Should I vibe code
Collects and connects short ideas into a browsable personal knowledge graph
The AI that connects your ideas is an embedding and a cosine. What you're paying for is two taps to capture.
?
Their verdict, the All features (paid monthly) price and the build-time estimate come from their entry, MIT-licensed. Checked 2026-08-04.
?
Our verdict, the regret score and everything below it. Editorial and unsponsored — nobody can pay to be moved.
The honest answer
why the verdict is what it is
Underneath the language about reflection and flow, Napkin is a text box, a search index and a shuffle. Capture a quote in one gesture, store it, find it again, and now and then be shown something you had forgotten. That is a genuine evening on SQLite with FTS5, and the "personal AI that understands your ideas" is an embedding call and a cosine similarity — the part that felt like magic in 2023 is now a library import. Your version can beat the subscription on the one axis that matters for a lifelong archive: the notes stay in files you can read without the app. Two things deserve actual thought. If you compute embeddings through a hosted API, every half-formed thought about a colleague becomes a request body on someone else's server, so run the model locally or decide knowingly. And this category dies of friction, not of features: if capture takes three taps the habit is gone inside a fortnight, which means the keyboard shortcut is the product and the graph is decoration.
What actually breaks
not "if". the specific failures.
- Capture friction, the only metric that matters here — three taps instead of one and the habit is dead within a fortnight
- Sync between phone and laptop, the classic, and the reason most personal capture tools quietly become one-device tools
- Search once there are four thousand notes and substring matching stops being enough to find the thing you half remember
- The connection engine without a similarity threshold, which discovers that every note is related to every other note
- Your privacy boundary, if embeddings are computed by a hosted API — a fleeting note about a colleague is now in someone's request logs
- Backups, because an idea archive is a single file you will not notice losing until the day you go looking for something in it
- The app itself, eventually, which is exactly why the notes should be readable without it
Is that you?
the verdict is a default, not a law
- Notes are plain text or Markdown on disk and the archive is readable with grep and nothing else
- Full-text search is SQLite FTS5 rather than a vector database you have to operate
- Capture is one keystroke on the desktop or one share-sheet tap on the phone, and you measured it
- Backup is a directory sync you can verify by opening the folder and looking
- You want it on three devices and are not prepared to adopt a sync library or accept file-sync conflicts
- The notes are about other people and you are about to POST them to a hosted embedding endpoint without thinking about it
- The only copy lives inside an app database with no export, which is the one mistake here that compounds with time
- You are building the knowledge graph before the capture path, which is the usual way this project ends unused
If you build it anyway
the checklist, then the prompt that enforces it
- Make the file the product. Markdown or plain text on disk, one note per file or one append-only log, so the archive outlives every version of the app including the last one.
- Build capture before storage design gets interesting. Time yourself: idea to saved in under two seconds, or the tool will not survive a busy week.
- Use SQLite FTS5 for search and treat the index as disposable — rebuildable from the files at any time. Never let the index become the only place a note exists.
- If you want similarity, run the embedding model locally. If you use a hosted one, say so in the README and add a per-note flag to exclude something from it.
- Put a threshold on the connection engine and show fewer, better links. Related-to-everything is the same as related-to-nothing, and it is what makes these tools feel like noise.
- Back up by copying a directory, and test the restore once. An idea archive fails silently: you do not notice it is gone until you go looking, which is months later.
I want a personal idea-capture app: quick capture, search, and a bit of AI that surfaces
related notes. Build it in this order and tell me if I try to skip ahead.
1. Storage format first. Plain Markdown files in one directory, one note per file, with the
timestamp in the filename and any tags in frontmatter. Do not put note text in a database
as the primary copy. If I ask for that, push back: this archive should outlive the app.
2. Then capture, and treat it as the hardest requirement. A single global keystroke opens a
box, and closing it saves. Measure the time from keypress to file on disk and tell me the
number. If it is over two seconds, fix it before building anything else.
3. Then search: SQLite FTS5, with the index built from the files and fully rebuildable from
them. The index is a cache, never the source of truth.
4. Then, and only then, similarity. Default to a local embedding model. If I ask for a hosted
API, stop and tell me plainly that every note becomes a request body on someone else's
server, and add a per-note exclusion flag before you wire it up.
5. Apply a similarity threshold and cap suggestions at five. Explain in the README what the
threshold is and how to change it. Suggesting everything is the same as suggesting nothing.
6. No accounts, no server, no telemetry. If a feature needs the network, name it and why.
7. Do not build sync. If I ask for a second device, explain the choice between adopting a sync
library and accepting conflicted copies from file sync, and let me pick knowingly.
8. Write a backup and restore procedure into the README, then walk me through restoring from
scratch once. An idea archive fails quietly and is missed months later.
9. Out of scope unless I ask: images, a graph visualisation, publishing, sharing, mobile
widgets, and anything that turns my private notes into a product surface.
10. When you are done, show me how to answer "what did I write about X" using only grep, with
the app closed.
That one keeps you out of trouble. For the prompt that actually builds it, canivibecodeit.com has one.
their build prompt ↗Or don’t build it
the boring option, and the way back out
The ritual is what you want rather than the storage — the curated daily flow, the resurfacing that feels like someone chose it — and you are on an iPhone. At $5 a month, or $40 a year, you are buying taste and a capture path someone polished, not technology. Go in with eyes open about the vendor, though: the pricing page 404s, the old web app is down, and the price is only discoverable in an App Store listing. That is not fatal for a $5 app, but it is another argument for keeping a copy of your notes in a format you can read without them.
$5/mo is cheaper than your weekend.
If you followed the file-first rule, there is no exit to plan: the archive is a folder of Markdown that Obsidian, Logseq, Memos, an editor or grep can all read tomorrow. Delete the app and you have lost a search box, not a decade of notes. That is the entire reason to build this rather than rent it.
MIT self-hosted quick-capture note tool, Markdown-native and close to this product's actual shape.
Local-first outliner and knowledge graph that stores everything as plain Markdown files.
Questions
The AI is the whole selling point. Can I really replace that?
The mechanism, yes. Embedding each note and ranking by cosine similarity is a small amount of code and there are good local models. What you are unlikely to match quickly is the editorial judgement — which three of the forty related notes to show, and when to show them. That is tuning and taste, not architecture, and it is what a well-made version of this is actually selling.
Why does a note-taking app score anything above zero?
Two things. Your notes are yours and often about other people, so sending them to a hosted embedding API moves them somewhere you do not control. And an idea archive is unusually easy to lose without noticing — there is no error message, just a folder you stop being able to find things in. Keep it local and keep it in files and both numbers drop.
Is Napkin itself still around?
Yes, though the website has retreated. The iPhone app was rebuilt and shipped in May 2026, and the App Store listing says a subscription is required to use it. But the site's pricing page 404s, the old web app returns a server error, and the only price you can find is in the App Store. Alive, small, and iPhone-first.
Every week, someone ships something they shouldn’t have.
New verdicts, the worst thing that landed in the trap, and the occasional incident report. No other email, ever.
Semantic search over fifty notes is magic. Over five thousand it's a ranking problem you have no way to measure.
Your inbox is a password-reset machine. That is the credential store you just gave a weekend agent write access to.
RAG over your own files is a weekend. Then you change the embedding model and re-index four years of your life.
last reviewed 2026-08-05 · verdict is editorial and unsponsored · shared entry data from canivibecodeit under MIT · not legal advice