Dayward AIWeChat Mini Program
Context Engineering in 5 Days
Week 1 · D2🔒 LockedAbout 3 hours

System Prompts and the Instruction Hierarchy: the Right Altitude, Persistent Instruction Files, Progressive Disclosure, Less Is More

A system prompt written too fine goes brittle; written too coarse, it drifts. Today find the altitude in between: organize the prompt into blocks, move long-lived rules into persistent instruction files, and let detail unfold on demand through progressive disclosure.

Today's goals 0/3

Sign in to tick these off and save your progress.

Today's Goals

  1. Judge whether a system prompt is written too fine or too coarse, and give a concrete fix for each
  2. Organize a system prompt into blocks, and say which content should move down into a persistent instruction file
  3. Iterate a system prompt with a minimal-start, failure-driven approach, and record the reason each rule was kept

Of the four pieces measured yesterday, the system prompt is the one resent verbatim every turn. What makes it special is compounding: change it once and every turn of the whole session benefits. So we start there. Come back when you have read it and tick off the three goals.

Plain-Language Walkthrough

Should the packing list go down to socks, or only name categories

Before a trip you write a packing list. There are two ways to write one, and each is bad in its own way.

The first is extremely fine: "two white shirts, the blue one on top; three pairs of socks, one of them athletic, tucked into the shoes; charger in the second slot of the left side pocket." The problem with that list is not that it is verbose but that it is brittle. Shorten the trip to four days and the whole list is void; the hotel offers laundry and half the entries lapse. Either you rewrite it every time or you pack from a list that is already wrong.

The second is extremely coarse: "bring enough clothes, bring the electronics, don't forget the important documents." That list never goes stale, because it says nothing. Pack from it and the result is identical to packing without it.

The system prompt is that list, and both bad ways of writing it are extremely common in real projects. The too-fine kind hard-codes business logic into natural language — branches that should be decided in code written out as a string of if-then clauses. The too-coarse kind is a passage that reads impeccably and changes no model behavior when deleted.

For the position in between, Anthropic's engineering blog gives a good word: altitude. Fly too low and you see every tree but not the road; fly too high and the whole forest is in view but you do not know which way to go. The right altitude is specific enough to steer behavior effectively and flexible enough to leave the model room to judge.

How do you tell what altitude you are flying at? There is a self-check that almost never fails: could you write an assertion that automatically checks whether this rule was followed?

  • You cannot ("the tone should be natural," "answers should be clear and well organized") — too high; this line is filler.
  • You can, but the assertion has to enumerate seven cases ("if the status is A say this, if B say that…") — too low; this belongs in code or a reference file.
  • You can, and the assertion is short ("no exclamation marks in a reply," "a cited policy must carry its clause number") — the altitude is right.

You will use that self-check exactly 43 times today, because everything below rests on judging each rule one at a time — and the overwhelming majority of teams' system prompts have never been judged rule by rule at all.

Sign in and link WeChat to read

Sign in, then link your WeChat mini program account on the account page. Linked accounts share one free article per day. Unlock additional articles with rewarded videos in the mini program.

Mini program QR code

Scan in WeChat to open Dayward AI. Open the same course and chapter, choose the rewarded video, then refresh this page after completion.

Interview questions

  • How do you calibrate the altitude of a system prompt, and what goes wrong at each extreme?系统提示的高度怎么把握?写太具体和写太笼统各会出什么问题?
    Common in ChinaCommon overseasIntermediate#system-prompt#altitude

    🔒 Unlock this chapter to see the answer keys.

  • What belongs in the system prompt, what belongs in a persistent instruction file, and what should be loaded on demand?哪些内容该进系统提示,哪些该进持久指令文件,哪些该按需加载?
    Common in ChinaCommon overseasIntermediate#instruction-hierarchy#progressive-disclosure

    🔒 Unlock this chapter to see the answer keys.

  • How do system prompts keep growing, and how would you stop it?系统提示越写越长是怎么发生的?你会怎么止住这个过程?
    Common in ChinaCommon overseasBasic#prompt-bloat#maintenance

    🔒 Unlock this chapter to see the answer keys.

Comments