Prompts I Keep Coming Back To
Living document - last updated 2026-05-21. Started today. I add to this when a prompt earns its place.
What this is
A commonplace book of prompts.
Not prompts I wrote - prompts I’ve come across in other people’s posts, threads, repos, and talks, and found myself reaching for over and over. The kind that earn a spot in muscle memory because they keep producing useful output for the work I actually do.
This page exists for one reason: I kept losing them. See a great prompt, use it once, need it again a month later, reconstruct it badly from memory. From now on, when one earns its third use, it lives here.
Where I know the source, I credit it. Where I don’t, the source line stays open until I rediscover it. Each entry is just the prompt itself - copy-pasteable - and, when it helps, a short note on why I like it.
Quick reference
| Prompt | What it’s for | Section |
|---|---|---|
| ultrathink | Trigger deeper reasoning from Claude | Magic words |
| diary | Surface step-by-step reasoning instead of compressed summary | Magic words |
| great textbook | Counter GPT-5’s tendency to be cryptic | Magic words |
| BURN THE BOATS | Stop agents from doing unnecessary backward-compat work | Magic words |
| Project kickoff interview | Get past surface requirements at the start of a project | Full prompts |
Magic words
Short phrases I drop into prompts to nudge model behavior. From a friend at DeepMind: reasoning models aren’t only better because they “reason” - they’re also better because outputting more tokens means more compute budget. So anything that makes a model think longer or output more helps, to a degree. These are user-triggered versions of that effect.
ultrathink
A trigger for deeper reasoning. Drop it into any Claude prompt where you want expanded thinking - Claude treats it as a signal to spend more compute budget on the response.
plan this refactor - ultrathinkdiary
Forces step-by-step thinking. Makes uncertainty explicit. Prevents the model from collapsing everything into a dense summary.
Instead of jumping straight to a compressed explanation, the model has to walk through how understanding evolved over time: observations, hypotheses, dead ends, and conclusions. A side effect I’ve noticed: the reasoning trace also surfaces loopholes in conclusions before reaching them.
explain this bug as a diary entrygreat textbook
For countering GPT-5’s tendency to be cryptic. Forces pedagogical, clear, build-up-understanding-from-zero explanations instead of dense compressed summaries.
I discovered this one going through a bug. The agent’s summary was technically correct but unreadable. The reframing produced a clean walk-through I could actually hand to a teammate.
write an explanation of this bug, like a great textbookBURN THE BOATS
For when agents do unnecessary backward-compatibility work. Drops the legacy-preservation instinct and lets the model rewrite freely.
refactor this module - BURN THE BOATSI’ve used it four or five times in 2025. Works every time.
Full prompts
Project kickoff interview
I am about to start this project, "PROJECT NAME". You have all the
context using the hopocalypse MCP server. Interview me until you have
95% confidence about what I actually want, not what I think I should
want.The load-bearing line is “what I actually want, not what I think I should want.” Without it, the model interviews for surface requirements I’ve already articulated. With it, the interview interrogates the framing I came in with — which is usually where the assumptions are hiding.
The “hopocalypse MCP server” line is just how I load context. Swap it for whatever fits your setup:
read the README at @/path/to/spec.mdpull issue #42 using gh clihere's the design doc, attacheduse the Atlassian MCP to read the Confluence page titled "X"- or just paste the context inline before this prompt
Or drop the context line entirely if the model already has what it needs. The interviewing pattern is the part that matters.
On my radar
Half-remembered prompts I want to write up next. Placeholders until they get full entries above.
yaml dsl- a structural framing trick I keep using; need to articulate the load-bearing parts.for a new intern- a thoroughness trigger; need to write up when it does and doesn’t help.
How I add to this
A prompt earns a spot here after three uses and after I can articulate (even in one line) what makes it land. Until then it lives in On my radar above as a name, a thread link, or a vague memory of a phrase that worked.