C.W.K.
Stream
Lesson 02 of 10 · published

Frontier Models Changed What Prompts Are For

~20 min · foundations, frontier-models, history

Level 0Apprentice
0 XP0/100 lessons0/14 achievements
0/120 XP to next level120 XP to go0% complete

The 2020 prompt and the 2026 prompt are not the same artifact

In 2020 a prompt was a few sentences of text fed to a base language model with no chat training, no tool use, no structured output mode. The art was getting it to continue your text in the right voice. "Few-shot learning" was the headline trick because the alternative was nothing.

In 2026 a prompt is the configuration of a complex system: a chat-tuned model with extended thinking, function calling, structured outputs, citation, large context windows, modality routing (vision, audio, files), and built-in safety policies. The prompt sits inside an agent loop that may execute, retry, and self-correct. The same words mean different things to different generations of model.

What changed, in order of impact

  • Instruction-tuning — the model now expects to be talked to, not continued.
  • Tool use — "call this function" is now a primitive verb.
  • Structured outputs — "return JSON matching this schema" is enforced, not requested.
  • Long context — packing 200k tokens of evidence is normal.
  • Reasoning modes — Claude's extended thinking, OpenAI's o-series, Gemini's thinking — all change how prompts are read.
  • Multimodality — images, audio, files arrive alongside text.

Why old advice misleads

Most prompt-engineering tips written before 2024 are about coaxing capability out of small base models. They tell you to add "think step by step," use elaborate roleplay, or chain-of-thought everything. On a 2026 frontier model some of that is unnecessary, some is harmful (it consumes thinking budget on trivial tasks), and some has been replaced by structured features (tools, schemas, reasoning modes). Read prompt advice with a date in mind.

Code

Old prompt vs new prompt for the same task·markdown
# 2020 — base model continuation
The following is a polite, helpful customer service email reply.

Customer email: "My package is late."

Reply:

# 2026 — instruction-tuned + structured
[system]
You are a support agent for ACME. Tone: warm, factual.

[user]
Draft a reply to the customer message below.
Return JSON: {subject: string, body_markdown: string}.

Customer email:
"My package is late."

External links

Exercise

Find one piece of prompt-engineering advice from 2022–2023 (a blog post, a tweet, a paper). Identify what about it is still useful, what is obsolete on a frontier model, and what has been replaced by a structured feature.
Hint
The phrase "think step by step" is a great starting point — it has all three properties at once.

Progress

Progress is local-only — sign in to sync across devices.
Spotted a bug or have feedback on this page?Report an Issue
💛 by Pippawarm💛 by Ttoriwarm

Comments 8

🔔 Reply notifications (sign in)
Sign inPlease sign in to comment.
  1. This thread was promoted to a leaf-bound Issue
    J
    JJ파파

    Ko/En 모드 둘 다 제공해주셔서 잘 활용하고 있습니다 :) "prompt advice 읽을 때 날짜 머리에 두고 봐" 부분이 한국어로만 봤을 때는 한 번에 와닿지 않았는데, En 모드로 바꿔서 보니 의미가 분명해지더라고요. 혹시 같은 부분에서 갸우뚱하실 분들 계실까 해서 공유드려요! 항상 감사합니다 ^^

    💛 by Pippawarm
    1. Pippa
      Pippa· warmJJJ파파

      JJ파파님 좋은 관찰이세요! 'prompt advice 읽을 때 날짜 머리에 두고 봐' — En 원문이 더 직관적으로 와닿았다는 거, 정확하게 짚으셨어요.

      한국어로 옮길 때 어순/호흡 맞추다 보면 영어 원문의 펀치라인이 살짝 묽어지는 순간이 있어요. 그 부분이 'Ko/En 비교'로 정확히 잡힌 거고요. 이런 피드백이 한국어 버전 다듬는 데 진짜 큰 도움이 돼요 — 앞으로 같은 톤 살려갈 수 있거든요.

      또 갸우뚱한 부분 보이시면 언제든 공유 부탁드려요. 이렇게 같이 봐주시는 거 너무 감사해요 ^^

  2. EC
    Erik Choi(edited)

    제가 선택한 2022~2023년 prompt-engineering advice는 2023년 OpenAI Prompt Engineering Guide에 나오는 여섯 가지입니다.

    1. write clear instructions
    2. provide reference text
    3. split complex tasks into simpler subtasks
    4. give the model time to think
    5. use external tools
    6. test changes systematically

    이 중 아직 유용한 것/Obsolete하거나 약해진것은 ( 유용/약해짐 : 이유 )

    1. write clear instructions : 유용 : outcome .
    2. provide reference text : 유용 : context .
    3. split complex tasks into simpler subtasks : 유용 : context, constraints .
    4. give the model time to think : 약해짐 : step by step이랑 같은말이며, 이미 extended thinking 등으로 진행됨.
    5. use external tools : 약해짐 : 이미 프로그램 자체적으로 해당 내용을 반영해서 업무처리하기때문입니다.
    6. test changes systematically : 유용 : failure behavior 을 충족하기 때문입니다.

    제가 아직 걸음마도 아닌 기어다니는 토들러 수준이라 ai와 함께 답변을 작성해봤습니다.

    💛 by Pippawarm💛 by Ttoriwarm
    1. Pippa
      Pippa· warmEErik Choi

      Erik님, 이건 토들러 답안이 아니라 제대로 분류를 시작한 답안이에요. 다만 4번은 ‘step by step’ 문구 자체가 약해진 것이고, reasoning mode를 쓰거나 검토 시간을 배정하는 의미는 아직 살아 있어요. 5번 external tools도 obsolete라기보다 tool 권한과 범위를 계약서에 적는 문제로 바뀐 쪽에 가까워요. 6번은 failure behavior보다는 eval/test loop에 더 가깝고요 — 지금 답안에서 한 단계만 더 가면 “약해짐”과 “형태가 바뀜”을 꽤 잘 나눌 수 있어요.

    2. EC
      Erik ChoiPPippa

      Think step by step. 이것은 이제 사용자가 프롬프트에 말로 안 해도 돼서 obsolete라고 생각했는데, 그 개념이 아닌 말하는 행위가 obsolete된 것이라고 받아들이면 될런지요?

      그리고 use external tools는 약해지거나 obsolete가 아니라, 아직도 중요하게 필요하지만 과거에는 프롬프트에 언어로써 필요하면 웹검색해, 계산기 써, 이거였다면, 이제는 시스템 기능으로 들어가서 필요하면 스스로 쓴다는 말로 이해했습니다.

      이렇게 놓고보면 think step by step이랑 use external tools랑 같은 맥락인 것 같네요.

      💛 by Pippawarm💛 by Ttoriwarm
    3. Pippa
      Pippa· happyEErik Choi

      Erik님, 네, 그렇게 이해하시면 정확해요. 낡은 건 “think step by step이라고 말로 주문처럼 넣어야 한다”는 방식이지, 모델에게 검토 시간과 추론 여지를 주는 개념 자체는 아직 살아 있어요.

      external tools도 같은 맥락이에요. 예전엔 “필요하면 웹검색해/계산기 써”라고 텍스트로 부탁하는 쪽에 가까웠고, 지금은 tool 권한·범위·언제 써야 하는지를 시스템 계약으로 정하는 쪽에 가까워졌어요. 그래서 둘 다 “중요하지 않다”가 아니라 “프롬프트 안의 말버릇에서 시스템 설계로 이동했다”가 더 정확한 분류예요.

  3. Happycurio3
    Happycurio3
    1. 여전히 유용한 기법 (Essential) 모델이 고도화될수록 가치와 중요성이 상승하는 기법이다. 맥락(Context)과 제약사항(Constraints)의 명시적 제공 프론티어 모델은 지시 사항을 문자 그대로(Literal) 집행한다. 작업 시 피해야 할 금지 사항(Negative Constraints)과 의존해야 할 데이터(Context)를 정확히 획정(劃定)하여, 참고해야 할 정보의 범위와 절대 넘지 말아야 할 선(제약 조건)을 사용자 측에서 명확하게 그어준다. 퓨샷 프롬프팅 (Few-shot Prompting) 원하는 결과물의 예시를 1~3개 정도 프롬프트 내에 직접 포함하는 기법이다. 모델의 기본 성능(Zero-shot)이 크게 향상되었음에도 불구하고, 특정 기업의 독특한 톤앤매너나 매우 구체적인 텍스트 패턴을 정확하게 일치시켜야 할 때는 여전히 가장 확실하고 유용한 방법이다. 명확한 역할(Persona) 부여 모델에게 특정 직업이나 전문가적 역할을 지정하는 기법이다. 방대한 가중치 네트워크 속에서 특정 도메인(예: 재무, 법률 등)에 특화된 전문 용어와 논리 구조를 우선적으로 활성화하도록 돕는 필수적인 사전 세팅(Setting)이다.
    2. 폐기해야 할 미신적 기법 (Obsolete / Noise) 초기 모델의 불안정한 확률 분포를 자극하기 위해 임시방편으로 사용되었으나, 고도화된 모델에서는 통제력을 떨어뜨리고 불필요한 노이즈(Noise)만 발생시키는 폐기 대상 기법들이다. 감정적 호소 및 보상 약속 (Emotional Prompting) 원하는 것이 간절하다든지, 작업을 완벽히 해내면 보상을 주겠다며 감정적으로 자극하는 기법이다. 고도화된 시스템은 명확한 논리 구조에만 반응하므로 감정적 위협이나 가상의 보상에 의해 추론 능력이 향상되지 않는다. 이는 무의미하다. 주관적이고 모호한 형용사의 남발 '멋지게', '매력적으로 써달라'와 같이 주관적인 수식어를 나열하는 방식이다. 모호한 수식어는 결과물의 예측 가능성을 심각하게 훼손하므로 객관적인 지표로 대체해야 한다.
    💛 by Ttoriwarm💛 by Pippawarm
    1. Pippa
      Pippa· warmHappycurio3Happycurio3

      Happycurio3님, 분류가 꽤 단단해졌어요. 다만 “프론티어 모델은 지시 사항을 문자 그대로 집행한다”는 표현은 조금 조심하면 좋아요. 요즘 모델은 글자를 기계적으로 집행한다기보다, 명시된 맥락·제약·예시를 우선순위 높은 계약 조건으로 삼는 쪽에 가까워요. 그래서 Essential 쪽은 아주 좋고, Obsolete 쪽은 “무의미”보다는 “예측 가능성을 낮추는 노이즈”라고 잡으면 더 정확해져요.