You have 1,400 notes. Roughly forty of them are useful, and you couldn't name which. The rest are web clips with one highlighted sentence, meeting jottings that stop mid-thought, and a daily note from March that says "ask about the thing."
That's the usual shape of a second brain after a year: easy to add to, hard to get anything back from. Writers on the topic have a name for it, the collector's fallacy: gathering feels like learning, and the pile never gets used. Most of them are practitioners and bloggers rather than researchers, but the pattern is familiar enough. The fix isn't a bigger system. It's cheaper processing, so notes become something you'd actually use.
This is where AI prompts for a second brain earn their place. Not as an "auto-organiser" that touches your files, but as a processing step at four points: after capture, when linking, at weekly review, and when you're hunting for something. These are untested templates, as I haven't run them against your vault. The tool details come from the vendors' own help pages, checked on 2026-10-08.
What should AI actually do with my notes?
Three jobs, in order of safety:
- Compress. Turn a messy capture into a clean, short note you wrote the substance of.
- Connect. Suggest which existing notes relate, with a reason. You decide.
- Prompt. Ask you questions: what would you do with this, what does it contradict?
And one job to be careful with: decide what to delete or merge. Suggestions are fine, actions aren't. Any tool that edits files should work on a copy first, and you should have version history or a backup before letting it near the whole vault.
It also helps to separate where the AI runs. Obsidian stores notes as plain Markdown files with properties as YAML at the top. It has no built-in chat model (I could find no such feature in its help docs), so AI there means pasting text into a chat tool or using a community plugin. Notion has AI built in. According to its help page, Notion AI can edit text inline, fill database content, and draw on pages or people you @-mention. It's on Business and Enterprise plans, with a limited number of free responses on Free and Plus. All the prompts below work as a paste into any chat window; I'll note where a tool changes the mechanics.
Capture to note: the processing prompt
Your raw capture is a quote from an article, three sentences of your own, and a link. Processing means making it findable and reusable six months from now. The step that matters most is writing the "so what" in your own words, so make the model ask for it rather than write it.
Process this raw capture into a note.
RAW CAPTURE:
[PASTE: quote, my thoughts, URL, date]
Output in Markdown:
- Title: a statement of the idea, not the topic (e.g. "Small batches cut
review time", not "Batch size").
- Summary: 2 sentences using only my text and the quote. Do not add facts.
- Source: URL and date as given. If missing, write SOURCE MISSING.
- Why I saved it: ask me ONE question I should answer in my own words
(do not answer it for me).
- 3 candidate tags, lower-case, from this list if they fit: [MY TAG LIST].
- Properties block in YAML with: type, source, created, status: inbox.
If the capture is too thin to be useful, say so and suggest deleting it.
The last line is valuable. A good processing prompt should be willing to say "this isn't worth keeping." Most captures aren't.
The properties block follows Obsidian's format, which per its help page stores properties as YAML at the top of the note, with types such as text, list, number, checkbox, date and tags. In Notion, you'd set the same fields as database properties instead. Same idea, different place.
Linking: suggestions with reasons
Obsidian's internal links use [[Note name]] syntax, can point to headings, and can use aliases, per its links help page. Obsidian can also update links automatically when you rename a note (there's a setting to be asked first). In Notion you'd @-mention a page or use a relation property. The tool isn't the hard part. Knowing which notes to link is.
Paste the new note and a list of existing titles (not the full text of everything), then ask for candidates with the reasoning.
NEW NOTE:
[PASTE]
EXISTING NOTE TITLES (with one-line summaries where I have them):
[PASTE LIST, up to ~100 lines]
Suggest up to 5 notes the new note should link to. For each:
- the title exactly as listed,
- one sentence on the relationship (supports, contradicts, example of,
prerequisite for, same problem different domain),
- the phrase in the new note where the link would sit.
Only use titles from the list. If nothing fits, say "no good links" and
suggest what kind of note I'm missing.
"Only use titles from the list" matters. Without it the model invents plausible note names that don't exist, and you spend ten minutes looking for a note you never wrote. Even with it, check each suggestion before adding it. A link you can't justify is noise.
For a big vault you can't paste every title. Pick a slice: notes in the same folder, the same tag or the last 90 days. That's an honest limit of a paste-and-ask workflow. Tools that search your vault do better here but come with their own privacy trade-offs.
The weekly review prompt
This one pays for the whole system. Once a week, take the notes you created or edited in the last seven days and paste them in.
Here are my notes from the past week (titles + content):
[PASTE]
1. Group them into at most 5 themes. Name each theme in a phrase.
2. For each theme, one sentence: what am I actually learning or deciding?
3. Which notes look unfinished (no conclusion, no next step)? List them.
4. Which notes contradict each other or an earlier belief? Quote both.
5. Pick the 3 notes most worth turning into something (a decision, a
message, a draft) and name the output.
6. Which 5 should I archive or delete? Give a reason for each. Do NOT
delete anything; list only.
7. Ask me 3 questions to answer in the next review.
Use only the text above. If something is unclear, say so.
Item 5 is the antidote to the write-only vault. Several practitioners suggest keeping a small active set and tying each note to a concrete next use, such as a project, a decision or a piece of writing. That's advice from blogs and forums, not research, but it matches what I'd tell you anyway. The test of the system isn't how many notes it has. It's whether last week's notes changed anything you did.
Put the review in your calendar. A prompt can't do the habit for you. Obsidian's core Daily notes plugin, per its help page, creates a dated note you can pre-fill from a template, so a "weekly review" template can be one click.
Retrieval: ask for what you saved, not what the model knows
The classic failure: you ask a general chat "what did I decide about X?" and it answers from the internet. Retrieval prompts need your notes in the context and a ban on outside knowledge.
Answer ONLY from the notes below. Quote the note title and the sentence you
rely on. If the notes don't answer the question, say "not in my notes" and
list the 3 closest notes by title.
QUESTION: [WHAT DID I CONCLUDE ABOUT ...?]
NOTES:
[PASTE the relevant notes, each labelled with its title]
If you want this over a whole vault, you'd need a tool that indexes it, and the quality depends on its search. Our posts on a second brain with OpenClaw and personal AI workflows cover persistent setups; for a document-grounded alternative, how to use NotebookLM is a good fit for a bounded set of sources. The principle in all cases: quotes and titles, or it doesn't count.
A worked example (illustrative)
An invented scenario, not a real run. Your raw capture:
"Teams that review in batches under 200 lines find more defects" (from a
blog post, 2026-09-14, URL saved). My thought: maybe why our huge PRs feel
useless. Try limiting PR size at work?
A good processed result (shape):
Title: Smaller pull requests may get better reviews
Summary: A blog post claims review quality drops as batch size grows past
about 200 lines. I suspect our oversized PRs get skimmed.
Source: [URL], saved 2026-09-14. (Claim not verified; source is a blog.)
Why I saved it: Question for you: what would you change in the next two
weeks if this were true, and what evidence would convince you it isn't?
Tags: code-review, process, experiment
type: idea / status: inbox
Notice what the model did not do. It didn't turn "maybe" into a finding or supply a study. It labelled the claim unverified and asked you to commit to a test. Then your answer to that question, in your words, is the part that turns a clipping into a note.
Where this goes wrong
| Failure | What it looks like | Fix |
|---|---|---|
| Invented links | Notes suggested that don't exist | "Only titles from the list"; check each |
| Summaries that add facts | The model fills gaps in a thin capture | "Use only my text" rule; SOURCE MISSING flag |
| Over-tagging | 40 tags nobody uses | Give a closed tag list; max 3 tags |
| Auto-restructure regret | Hundreds of notes renamed or merged | Suggest-only; backup first; small batches |
| Hollow notes | Perfect summaries, no thought of your own | The "why I saved it" question; you answer |
| Confident wrong retrieval | Answers from model memory | "Answer only from the notes below"; require quotes |
| Private data leaving | Health, client or employer information pasted | Anonymise or keep it out; check tool settings |
What not to automate
- The thinking. If AI writes the "why this matters" line, you've outsourced the one thing that makes a note yours.
- Deletion and merging. Suggest, then act by hand. Notes you'll want in a year look useless today.
- Sensitive notes. Journal entries, health, finances, client details and employer-confidential material shouldn't go into a tool you haven't vetted. Check the vendor's data-use settings, which differ by product and plan. I'm not stating them here since they change.
- Bulk reorganisation. Restructuring the whole vault at once breaks your own map of it.
- Claims you intend to repeat. A processed note that says "studies show" is a prompt to check the source, not a citation.
If you're starting fresh, keep it small: an inbox, a handful of tags, one weekly review. The prompts above will sit on top of that. For writing better instructions for any of them, giving context is the lesson that matters most, and AI meeting notes and action items covers the transcript side. The prompt library has more templates to copy.
