Getting Good Results From Your Sources
When an AI creation comes back thin, generic, or about the wrong thing, the instinct is to blame the prompt and try a different model. It is almost never the prompt.
Scholarly builds everything from what it can actually read in your source. A 200-page textbook where only the first chapter is readable produces a lecture about the first chapter. A screenshot of a sign-in page produces nothing worth having. This guide is about getting that part right — it is the single highest-leverage thing you can do.
Scholarly Checks Before You Spend a Credit
You no longer have to guess. When you pick a source in a creation window, Scholarly looks at it first and tells you if it will not read well, naming the problem and suggesting what to use instead.
You will see a note when a source turns out to be:
- A sign-in or paywall page — the save captured the login screen, not the article behind it.
- Mostly navigation — menus, cookie banners, and repeated headers with little actual content.
- Unreadable text — a scan or export whose text cannot be extracted.
- Empty — nothing to work from at all.
The note is advice, not a block. You can go ahead anyway if you know something it doesn't. If you upload a fresh file with the same problem, the note appears on the progress page instead.
Audio and video with no speech
A clip whose soundtrack is only music or applause tells you the moment you pick it that it has no spoken audio, rather than running a creation that could only come back empty. No credit is spent.
The same applies after the fact:
- A meeting recording that picked up no speech is reported as exactly that, and its audio is kept so you can play it back and check for yourself. It does not count toward your weekly meetings.
- An audio file in your library with only music or applause is reported as having no speech, instead of producing a transcript made of sound cues.
On the Free plan, a file that genuinely cannot be read — a blank photo, a PDF or text file with nothing to extract, audio with no speech — returns your free upload so you can add a better one. The unreadable file moves to Deleted.
What Reads Well
| Source | Reads well | Reads badly |
|---|---|---|
| Exported from Word, LaTeX, or a publisher; real selectable text | A photo of pages; a fax-quality scan; a form with no text layer | |
| Slides | A real .pptx or PDF export, speaker notes included | A screenshot of a slide |
| Audio / video | Clear speech, one or a few speakers | Music only; heavy crosstalk; a phone across a large room |
| Web page | The article itself | The page behind a login; a feed or index page |
| Photos of notes | Flat, well-lit, in focus, one page per shot | Angled, shadowed, several pages at once |
| Your own notes | Pasted straight in as text | A screenshot of the notes |
Two things matter more than resolution or file size: is the text extractable, and is the content actually on the page you gave it.
Scanned PDFs
Scanned pages work — Scholarly reads them visually when there is no text to extract, including diagrams, equations, tables, and handwriting. Large scanned documents render reliably.
But visual reading is slower and less exact than real text. If you have the original digital file, use it. If you only have the scan, keep the pages flat and the contrast high.
Pick the Pages That Matter
This is the setting with the biggest effect on quality, and the one most often skipped.
On a long PDF, the page picker in the creation window reaches every page of the document. If your plan reads 32 pages per document, you can choose any 32 consecutive pages — pages 90 to 121, say — not just the first 32.
So on a 400-page textbook, don't create from "the textbook". Create from chapter 9. The output will be dramatically better, and you will see what you expected.
Page limits by plan are in Plans and Limits.
Check It in Chat First — Ten Seconds
Before you spend a credit on a long creation, ask about the source in chat:
- Open chat and type @, then pick the file.
- Ask something only that document can answer — "what does section 4 require?", "what were the three action items?", "what method did they use?"
- Look at the Sources button under the answer. It lists what the reply actually cited, and each entry opens the page it came from.
If the answer is vague or the cited pages look wrong, the source is the problem. Fix it now rather than after a 20-minute render. See Sources and Citations in Chat.
Fixing a Source That Doesn't Read
| Symptom | Fix |
|---|---|
| A scan with no readable text | Re-export from the original, or use the page picker on the clearest pages |
| You saved a sign-in page | Copy the article text and paste it in instead |
| A web page came through as menus | Paste the body text, or use the PDF or print view |
| Audio has no speech | Check you uploaded the right file; music-only clips cannot be transcribed |
| A recording is hard to follow | Re-record closer to the speaker; one clear mic beats a room mic |
| The file is too large for your plan | Every upload and create screen shows the file's size beside your plan's limit and names the plan that accepts it |
| "Choose file again" appears | The file moved or was changed by iCloud, Drive, or another app after you picked it — pick it again |
| The output covered the wrong part | Use the page picker, or split the document |
Documents, text files, and images up to 50 MB are read as soon as you choose them, so a file that another app moves afterwards still uploads when you press Create.
More Sources Isn't Better
Every creation tool can combine several sources — one on Free, up to ten on Premium, up to twenty on Laureate. Use that for material that genuinely belongs together: the chapter plus the lecture plus your notes on it.
Attaching eight unrelated documents does not produce a richer creation. It produces a vaguer one, because the AI has to find something all eight have in common. Narrow beats broad nearly every time.
Where the Suggestions Come From
Once a creation finishes, the next-step suggestions on the progress page follow the source. A recorded lecture, a contract, and a slide deck each get the formats that suit them — which is another reason a well-read source pays off twice.
Related
- Uploading Content — every way to bring a source in.
- Supported File Formats — the full format list and size limits.
- Paste Text — when a file won't cooperate.
- Sources and Citations in Chat — checking what an answer actually used.
- When a Creation Fails — every failure reason and the matching fix.
- Working with PDFs — page ranges, bookmarks, and PDF chat.
- Plans and Limits — pages per document, sources per creation, file sizes.
- Choosing What to Create — matching the output to the material.