The 24 principles
pstack writes engineering experience as 24 short principles. When the AI drifts, saying a name pulls it back.
An analogy
Like a team's play calls. The coach shouts one, and everyone knows what to do.
In pstack
You don't call principles; /pstack:poteto-mode reads them at the start of every task. The names are for you, to correct course. Say it claims success just because it compiled. You reply: "apply prove it works. Run the real import flow and show me the records it wrote." Its reply should say which decision that changed. If it can't, it didn't really use it.
The author's advice: don't memorize them. Skim them now. When the AI makes a mistake a principle could have prevented, come back and look it up. That's how they stick.
All 24 (type the English name as-is so the AI recognizes it)
| Name | In plain words |
|---|---|
| laziness protocol | Core · Delete what you can; change as little as possible |
| foundational thinking | Core · Pick the right data shape, then write logic |
| redesign from first principles | Core · Redesign as if it was there from day one |
| attack the premise | Core · Two fixes failed? Test the premise |
| subtract before you add | Core · Remove what's useless, then add |
| minimize reader load | Core · Less to jump through, less to remember |
| outcome-oriented execution | Core · Go straight to the goal; no stopgap code |
| experience first | Core · User experience first; fewer, better features |
| exhaust the design space | Core · No precedent? Build a few prototypes and compare |
| build the lever | Core · Build a tool instead of repeating by hand |
| model the domain | Architecture · Keep business rules in one place, not scattered ifs |
| boundary discipline | Architecture · Check outside data where it comes in |
| type system discipline | Architecture · Make wrong states impossible to write |
| make operations idempotent | Architecture · Half-run or run twice, same result |
| migrate callers then delete legacy apis | Architecture · Move the callers, then delete the old API |
| separate before serializing shared state | Architecture · Stop sharing first; lock only if you must |
| prove it works | Verification · Prove it with something real; compiling doesn't count |
| fix root causes | Verification · Fix where things first went wrong |
| sequence verifiable units | Verification · Small pieces; check each before the next |
| test behavior not implementation | Verification · Test what users can see |
| explain the number | Verification · Find out what's limiting the number |
| guard the context window | Delegation · Heavy reading goes to helper AIs |
| never block on the human | Delegation · Do what can be undone; ask before what can't |
| encode lessons in structure | Meta · Said the same reminder twice? Make it a check |
Common misconceptions
- "You have to call the principles one by one": no,
/pstack:poteto-modereads them every time. The names are for you to correct course. - "If I say a principle's name, the AI follows it": check whether its reply says which decision that principle changed.
Try it: Pull it back on track in one sentence
Related: /pstack:poteto-mode, Evidence (not "should be fixed")