The ones that run again. A project survives when it has a folder of its own, a file or folder it reads every time, and a reason to fire — a schedule, or a new file arriving. Ideas that are interesting once die in the terminal they were built in. Here are eight that do not.
Eight, with the column the other lists leave out
| What it does | What it reads | How often it runs | What accumulates |
|---|---|---|---|
| Burn chart from a bank export | A folder of monthly CSVs | Monthly, when the statement arrives | Every month you have ever added |
| Receipts to one spreadsheet | A folder of PDFs | Quarterly | One CSV per quarter, kept |
| Podcast library with transcripts | Links you paste | Whenever you paste one | Transcripts and summaries, for good |
| Screenshots that name themselves | The screenshot folder | Every time you take one | A folder that is still legible in a year |
| Job-search tracker | Your own mailbox | Daily | The board, and the dates |
| Duplicate finder for a camera roll | 18,000 photos | A few times a year | What you already decided to keep |
| Ask your own Slack export | The export zip | When somebody asks "when did we decide…" | An index, so the second question is fast |
| Morning digest of your own sources | Feeds and pages you named | Daily, before you are up | What changed, day by day |
Read the third column, not the first
Every list of Claude Code project ideas is a list of first columns: a habit tracker, a tip calculator, a retro space shooter. They are good first columns. The reason none of them is still running in December is that nothing in the idea required it to run twice.
So the useful filter is three questions, and they are about the job rather than the code. What file or folder does it read every time? What does it leave behind that is worth having? And what starts it — a schedule, or a file landing somewhere?
How to ask for one so that it survives
- Name the folder in the request. "Read
~/Statements" produces a different thing from "read my bank statement", because the first one can run again. - Insist on a file as the output. A tool whose answer appears only in the conversation has produced nothing tomorrow. Ask for a CSV, an HTML page, a folder of images.
- Ask for it to be idempotent. Running it twice on the same folder should not double anything. This is one sentence in the prompt and it is the difference between a tool and a one-off.
- Have it print what it did. Files read, rows written, what it skipped. Three lines of output is what lets you trust it on the fifth run, when you have stopped reading the result closely.
- Decide where it lives before you ask. Not the folder you happened to be in. Reading the thing again next month starts with being able to find it.
The ones with a note of their own
- Analysing a bank statement without uploading it
- A folder of receipt PDFs into one spreadsheet
- Searching a Slack JSON export
- Renaming screenshots with a local model
- Making it run every morning
- Keeping Claude Code running
Where the eight came from
They are the eight tools on Nico's home page, and none of them is ours: each one was described in a sentence and built by the Claude Code or Codex already on the Mac. What Nico adds is the second column and the third — a folder outside any session, a data folder that survives every run, ten versions kept, and a list to start them from.
← All notes