Bad
description: I can help you translate subtitles
description: You can use this to translate subtitle files
```

Include both **what** the skill does and **when** to use it (triggers/keywords).

### Script Directory Template

Every SKILL.md with scripts MUST include this section:

```
## Script Directory

**Important**: All scripts are located in the `scripts/` subdirectory of this skill.

**Agent Execution Instructions**:
1. Determine this SKILL.md file's directory path as `SKILL_DIR`
2. Script path = `${SKILL_DIR}/scripts/<script-name>.ts`
3. Execute via `bun ${SKILL_DIR}/scripts/<script-name>.ts`
4. Replace all `${SKILL_DIR}` in this document with actual values

**Script Reference**:
| Script | Purpose |
|--------|---------|
| `scripts/main.ts` | Main entry point |
| `scripts/other.ts` | Other functionality |
```

When referencing scripts in workflow sections, use `bun ${SKILL_DIR}/scripts/<name>.ts` so agents can resolve the correct path.

### Progressive Disclosure

For skills with extensive content, use separate reference files:

```
skills/daftai-example/
├── SKILL.md              # Main instructions (<500 lines)
├── references/
│   ├── glossary-en-zh.md # Term mappings
│   ├── config/
│   │   ├── preferences-schema.md
│   │   └── first-time-setup.md
│   └── workflow-details.md
└── scripts/
    └── main.ts
```

In SKILL.md, link to reference files (one level deep only):

```
**Glossary**: See [references/glossary-en-zh.md](references/glossary-en-zh.md)
```

## Extension Support (EXTEND.json / EXTEND.md)

New or migrated preference-aware skills must prefer `EXTEND.json` for reading and new saves, with legacy `EXTEND.md` read compatibility. Preserve scope priority: explicit input > project JSON > project Markdown > user JSON > user Markdown > defaults. Read only the first existing file, validate JSON, and report invalid configuration instead of falling back silently. Never migrate files merely by reading them.

Current implementation: `daftai-chinese-copywriting` is JSON-first; `daftai-chinese-technical-writing` needs no preferences. Other existing skills retain their documented Markdown behavior until their readers and writers are migrated. The legacy guidance below applies to those unmigrated skills.

Skills that need user preferences SHOULD support EXTEND.md configuration. This provides persistent settings (target language, glossary, style, etc.) without asking every time.

### How It Works

Check EXTEND.md existence (priority order):

```bash
# macOS, Linux, WSL, Git Bash
test -f .daftAI-skills/daftai-<skill-name>/EXTEND.md && echo "project"
test -f "$HOME/.daftAI-skills/daftai-<skill-name>/EXTEND.md" && echo "user"
```

| Path | Location |
|------|----------|
| `.daftAI-skills/daftai-<skill-name>/EXTEND.md` | Project directory |
| `$HOME/.daftAI-skills/daftai-<skill-name>/EXTEND.md` | User home |

| Result | Action |
|--------|--------|
| Found | Read, parse, apply settings. On first use in session, briefly remind: "Using preferences from [path]. You can edit EXTEND.md to customize." |
| Not found | **MUST** run first-time setup (see below) — do NOT silently use defaults |

### First-Time Setup (BLOCKING)

**CRITICAL**: When EXTEND.md is not found, you **MUST** run the first-time setup before ANY workflow steps. This is a **BLOCKING** operation.

Full reference: `references/config/first-time-setup.md`

Ask user all preference questions in **ONE** call (not sequentially). After user answers, create EXTEND.md at the chosen location, confirm "Preferences saved to [path]", then continue.

### Parameter Priority Chain

User input (explicit flags/instructions) > Project EXTEND.md > User EXTEND.md > Built-in defaults

### Adding EXTEND.md to a Skill

Every skill with preferences MUST include:

1. **Preferences section** in SKILL.md (or Step 1.1 in workflow):
   - EXTEND.md paths and lookup logic
   - Supported configuration options
   - Link to `references/config/preferences-schema.md`

2. **First-time setup** (`references/config/first-time-setup.md`):
   - Questions to ask user (all in ONE call, not sequentially)
   - EXTEND.md template to generate
   - Save location choice (project vs user home)

3. **Extension Support section** at the end of SKILL.md:
   ```
   ## Extension Support
   Custom configurations via EXTEND.md. See **Preferences** section for paths and supported options.
   ```

**Notes**:
- Replace `<skill-name>` with actual skill name (e.g., `daftai-video-subtitler`)
- Use `$HOME` instead of `~` for cross-platform compatibility (macOS/Linux/WSL)
- Use `test -f` for explicit file existence check

## Output Path Convention

Skills that produce output files should follow consistent naming.

### Output Directory

Each session creates an independent directory. Even the same source file generates a new directory per session.

```
<output-dir>/<topic-slug>/
```

- `<topic-slug>`: Generated from content topic (2-4 words, kebab-case)

### Slug Generation

- Extract main topic from content (2-4 words, kebab-case)
- Example: "Introduction to Machine Learning" → `intro-machine-learning`

### Conflict Resolution

If the output directory already exists:
- Append timestamp: `<topic-slug>-YYYYMMDD-HHMMSS/`
- Example: `ai-future` exists → `ai-future-20260118-143052`
- Never overwrite existing results

### File Naming Convention

Output filenames MUST be descriptive and include meaningful context:

- **Format**: `[NN-]<type>-[slug].<ext>`
  - `NN`: Two-digit sequence number if ordered (01, 02, ...)
  - `<type>`: File type (translated, bilingual, extracted, etc.)
  - `[slug]`: Descriptive kebab-case slug

- **Examples**:
  ```
  video_zh.srt
  video_bilingual.srt
  01-translated-intro.srt
  02-translated-chapter1.srt
  ```

- **Slug Rules**:
  - Derived from file purpose or content (kebab-case)
  - Must be unique within the output directory
  - 2-5 words, concise but descriptive
  - Include language code if applicable: `_zh`, `_en`, `_ja`

### Source Files

- Copy all sources to output directory with naming: `source-{slug}.{ext}`
- Original source files remain unchanged

## Inter-Skill Delegation

Skills that need functionality from other skills MUST delegate rather than re-implementing.

### Delegation Flow Template

Use this template when a skill needs to call another skill:

```
### Step N: Delegate to <skill-name>

**Skill Selection**:
1. Check available skills in `skills/` directory
2. Read selected skill's SKILL.md for parameter reference
3. If multiple skills available, ask user to choose

**Execution Flow**:
1. Call selected skill with required parameters
2. Handle failures gracefully with retry logic
3. Provide clear progress feedback to user
```

### Best Practices

- Always read the target skill's SKILL.md before calling
- Pass parameters exactly as documented in the skill
- Handle failures gracefully with retry logic
- Provide clear progress feedback to user

## Code Style

- TypeScript throughout, no comments unless complex logic
- Only Node.js built-in modules — no npm dependencies
- Async/await patterns
- Type-safe interfaces