# Skill: Conduct multi-source AI deep research with verified, dated, cited output

> Use this skill when you need fresh, fact-checked information from the web -- not a memory-based answer -- and must produce a report where every claim has a source link and a date.

Source: sodigi-learn -- vibecoding/glubokoe-issledovanie - https://sodigi.io/learn/vibecoding/glubokoe-issledovanie

## When to use
- The user asks a question where freshness matters: prices, current events, library status, market options.
- The task requires comparing multiple alternatives before making a recommendation.
- A previous AI answer gave facts with no links or no date and the user needs verification.
- The same information query recurs on a schedule (set up monitoring instead of repeating manually).

## Core rules
- Every fact in the output must have a URL and a publication/access date. No link = no fact.
- Split the question into 3-5 sub-questions; search each one separately.
- Gather at least 15 sources; prefer primary sources and official pages over blogs; prefer blogs over comments.
- Read the 3-5 most important pages in full -- not just the title and snippet.
- If only one source confirms a claim, label it explicitly as "not verified -- single source."
- If data is missing, say "insufficient information" -- never invent a plausible-sounding number.
- Separate four layers in the final report: sourced fact | your supplied data | reasoned guess | recommendation. Never blend them.
- For cheap one-fact lookups (is this library still maintained? current price?) use a quick search, not a full multi-source breakdown.

## Procedure
1. Ask 1-2 clarifying questions before searching: purpose (learn / decide / write), region, budget, required depth.
2. Decompose the main question into 3-5 focused sub-questions.
3. For each sub-question run several differently worded queries; collect 15-30 sources total.
4. Open and read the top 3-5 pages in full; note date, author, domain.
5. Assemble the report: short executive summary at the top, then sections per sub-question, then source list.
6. Attach a link and date to every factual statement.
7. Flag single-source claims and any data gaps honestly at the end of the report.
8. If the question is recurring, propose a monitoring schedule to automate future checks.

## Ready-to-use prompt
```
Do deep research: [your question here].

Work like this:
1. First ask me 2 clarifying questions (region, purpose, depth).
2. Split the task into 3-5 sub-questions.
3. For each sub-question find several sources -- at least 15 total.
4. Read the most important pages in full, not just headlines.
5. Every fact in the answer gets a link and a date. A single-source claim is labelled "not verified."
6. Split the report into four sections: verified facts | my supplied data | your inferences | recommendation.
7. End with a source list and an honest note on any gaps.
```

## Pitfalls
- Accepting an answer that has no links -- any confident number without a source may be invented.
- Forgetting to ask for the date -- for prices, service status, or news, an undated answer is useless.
- Concluding from one source -- one blog is an opinion, not evidence.
- Using full deep research for a trivial one-fact lookup -- one targeted query is enough and cheaper.
- Letting the AI blend facts and guesses in a single paragraph -- demand the four-layer separation.
- Manually repeating the same search weekly -- if it recurs, set up monitoring once.