Inspect Supabase schema and logs via MCP
Connect Supabase MCP and ask for table shapes, RLS notes, and recent error logs.
Data·Easy·0 saves
Setup USE CASEDataSimple
“13,425 art prompts, extracted and sorted”
Julie McCann had a Grok Bot go through 557 chats across all her projects, extract every art prompt, and sort them by subject into one searchable library — 13,425 in total. When it hit context limits it switched to smaller blocks on its own and finished the ~30-hour job.
How it runs
5 steps. Do them in order.
- 01Point the bot at all of your projects and chats — the places your prompts are buried.
- 02Tell it exactly what to pull (here: every art prompt) and how to organize it — Julie had it sort by subject into one searchable file.
- 03Tell it up front to work in **smaller blocks** and checkpoint as it goes. Julie's run timed out twice before the bot switched to smaller, more manageable chunks on its own; asking for that from the start avoids the false starts.
- 04Let it run — a job this size took ~30 hours across 557 chats; check in and re-point it if its computer needs an update mid-run.
- 05Get back one compiled, sorted, searchable library you can mine for future work.
What you get
It's the boring superpower: an agent with the patience to walk 557 chats and turn years of scattered work into one searchable asset. The clever part is the failure handling — after timing out it shrank the job into blocks and kept going, which is the difference between a big extraction that finishes and one that dies halfway. Anyone sitting on a pile of past prompts, snippets, or decisions can point this at their own history.
Prompt
Copy it. Change the names. Keep the job.
Prompt
Go through all of my projects and chats and build me one searchable library of every prompt I've written. 1. Work project by project, chat by chat. Extract every prompt you find (for me: AI art prompts — the generation prompts, not the replies). 2. Work in SMALL blocks and checkpoint after each one. Do not try to hold everything at once — if a block is large, split it further. Save progress as you go so a timeout never loses work. 3. Normalize each entry: the prompt text, which project/chat it came from, and a subject tag. 4. Sort the whole set by subject and compile it into one searchable file (and tell me the total count). Rules: extract only what's actually there — never invent or "improve" a prompt. If you hit a limit, shrink the block and continue rather than stopping. Give me a progress note after each project so I can see how far along you are.
Use withGrok
via @juliezmccann · Tweet by @JulieZMcCann
Tips
- ·A Curator reconstruction of Julie's task. To adapt: point the bot at your own projects, tell it what to extract, and — the key lesson — tell it to work in smaller blocks and checkpoint so timeouts don't lose progress on a long job.
- ·none
- ·readonly
Related setups
Same category. Different job.
Five reasons they left, five things to ship this week
After the win-back emails, Liam's bot analysed overnight why people left and turned that into a 5-step plan. Same tweet as the win-back campaign — this is the research half of the job.
Data·Moderate·0 saves
Setup An hourly brief that says nothing when nothing changed
From the same Gmail-connected setup, he created standing agents, including an hourly political update on his work topics and a rule to always check with him on key decisions. No posting. Cite sources. Stay quiet if nothing new.
Data·Moderate·0 saves
Setup