AI research: keep the source, then ask the model

A research workflow fails when the only copy of a paper page or a docs section is a chat transcript. The transcript paraphrases. The page is the source. WebInk’s role is to freeze that page as Markdown with a URL, so a later question to ChatGPT, Claude, or Gemini can be checked against a file you still have.

Capture one source at a time

Use the URL to Markdown tool for a public link: a blog post, an RFC mirror, an open docs page. Use the extension when you are already signed in, which is common for library portals and product dashboards. Do not point either one at a list of fifty URLs. WebInk converts a page, not a crawl.

Read the preview before you save. Research notes are only as good as the extraction. A sidebar of “related papers” inside the body will be cited later as if it were the article. Cut it in the preview, or fix the capture, before the file enters a folder you trust.

Put the citation in the file

A template should carry the title, the source URL, and the published time when the page has one. Then the body. That front matter is what you search in six weeks. The templates page shows the fields. Quote titles that contain colons so the YAML stays valid.

Download a single .md file for the text. In the extension, choose Download .zip when a figure has to travel with the note. The save guide spells out filenames, the images/ rewrite, and the toast you get when a fetch fails. Remote image links are fine for a source you can reopen. They are a bad archive if the host rotates URLs.

Ask the model about the note

Send the Markdown, not the original HTML. The send guide explains which assistants get a prefilled prompt and which ones receive a copy. Ask for claims that quote a heading or a sentence from the note, and keep the file when the answer matters. If the model invents a number, the note is how you notice.

Do not paste paywalled or confidential text into a public chat just because it converted cleanly. Conversion is not permission. For a page that should never be fetched by the URL tool, stay on the extension and decide locally whether the prompt leaves your machine.

The token comparison on HTML vs Markdown for LLMs is an illustration, not a meter. Use it to decide the format. Use the saved note to decide the fact.

One afternoon, three sources

Suppose you are comparing how three public posts describe cache invalidation. Give each post its own file. A shared scratch buffer is how quotations get swapped between authors.

1. For each public URL, run URL to Markdown. Read the preview far enough to see whether a related-posts list was pulled into the body. Delete that list in the preview before it becomes a “source” the model can cite.
2. Download the .md. The web tool slugifies the title, so “Cache invalidation notes” becomes a hyphenated lowercase file. Rename to something stable if your folder convention is a date prefix, for example 2026-04-02-example-cache.md.
3. Open the file and add a source line if the download is body-only. On the extension, switch to Basic Frontmatter before you download so title, url, and publishedTime are written for you. Quote the title in that template when the headline contains a colon. The substitution does not add quotes.
4. If a diagram matters, repeat the capture in the extension after the image has loaded and use Download .zip. Unzip into sources/example-cache/ so images/ stays next to the note. The save guide has the rewrite rule.
5. Ask the model about one file at a time. A prompt that worked for this comparison:

Using only the note below, list claims about cache invalidation.
Quote the sentence you used. If a claim is not in the note, say it is absent.

{{markdown}}

That block is a prompt template for the extension, which substitutes {{markdown}}. On the web tool you can paste the same instruction in front of the downloaded file. The stock Summarize prompt will not do this job. It sends the URL, which the send guide lists next to the other five built-in prompts.

Check the answer against the file

Read every number in the model’s reply and find it in the note. If it is not there, the model added it. A useful research note records that rejection in a short section you write yourself, under the captured body, so the next session does not rediscover it.

## Check on 2026-11-09

Asked which TTL the post recommends.
The note says writers delete the key on update.
It does not recommend a TTL. Rejected the model's "five minutes" figure.

Keep that section in your words. Do not paste the whole transcript back into the source file. The transcript is long, and it mixes the model’s guesses with the page. The capture stays the capture. Your check stays a few lines at the bottom.

When a second post disagrees, link the two files with ordinary Markdown links and a sentence about the disagreement. WebInk does not build that link graph. You do, after both previews were honest.

Extension controls that help a reading session

The side panel can translate the article Markdown into a language you pick, including English, Chinese, Japanese, Spanish, and the other languages in that menu. Translation runs on the body, and it requires the selected template to contain {{markdown}}. If it does not, the panel reports that the current template does not contain markdown content. The Default template is only that variable, so it qualifies. The Markdown Link template does not, because it stores a link and an excerpt.

AI Fix is the wrench on the Markdown tab. It tries to repair fences, tables, leftover HTML, and obvious chrome such as share buttons. It is disabled above 100,000 characters. Read the result before you save. A formatting pass can also drop a line you meant to keep. It is not a substitute for reading the preview.

The table of contents, described on the TOC feature page, can be copied on its own or inserted into the note. For a long spec, a TOC at the top makes the later quotation easier to find. For a three-paragraph post, skip it.

When the source is a spec, not an essay

A standards page or a long API reference should not go to the model as one blob if you only need one operation. Highlight that section on the page and use a prompt that includes {{selection}}. The stock Explain Selection prompt does this, and it also appends the page URL. It refuses to send when the selection is empty. The selection text is the page’s plain text, so a highlighted code sample may lose its fence. If the fence matters, send {{markdown}} and name the heading you care about in the instruction.

Keep a running index note that is only links and one-line conclusions, separate from the captures:

# Cache reading index

- [Example post](2026-04-02-example-cache.md) — delete on write. No TTL recommendation.
- [Docs: cache](example-docs-cache.md) — same claim, plus a DEL snippet.

WebInk does not update that index. You add a line after each check. The index is what you reopen. The captures stay untouched so a quotation still points at the original sentences.

Limits

• One page per conversion. A literature review is a sequence of captures, not one request that accepts a bibliography.
• The URL tool and the API do not see your cookies. Library portals and dashboards belong on the extension.
• The extension removes a leading H1 from the body. If your notes are filed by the first heading, put # {{title}} in the template.
• Send to AI does not upload images. A vision question needs the chat product’s uploader, or a file from the ZIP’s images folder.
• WebInk does not score citations, talk to a reference manager, or watch a page for later edits. Reconvert when the page changes and keep the old file if the old quotation still matters.

If the destination folder is an Obsidian vault, continue with the Obsidian use case after the file is worth keeping. If the source is a docs site you will fetch again from a script, the developers use case is the handoff.

FAQ