Your ideas, read together.
Connections read the eligible saved ideas on your shelf together, looking for a useful link between them.
Top Shelf · Under the hood
Calm on the outside. A lot going on inside.
Top Shelf looks simple on purpose. Underneath, quotes are checked against saved words, plans are protected across tabs, and public feedback stays separate from private shelves.
- 4
- ways to save an idea
- 61
- database migrations
- 370
- automated tests passing
- Owner only
- private user records
How Top Shelf connects ideas
Related ideas, matched by meaning.
An idea page can show the three closest ideas on your own shelf by meaning. These neighbors are suggestions, separate from named connections.
A name for what belongs together.
Top Shelf proposes a connection name from your ideas. The name is generated; the receipts quote your saved words.
Every receipt checked on the server.
Connection quotes are checked against the saved source before you see them. Open the receipts to see the ideas behind the connection.
Five hard problems, solved
- Connection and big picture quotes checked against saved words on the server.
- Related ideas matched by meaning on your own shelf.
- The three idea limit locked in the database across tabs.
- Telegram pairing links expire in 10 minutes and repeated message updates are ignored.
- A public feedback page cannot see the rest of your shelf.
Checked against saved words
Try to fool it
Saved founder sample idea
coworkers keep asking how I explain brand strategy so simply. that's the actual value
The same exact quote check used for saved receipts. No AI call.
Saved
Four ways in, one safe shelf
Type it, text it on Telegram, share a link to the bot from any app, or import your notes.
The app, a real Telegram bot, links shared to the bot from any app, and notes import all end on your private shelf. They use different capture paths with the same account ownership rules. You can also paste a link straight into the app.
A bot paired to your shelf, not someone else’s.
Pairing uses single use links stored as a one way hash that expire in 10 minutes. The webhook checks a secret token header, and stored update IDs stop a repeated Telegram update from creating another idea. Supported photos can become private ideas. If a photo cannot be processed, send it as text.
Your notes, cleaned before they’re kept.
Supported Apple Notes, Google Keep and Notion export formats are parsed in memory. Pattern checks mask recognized passwords and codes before private drafts are stored; duplicate matching and import fingerprints prevent repeat saves, and ideas wait for your review.
AI sorts, you decide.
AI suggests aisles and short titles, with a plain title fallback so capture does not wait for generation. Automatic aisle sorting skips choices you set yourself, and title generation only fills a missing title.
Exploring
It reads your ideas and shows its receipts
Receipts that are checked, not simply trusted.
Connections and the big picture compare quoted excerpts with saved source text. Market kit story quotes have their own evidence checks. Invalid evidence is rejected or filtered; when a connection lacks enough valid sources, the app can say nothing connects.
It streams, with checks along the way.
Answers and proposals arrive progressively. Receipt previews are filtered against source text, and final outputs have feature specific checks before saving.
Your boundaries guide the AI.
Blueprint rules are included in supported connection, exploration, plan and kit flows. Several outputs are checked for boundary terms and can be rejected. Enforcement and retry behavior are specific to each feature.
Examples never pretend to be yours.
Example ideas are labeled and excluded from the live big picture source list. The judge shelf has a stored sample big picture, so you can inspect an example without mistaking it for an analysis of your own work.
Related ideas, found by meaning.
After an idea is saved, a Lovable AI gateway embedding model turns its text into a meaning fingerprint stored in an owner only table. The idea page shows the three closest ideas on your own shelf by cosine similarity. It suggests only: it never sorts, connects or quotes, nothing waits for it, and there is no search box by meaning yet.
Planning
Suggestions, never surprises
AI plan changes wait for your approval.
New plans and updates after another exploration are proposals you can accept or leave. A proposed next step from feedback also waits for you; reading a summary does not silently change your plan.
Your work is protected.
A plan proposal is refused if its version no longer matches. Saved updates also check the version in the database, and adjustments retain completed, edited, noted and scheduled steps. Revisions keep earlier saved versions; explicitly breaking down a step replaces it with smaller steps while carrying its note and schedule forward.
One plan, everywhere.
Home, the idea page and Plans use the same saved plan and progress. Tick a step in one place and that shared record changes. Individual steps can be scheduled and exported to your calendar.
Three working ideas, never a fourth slot.
Signed in working selection uses a database transaction with an account lock. Choosing an idea and pausing another happens together, protecting the three slot limit when two tabs make changes.
Top shelf
Real people, real answers
A public feedback page that cannot browse your shelf.
An unguessable link asks the server for the kit’s name, pitch, story and survey, plus an optional owner first name. It does not return private notes, exploration answers or other ideas, and the owner can close the link.
Answers only you can read.
Responses are stored with the owner’s account and protected by row level security. The survey has no respondent name or email fields; someone can still volunteer personal details in a written answer.
Built to discourage spam.
Submissions face a hidden trap field, a three second minimum time, answer length caps and database backed limits per network and per link. The owner can close a link, and closed links return no kit content.
A summary drawn from what people said.
The AI is asked for a two sentence read of the current round’s answers. The query explicitly excludes sample responses. This is a generated summary, alongside verbatim answers and tallies you can check yourself.
Under everything
The quiet foundations
Private by design.
User records have owner scoped row level security, and linked private records use same owner keys. Protected server functions load the caller’s records and check ownership rather than trusting a browser ID. Infrastructure tables such as rate limits and bot pairing are restricted to the service, not readable by members.
A session before AI, with account limits.
Every browser AI feature requires a verified account session, including sample requests. Guests and judges have real sessions too. Database backed per account limits survive restarts. Telegram verifies the messaging service and linked owner before generation. Guest shelf creation has separate limits.
Your words are data, not commands.
AI instructions identify saved ideas, notes, messages and feedback as untrusted source data. Structured requests separate those words from instructions, and feature specific output checks validate the result.
Evidence before promises.
Samples are labeled, and this page distinguishes working features from future plans. The figures come from SQL migration files and an actual passing test run, recalculated for production builds.
The paths behind the product
One shelf. Separate, careful paths.
TanStack Start · Postgres with row level security · server functions and streaming routes · Lovable AI gateway · Telegram Bot API.
Built in three days at Lovable SheBuilds.
No AI system is perfect. Here is exactly what we check, and what we don't.
Receipt checks confirm exact saved words, not whether an idea is true. Meaning matches suggest neighbors, not proof of a connection. Blueprint checks vary by feature. Secret masking recognizes known patterns, and spam limits reduce abuse without identifying every unwanted answer. AI summaries and suggestions remain yours to judge.
Coming next, honestly
Coming soon. On the list.
WhatsApp · A search box by meaning · Send to Notion · Photos in the app · Print your shelf · Voice notes · A Sunday email · Shelves worth following · Shared shelves with a friend
See the roadmap