Hydrafetch

View saved version →

---
name: extract-data-from-pages
description: Pull specific fields such as prices, names, dates or specifications out of one or more web pages as structured data. Use when the user wants values or a table rather than a summary, or the same fields from several pages.
---

Use this skill when the user wants particular values from web pages rather than the pages' full text.

1. Work out the fields the user wants and the pages that hold them. If the user has not given URLs, find them with `search`, or with `map` when they are on one site, and confirm the list with the user when it is long.
2. Call `extract` with the `urls` and either a `prompt` describing the fields or a JSON `schema` with them. Use a `schema` when the user wants a table with fixed columns or the same fields across several pages. Type numbers as numbers and keep fields that may be missing nullable.
3. Set `showSources` to true when several URLs are involved, so each value can be traced to its page. Set `showConfidence` to true when the user asks how sure the values are.
4. Present the result in the shape the user asked for, usually a table, with the source URL for each row.

When data is missing:

- Leave a field empty when the page does not state it, and say which fields were not found. Never fill a gap with a plausible value.
- If a page could not be read, list it separately instead of dropping it silently.

Do not:

- Use `extract` to answer a question about one value on one page. `scrape` that page and read the value instead.
- Extract from more pages than the user needs. Narrow the list first.

The user's explicit instructions take priority over this skill.

Referenced files: 1 — file references are metadata, not package contents.

View saved version →

---
name: find-current-information
description: Find and read current public information on the web, then answer with sources. Use when the user asks about recent events, news, releases or facts that change over time and does not give a specific URL.
---

Use this skill when the answer depends on information that may have changed recently and the user has not pointed at a specific page.

1. Call `search` with a short, specific `query` built from the user's request. Keep the default `limit` unless the user asks for more results.
2. Read the titles and snippets. Pick the results that actually address the question, and prefer primary sources such as the organisation's own site, official announcements and reputable outlets.
3. When snippets are not enough to answer, call `scrape` on the two or three most relevant results rather than all of them. Alternatively set `scrapeResults` to true on the search when every result needs reading.
4. Answer from what the pages say, and cite each source URL next to the claim it supports.

When results are weak:

- If the search returns nothing relevant, say so and suggest a narrower or differently worded query instead of answering from memory.
- If sources disagree, say that they disagree and show both, with links.
- If the search fails, tell the user it failed. Do not present older knowledge as current.

Do not:

- Make claims without a source you actually read in this conversation.
- Pad an answer with results that do not address the question.

The user's explicit instructions take priority over this skill.

Referenced files: 1 — file references are metadata, not package contents.

View saved version →

---
name: read-a-web-page
description: Read a public web page the user links to or names, and answer from its actual content. Use when the user shares a URL, asks to summarise, explain or quote a page, or asks what a specific page says.
---

Use this skill when the user gives a URL or names a specific page and wants to know what it says.

1. Call `scrape` with the `url`. Set `preferStructure` to true when the page's meaning lives in tables, lists or headings, such as pricing pages, comparison grids, documentation and spec sheets.
2. Answer from the returned markdown only. Quote or paraphrase what the page says, and link the URL you read.
3. If the user asked for a format, such as bullets or a table, build it from the page content.

When the result is thin or looks wrong:

- If the content is very short, is mostly a cookie notice, or says the page needs JavaScript, say what came back and offer to try again rather than guessing at the rest.
- If the page returned an error or was refused, tell the user plainly that the page could not be read. Do not fill the gap from memory.
- If the page redirected somewhere unexpected, mention the final URL.

Do not:

- Invent facts the page does not state, even when they seem likely.
- Read pages that need the user to be signed in. This tool only reads public pages; say so if the user asks for account or private data.
- Scrape a whole site when one page answers the question. If the user wants several pages from one site, list them with `map` first and read only the ones that matter.

The user's explicit instructions take priority over this skill.

Referenced files: 1 — file references are metadata, not package contents.

View saved version →

---
name: research-a-company
description: Build a short, sourced profile of a company from its website, including what it does, its brand and its key pages. Use when the user asks what a company does, wants a company overview, or wants its logo, colours or fonts.
---

Use this skill when the user names a company or gives its domain and wants to understand it. Work in this order and stop as soon as the question is answered.

1. Call `brand` with the company's `domain`. It returns the name, description, logos for light and dark backgrounds, colours and social profiles. If the user only asked for the logo, call `logo` with the `domain` instead.
2. If the user wants to know what the company offers or how it describes itself, call `scrape` on its homepage. Add its pricing or about page only when the question is about pricing or history; set `preferStructure` to true on pricing pages.
3. If the user wants to know what is on the site, call `map` with the homepage URL and show the relevant pages instead of reading them all.

Write the answer:

- Lead with a one or two sentence summary of what the company does, in its own words where possible.
- Then give only the details the user asked for, such as offering, pricing tiers, logo links, colours or fonts.
- Link every page you read.

When data is missing:

- If `brand` returns no description or no logo, say it was not found. Do not describe a logo or colour you have not seen.
- If the domain does not resolve or belongs to a different company than the user meant, say so and ask which company they meant.

Do not:

- Guess headcount, funding, revenue or founders. Report them only when a page you read states them, and cite that page.
- Read every page of the site when a single page answers the question.

The user's explicit instructions take priority over this skill.

Referenced files: 1 — file references are metadata, not package contents.