skilless.ai-research（熟练的 AI 研究）

Research Skill

Fetch internet content or conduct deep multi-source research.

Routing Table

When the user intent is ambiguous, ask one clarifying question before proceeding.

Research Depth Levels

Not every request needs a full investigation. Choose the appropriate depth based on complexity, or ask the user if unclear.

"Round" = one cycle of parallel tool calls. A single round can fire 3-5 searches/reads in parallel. Rounds are sequential — each round builds on what you learned in the previous one.

Decision guide:

• If the question can be answered with a single search result → L1
• If it involves comparing options or understanding a topic in depth → L2
• If it requires multiple perspectives, data synthesis, or the user explicitly says "research" / "investigate" / "deep dive" / "深度" → L3
• If you cannot determine the level, ask the user: "This could be a quick lookup or a deeper investigation — how thorough would you like me to be?"

Research Workflow (L2 and L3)

Step 1 — Define Scope

Before searching, clarify:

• What is the core question? What does the user actually need to know?
• What source types are needed — official docs, news, academic papers, user reviews?
• What dimensions matter (price, performance, compatibility, recency...)?
• How confident does the conclusion need to be?

If any of the above is unclear, ask the user before proceeding.

Step 2 — Create Research Plan (L3 mandatory, L2 recommended)

Decompose the task into research dimensions — each dimension is an independent angle of investigation. Use todo, tasks, todowrite, or equivalent tool to create a visible task list.

Example — "深度检索 AI coding 领域的创业公司，区分海内外，并简单判断投资潜力":

[ ] 1. 海外 AI coding 创业公司全景 (who, what, founded when)
[ ] 2. 国内 AI coding 创业公司全景
[ ] 3. 各公司融资情况、投资方、估值
[ ] 4. 产品对比：功能、定价、目标用户
[ ] 5. 市场格局和竞争态势
[ ] 6. 投资潜力评估（团队、技术壁垒、增长、风险）

Rules:

• L3 must have at least 3 dimensions, typically 4-6
• Each dimension will get its own dedicated search round(s)
• Do NOT skip this step — a task list forces thoroughness

Step 3 — Execute Per-Dimension (Multi-Round Search)

Execute each dimension as a focused investigation cycle:

For each dimension in the task list:

1. Mark in-progress — Before starting a dimension, update its todo status to in-progress
2. Search — Fire 1-3 parallel searches targeting this specific dimension
3. Read — Open and read the most relevant sources from search results
4. Record — Extract key data points and note sources
5. Mark complete — Immediately after finishing a dimension, mark it completed in the todo list
6. Adapt — If new sub-questions, gaps, or unexpected angles emerge, add new todo items to the list before moving on; do not silently absorb them

Dynamic todo updates are mandatory throughout execution:

• Never batch completions — mark each item done the moment it is finished
• If a dimension turns out to be larger than expected, split it into sub-items
• If a dimension turns out to be irrelevant, mark it skipped with a one-line reason
• The todo list is a live document — the user should always be able to glance at it and understand what has been done and what remains

Search angle variety within each round:

• Broad terms — Get the landscape, find authoritative sources
• Specific terms — Target precise data (version numbers, prices, specs)
• Contrarian terms — Search for "problems", "downsides", "alternatives", "vs" to find critical perspectives
• Recency terms — Add year or "2025" to filter outdated content
• Multi-language — Search in both English and Chinese when relevant

Parallelism: Within a single round, fire multiple tool calls in parallel (e.g., 3 searches at once, or 2 searches + 1 web read). Each round should maximize parallel execution.

If you have not met these minimums, keep researching. Do NOT shortcut.

Step 4 — Fact-Check

Do not skip this step.

• Cross-verify key data from multiple independent sources (aim for 2-3; more for L3)
• If only a single source exists, explicitly note: "Based on a single source — not independently verified"
• When sources contradict each other:

1. Identify the reason (timing difference? different versions? conflicting interests?) 2. State which source is more credible and why 3. If you cannot resolve the contradiction, present both sides and ask the user how they want to proceed

• Distinguish facts from opinions — opinions may be cited but must not be stated as facts
• Never fabricate or assume unverified data — say "insufficient evidence" instead

Step 5 — Synthesize Output

• Conclusion first — Lead with the most important finding or recommendation, then expand
• Cite every data point with [1], [2], etc., so the reader can verify
• Be honest about uncertainty — If evidence is limited, say so; do not force a conclusion
• End with a full source list

Contradiction and Uncertainty Handling

When you encounter any of the following during research, do not silently resolve it — surface it to the user:

Format for follow-up questions:

⚠️ Needs clarification: [concise description of the issue] Options: 1. [Option A] 2. [Option B] Which would you prefer, or would you like me to handle it differently?

Error Handling

When a tool fails, follow this protocol — never silently fail or fabricate content:

Core principle: If a tool fails and you cannot recover, tell the user what happened, what you tried, and suggest next steps. Never pretend it succeeded.

Output Format

Strict portable Markdown only. The output must render correctly in any Markdown editor (GitHub, Obsidian, Typora, VS Code, etc.).

Rules

1. No HTML tags — no <br>, <div>, <table>, <sub>, <sup>, or any HTML whatsoever
2. Table cells must be single-line plain text — no line breaks, no nested lists, no multi-line content inside a cell
3. If content does not fit single-line table cells, use a list instead of a table
4. Use blank lines before and after headings, tables, code blocks, and block quotes to ensure correct parsing
5. Do not use indented code blocks — always use fenced code blocks with triple backticks

Formatting Toolkit

Table Example (Use Only When Appropriate)

Tables are for comparing multiple items with short single-line values:

| Product | Price | Offline | Rating |
| ------- | ----- | ------- | ------ |
| A       | $99   | ✅       | 4.5/5  |
| B       | $199  | ❌       | 4.0/5  |
| C       | Free  | ✅       | 3.5/5  |

Do not use tables when: content is single-column, cells need multi-line text, or structure requires nesting.

Source Citations

Inline: [1], [2], etc.

At the end:

## Sources
[1] [Title](URL) — Key data description
[2] [Title](URL)

Context Management

• Long content (>2000 words): Summarize key information after extraction; do not paste raw content into the response
• Subtitles: Download to disk by default — only load into context when the user explicitly asks for content analysis (e.g. "summarize this video", "extract info about X")
• Multiple pages: Synthesize and integrate findings; do not stack raw page dumps

CLI Tools Reference

📁 Working directory: All commands run from ~/.agents/skills/skilless/. The cd ~/.agents/skills/skilless/ && prefix is shown in full for each command to ensure correct execution.

Search (Exa AI)

cd ~/.agents/skills/skilless/ && uv run scripts/search.py "your query"
cd ~/.agents/skills/skilless/ && uv run scripts/search.py "your query" 10

Web Reader (Jina Reader)

cd ~/.agents/skills/skilless/ && uv run scripts/web.py <url>

Video / Transcript Extractor

youtube.py — One-step subtitle + metadata extraction (recommended for most cases):

cd ~/.agents/skills/skilless/ && uv run scripts/youtube.py "<url>"

yt-dlp direct — Advanced usage (custom formats, audio-only, subtitle listing):

cd ~/.agents/skills/skilless/ && uv run yt-dlp "URL"
cd ~/.agents/skills/skilless/ && uv run yt-dlp --list-subs "URL"
cd ~/.agents/skills/skilless/ && uv run yt-dlp --write-subs --write-auto-subs "URL"
cd ~/.agents/skills/skilless/ && uv run yt-dlp -x --audio-format mp3 "URL"

Supported platforms (1700+ via yt-dlp): YouTube, Bilibili, TikTok, Twitter/X, Twitch, Vimeo, Dailymotion, Niconico, Rumble, Odysee, SoundCloud, Reddit, Instagram, Facebook, and many more.

Download path rules:

• Specific project directory (e.g. ~/codes/my-project/) → download to current working directory
• Home directory (~) or empty path → download to ~/Downloads/
• Never download to /tmp — requires special permissions, files may auto-delete

YouTube troubleshooting:

• "Sign in to confirm you're not a bot" or timeout → enable TUN mode proxy, or try --cookies-from-browser chrome, or use Bilibili as alternative

FFmpeg (Media Converter)

cd ~/.agents/skills/skilless/ && uv run scripts/ffmpeg.py <input> <output>
cd ~/.agents/skills/skilless/ && uv run scripts/ffmpeg.py video.mkv output.mp4
cd ~/.agents/skills/skilless/ && uv run scripts/ffmpeg.py audio.wav output.mp3
cd ~/.agents/skills/skilless/ && uv run scripts/ffmpeg.py input.mp4 output.mp4 -crf 28

Supports all common media formats (mp4, mkv, mp3, wav, flac, webm, avi, mov, etc.)

Cross-References

• Need a detailed report? → After completing research, invoke skilless.ai-writing to produce professional reports, articles, documentation, or any structured written content from your findings
• Research goal unclear? → Invoke skilless.ai-brainstorming to define scope, clarify questions, and explore approaches before starting a deep investigation