[OpenAI] GPT-5.2 ํ”„๋กฌํ”„ํŒ… ๊ฐ€์ด๋“œ: ์—”ํ„ฐํ”„๋ผ์ด์ฆˆ AI ์—์ด์ „ํŠธ๋ฅผ ์œ„ํ•œ ์ตœ์ ํ™” ์ „๋žต

Posted by Euisuk's Dev Log on January 17, 2026

[OpenAI] GPT-5.2 ํ”„๋กฌํ”„ํŒ… ๊ฐ€์ด๋“œ: ์—”ํ„ฐํ”„๋ผ์ด์ฆˆ AI ์—์ด์ „ํŠธ๋ฅผ ์œ„ํ•œ ์ตœ์ ํ™” ์ „๋žต

https://cookbook.openai.com/examples/gpt-5/gpt-5-2_prompting_guide

์„œ๋ก 

OpenAI์˜ GPT-5.2๋Š” ์—”ํ„ฐํ”„๋ผ์ด์ฆˆ ํ™˜๊ฒฝ๊ณผ ์—์ด์ „ํŠธ ์›Œํฌํ”Œ๋กœ์šฐ๋ฅผ ์œ„ํ•ด ์„ค๊ณ„๋œ ์ตœ์‹  ํ”Œ๋ž˜๊ทธ์‹ญ ๋ชจ๋ธ์ž…๋‹ˆ๋‹ค. ์ด์ „ ์„ธ๋Œ€์ธ GPT-5.1 ๋Œ€๋น„ ํ† ํฐ ํšจ์œจ์„ฑ, ์ถœ๋ ฅ ํฌ๋งทํŒ…, ๊ตฌ์กฐ์  ์ถ”๋ก  ๋Šฅ๋ ฅ์—์„œ ๋ช…ํ™•ํ•œ ๊ฐœ์„ ์„ ๋ณด์—ฌ์ฃผ๋ฉฐ, ํŠนํžˆ Tool Calling๊ณผ ๋ฉ€ํ‹ฐ๋ชจ๋‹ฌ ์ดํ•ด ๋Šฅ๋ ฅ์ด ๊ฐ•ํ™”๋˜์—ˆ์Šต๋‹ˆ๋‹ค.

์ด ๊ฐ€์ด๋“œ์—์„œ๋Š” GPT-5.2์˜ ํ•ต์‹ฌ ํŠน์„ฑ ๋ณ€ํ™”๋ฅผ ์‚ดํŽด๋ณด๊ณ , ์‹ค์ œ ํ”„๋กœ๋•์…˜ ํ™˜๊ฒฝ์—์„œ ์„ฑ๋Šฅ์„ ๊ทน๋Œ€ํ™”ํ•˜๊ธฐ ์œ„ํ•œ ํ”„๋กฌํ”„ํŒ… ํŒจํ„ด๋“ค์„ ๋ถ„์„ํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ ๊ธฐ์กด GPT-5/5.1 ํ”„๋กฌํ”„ํŠธ๋ฅผ GPT-5.2๋กœ ๋งˆ์ด๊ทธ๋ ˆ์ด์…˜ํ•˜๋Š” ์ „๋žต๊ณผ, ์žฅ๊ธฐ ์ปจํ…์ŠคํŠธ ์ฒ˜๋ฆฌ๋ฅผ ์œ„ํ•œ Compaction ๊ธฐ๋Šฅ๋„ ๋‹ค๋ฃน๋‹ˆ๋‹ค.


GPT-5.2์˜ ํ•ต์‹ฌ ํ–‰๋™ ๋ณ€ํ™”

GPT-5.2๋Š” ์ด์ „ ์„ธ๋Œ€ ๋ชจ๋ธ๋“ค๊ณผ ๋น„๊ตํ–ˆ์„ ๋•Œ ๋‹ค์Œ๊ณผ ๊ฐ™์€ ํŠน์„ฑ ๋ณ€ํ™”๋ฅผ ๋ณด์ž…๋‹ˆ๋‹ค.

๋” ์ฒด๊ณ„์ ์ธ ์Šค์บํด๋”ฉ(Scaffolding): GPT-5.2๋Š” ๊ธฐ๋ณธ์ ์œผ๋กœ ๋” ๋ช…ํ™•ํ•œ ๊ณ„ํš๊ณผ ์ค‘๊ฐ„ ๊ตฌ์กฐ๋ฅผ ์ƒ์„ฑํ•ฉ๋‹ˆ๋‹ค. ์ด๋Š” ๋ณต์žกํ•œ ํƒœ์Šคํฌ์—์„œ ์žฅ์ ์ด ๋˜์ง€๋งŒ, ๋ช…์‹œ์ ์ธ ๋ฒ”์œ„(scope)์™€ ์ƒ์„ธ๋„(verbosity) ์ œ์•ฝ์„ ํ•จ๊ป˜ ์ œ๊ณตํ•  ๋•Œ ์ตœ์ ์˜ ๊ฒฐ๊ณผ๋ฅผ ์–ป์„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.

์ „๋ฐ˜์ ์œผ๋กœ ๋‚ฎ์•„์ง„ ์ƒ์„ธ๋„: ์ถœ๋ ฅ์ด ๋” ๊ฐ„๊ฒฐํ•˜๊ณ  ํƒœ์Šคํฌ ์ค‘์‹ฌ์ ์ž…๋‹ˆ๋‹ค. ๋‹ค๋งŒ ์—ฌ์ „ํžˆ ํ”„๋กฌํ”„ํŠธ์— ๋ฏผ๊ฐํ•˜๊ฒŒ ๋ฐ˜์‘ํ•˜๋ฏ€๋กœ, ์›ํ•˜๋Š” ์ƒ์„ธ๋„ ์ˆ˜์ค€์„ ํ”„๋กฌํ”„ํŠธ์—์„œ ๋ช…ํ™•ํžˆ ์ง€์ •ํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค.

๊ฐ•ํ™”๋œ ์ง€์‹œ ์ค€์ˆ˜: ์‚ฌ์šฉ์ž ์˜๋„๋กœ๋ถ€ํ„ฐ์˜ ์ดํƒˆ(drift)์ด ์ค„์—ˆ์œผ๋ฉฐ, ํฌ๋งทํŒ…๊ณผ ๊ทผ๊ฑฐ ์ œ์‹œ๊ฐ€ ๊ฐœ์„ ๋˜์—ˆ์Šต๋‹ˆ๋‹ค.

Tool Calling ํšจ์œจ์„ฑ์˜ Trade-off: GPT-5.1 ๋Œ€๋น„ ์ธํ„ฐ๋ž™ํ‹ฐ๋ธŒ ํ”Œ๋กœ์šฐ์—์„œ ์ถ”๊ฐ€์ ์ธ Tool Action์„ ์ทจํ•˜๋Š” ๊ฒฝํ–ฅ์ด ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋Š” ํ”„๋กฌํ”„ํŒ…์„ ํ†ตํ•ด ์ตœ์ ํ™”ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.

๋ณด์ˆ˜์ ์ธ ๊ทธ๋ผ์šด๋”ฉ ํŽธํ–ฅ: ์ •ํ™•์„ฑ๊ณผ ๋ช…์‹œ์  ์ถ”๋ก ์„ ์„ ํ˜ธํ•˜๋Š” ๊ฒฝํ–ฅ์ด ์žˆ์œผ๋ฉฐ, ๋ชจํ˜ธํ•œ ์ƒํ™ฉ์—์„œ๋Š” ๋ช…ํ™•ํ™”(clarification) ํ”„๋กฌํ”„ํŠธ๋ฅผ ์ œ๊ณตํ•  ๋•Œ ์ฒ˜๋ฆฌ ๋Šฅ๋ ฅ์ด ํ–ฅ์ƒ๋ฉ๋‹ˆ๋‹ค.


ํ”„๋กฌํ”„ํŒ… ํŒจํ„ด

์ƒ์„ธ๋„์™€ ์ถœ๋ ฅ ํ˜•ํƒœ ์ œ์–ด

์—”ํ„ฐํ”„๋ผ์ด์ฆˆ ํ™˜๊ฒฝ๊ณผ ์ฝ”๋”ฉ ์—์ด์ „ํŠธ์—์„œ๋Š” ๋ช…ํ™•ํ•˜๊ณ  ๊ตฌ์ฒด์ ์ธ ๊ธธ์ด ์ œ์•ฝ์„ ์ œ๊ณตํ•˜๋Š” ๊ฒƒ์ด ์ค‘์š”ํ•ฉ๋‹ˆ๋‹ค. ๋‹ค์Œ์€ ์›ํ•˜๋Š” ์ƒ์„ธ๋„ ์ˆ˜์ค€์— ๋”ฐ๋ผ ์กฐ์ •ํ•  ์ˆ˜ ์žˆ๋Š” ์˜ˆ์‹œ์ž…๋‹ˆ๋‹ค.

1
2
3
4
5
6
7
8
9
10
<output_verbosity_spec>
- Default: 3โ€“6 sentences or โ‰ค5 bullets for typical answers.
- For simple "yes/no + short explanation" questions: โ‰ค2 sentences.
- For complex multi-step or multi-file tasks: 
  - 1 short overview paragraph
  - then โ‰ค5 bullets tagged: What changed, Where, Risks, Next steps, Open questions.
- Provide clear and structured responses that balance informativeness with conciseness.
- Avoid long narrative paragraphs; prefer compact bullets and short sections.
- Do not rephrase the user's request unless it changes semantics.
</output_verbosity_spec>

์ด ํŒจํ„ด์˜ ํ•ต์‹ฌ์€ ์ƒํ™ฉ๋ณ„ ๊ธฐ๋Œ€์น˜๋ฅผ ๋ช…์‹œํ•˜๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. ๋‹จ์ˆœ ์งˆ๋ฌธ์—๋Š” 2๋ฌธ์žฅ ์ดํ•˜, ๋ณต์žกํ•œ ๋ฉ€ํ‹ฐ์Šคํ… ํƒœ์Šคํฌ์—๋Š” ๊ฐœ์š” + 5๊ฐœ ์ดํ•˜์˜ ํƒœ๊ทธ๋œ ๋ถˆ๋ฆฟ ํฌ์ธํŠธ๋ผ๋Š” ์‹์œผ๋กœ ๊ตฌ์ฒด์ ์ธ ํ˜•์‹์„ ์ง€์ •ํ•ฉ๋‹ˆ๋‹ค.

๋ฒ”์œ„ ์ดํƒˆ(Scope Drift) ๋ฐฉ์ง€

GPT-5.2๋Š” ๊ตฌ์กฐํ™”๋œ ์ฝ”๋“œ ์ƒ์„ฑ์— ๊ฐ•ํ•˜์ง€๋งŒ, ํ”„๋ก ํŠธ์—”๋“œ ํƒœ์Šคํฌ์—์„œ ์ตœ์†Œํ•œ์˜ UX ๋ช…์„ธ๋‚˜ ๋””์ž์ธ ์‹œ์Šคํ…œ์„ ๋„˜์–ด์„œ๋Š” ์ฝ”๋“œ๋ฅผ ์ƒ์„ฑํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋ฒ”์œ„ ๋‚ด์— ๋จธ๋ฌผ๊ธฐ ์œ„ํ•ด ์ถ”๊ฐ€ ๊ธฐ๋Šฅ๊ณผ ๋น„์ œ์–ด ์Šคํƒ€์ผ๋ง์„ ๋ช…์‹œ์ ์œผ๋กœ ๊ธˆ์ง€ํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค.

1
2
3
4
5
6
7
8
<design_and_scope_constraints>
- Explore any existing design systems and understand it deeply. 
- Implement EXACTLY and ONLY what the user requests.
- No extra features, no added components, no UX embellishments.
- Style aligned to the design system at hand. 
- Do NOT invent colors, shadows, tokens, animations, or new UI elements, unless requested.
- If any instruction is ambiguous, choose the simplest valid interpretation.
</design_and_scope_constraints>

๊ธฐ์กด GPT-5.1์˜ <design_system_enforcement> ๋ธ”๋ก์„ ์‚ฌ์šฉํ•˜๋Š” ๊ฒฝ์šฐ, โ€œno extra featuresโ€์™€ โ€œtokens-only colorsโ€ ์กฐํ•ญ์„ ์ถ”๊ฐ€ํ•˜์—ฌ ๊ฐ•์กฐํ•˜๋Š” ๊ฒƒ์ด ํšจ๊ณผ์ ์ž…๋‹ˆ๋‹ค.

์žฅ๋ฌธ ์ปจํ…์ŠคํŠธ์™€ ๋ฆฌ์ฝœ ๊ฐœ์„ 

10k ํ† ํฐ ์ด์ƒ์˜ ๊ธด ์ž…๋ ฅ(์—ฌ๋Ÿฌ ์ฑ•ํ„ฐ์˜ ๋ฌธ์„œ, ๊ธด ์Šค๋ ˆ๋“œ, ๋‹ค์ˆ˜์˜ PDF ๋“ฑ)์„ ์ฒ˜๋ฆฌํ•  ๋•Œ๋Š” ๊ฐ•์ œ ์š”์•ฝ๊ณผ ์žฌ๊ทธ๋ผ์šด๋”ฉ(re-grounding) ํŒจํ„ด์ด ์œ ์šฉํ•ฉ๋‹ˆ๋‹ค. ์ด ํŒจํ„ด์€ ๊ธด ์ปจํ…์ŠคํŠธ์—์„œ ์ •๋ณด๋ฅผ ๋†“์น˜๋Š” โ€œlost in the scrollโ€ ์˜ค๋ฅ˜๋ฅผ ์ค„์ด๊ณ  ๋ฆฌ์ฝœ์„ ๊ฐœ์„ ํ•ฉ๋‹ˆ๋‹ค.

1
2
3
4
5
6
7
8
9
<long_context_handling>
- For inputs longer than ~10k tokens (multi-chapter docs, long threads, multiple PDFs):
  - First, produce a short internal outline of the key sections relevant to the user's request.
  - Re-state the user's constraints explicitly (e.g., jurisdiction, date range, product, team) 
    before answering.
  - In your answer, anchor claims to sections ("In the 'Data Retention' sectionโ€ฆ") 
    rather than speaking generically.
- If the answer depends on fine details (dates, thresholds, clauses), quote or paraphrase them.
</long_context_handling>

์ด ์ ‘๊ทผ๋ฒ•์˜ ํ•ต์‹ฌ์€ ์„ธ ๋‹จ๊ณ„๋กœ ๊ตฌ์„ฑ๋ฉ๋‹ˆ๋‹ค. ์ฒซ์งธ, ๊ด€๋ จ ์„น์…˜์˜ ๋‚ด๋ถ€ ์•„์›ƒ๋ผ์ธ์„ ๋จผ์ € ์ƒ์„ฑํ•ฉ๋‹ˆ๋‹ค. ๋‘˜์งธ, ์‚ฌ์šฉ์ž์˜ ์ œ์•ฝ ์กฐ๊ฑด์„ ๋‹ต๋ณ€ ์ „์— ๋ช…์‹œ์ ์œผ๋กœ ์žฌ์ง„์ˆ ํ•ฉ๋‹ˆ๋‹ค. ์…‹์งธ, ์ฃผ์žฅ์„ ๊ตฌ์ฒด์ ์ธ ์„น์…˜์— ์•ต์ปค๋งํ•˜์—ฌ ์ผ๋ฐ˜์ ์ธ ์ง„์ˆ ์„ ํ”ผํ•ฉ๋‹ˆ๋‹ค.

๋ชจํ˜ธ์„ฑ ์ฒ˜๋ฆฌ์™€ ํ™˜๊ฐ(Hallucination) ์œ„ํ—˜ ์™„ํ™”

๋ถˆ๋ช…ํ™•ํ•œ ์š”๊ตฌ์‚ฌํ•ญ, ๋ˆ„๋ฝ๋œ ์ œ์•ฝ ์กฐ๊ฑด, ๋˜๋Š” ์ตœ์‹  ๋ฐ์ดํ„ฐ๊ฐ€ ํ•„์š”ํ•˜์ง€๋งŒ Tool์ด ํ˜ธ์ถœ๋˜์ง€ ์•Š๋Š” ์ฟผ๋ฆฌ์—์„œ ๊ณผ์‹ (overconfident) ํ™˜๊ฐ์ด ๋ฐœ์ƒํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.

1
2
3
4
5
6
7
8
9
10
11
<uncertainty_and_ambiguity>
- If the question is ambiguous or underspecified, explicitly call this out and:
  - Ask up to 1โ€“3 precise clarifying questions, OR
  - Present 2โ€“3 plausible interpretations with clearly labeled assumptions.
- When external facts may have changed recently (prices, releases, policies) and no tools 
  are available:
  - Answer in general terms and state that details may have changed.
- Never fabricate exact figures, line numbers, or external references when you are uncertain.
- When you are unsure, prefer language like "Based on the provided contextโ€ฆ" 
  instead of absolute claims.
</uncertainty_and_ambiguity>

๋ฒ•๋ฅ , ๊ธˆ์œต, ์ปดํ”Œ๋ผ์ด์–ธ์Šค, ์•ˆ์ „ ๊ด€๋ จ ์ปจํ…์ŠคํŠธ์™€ ๊ฐ™์€ ๊ณ ์œ„ํ—˜ ์ถœ๋ ฅ์—๋Š” ์ถ”๊ฐ€์ ์ธ ์ž๊ธฐ ์ ๊ฒ€(self-check) ๋‹จ๊ณ„๋ฅผ ์‚ฝ์ž…ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.

1
2
3
4
5
6
7
8
<high_risk_self_check>
Before finalizing an answer in legal, financial, compliance, or safety-sensitive contexts:
- Briefly re-scan your own answer for:
  - Unstated assumptions,
  - Specific numbers or claims not grounded in context,
  - Overly strong language ("always," "guaranteed," etc.).
- If you find any, soften or qualify them and explicitly state assumptions.
</high_risk_self_check>

Compaction: ์œ ํšจ ์ปจํ…์ŠคํŠธ ํ™•์žฅ

Tool์„ ๋งŽ์ด ์‚ฌ์šฉํ•˜๋Š” ์žฅ๊ธฐ ์‹คํ–‰ ์›Œํฌํ”Œ๋กœ์šฐ๊ฐ€ ํ‘œ์ค€ ์ปจํ…์ŠคํŠธ ์œˆ๋„์šฐ๋ฅผ ์ดˆ๊ณผํ•  ๋•Œ, GPT-5.2 with Reasoning์€ /responses/compact ์—”๋“œํฌ์ธํŠธ๋ฅผ ํ†ตํ•œ ์‘๋‹ต Compaction์„ ์ง€์›ํ•ฉ๋‹ˆ๋‹ค.

Compaction์€ ์ด์ „ ๋Œ€ํ™” ์ƒํƒœ์— ๋Œ€ํ•ด ์†์‹ค ์ธ์‹(loss-aware) ์••์ถ•์„ ์ˆ˜ํ–‰ํ•˜์—ฌ, ํƒœ์Šคํฌ ๊ด€๋ จ ์ •๋ณด๋ฅผ ๋ณด์กดํ•˜๋ฉด์„œ ํ† ํฐ ํ’‹ํ”„๋ฆฐํŠธ๋ฅผ ๋Œ€ํญ ์ค„์ด๋Š” ์•”ํ˜ธํ™”๋œ ๋ถˆํˆฌ๋ช…(opaque) ์•„์ดํ…œ์„ ๋ฐ˜ํ™˜ํ•ฉ๋‹ˆ๋‹ค. ์ด๋ฅผ ํ†ตํ•ด ๋ชจ๋ธ์ด ์ปจํ…์ŠคํŠธ ์ œํ•œ์— ๋„๋‹ฌํ•˜์ง€ ์•Š๊ณ  ํ™•์žฅ๋œ ์›Œํฌํ”Œ๋กœ์šฐ์—์„œ ์ถ”๋ก ์„ ๊ณ„์†ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.

์‚ฌ์šฉ ์‹œ๋‚˜๋ฆฌ์˜ค:

  • ๋‹ค์ˆ˜์˜ Tool Call์ด ํฌํ•จ๋œ ๋ฉ€ํ‹ฐ์Šคํ… ์—์ด์ „ํŠธ ํ”Œ๋กœ์šฐ
  • ์ด์ „ ํ„ด์„ ์œ ์ง€ํ•ด์•ผ ํ•˜๋Š” ๊ธด ๋Œ€ํ™”
  • ์ตœ๋Œ€ ์ปจํ…์ŠคํŠธ ์œˆ๋„์šฐ๋ฅผ ๋„˜์–ด์„œ๋Š” ๋ฐ˜๋ณต์  ์ถ”๋ก 

์ฃผ์š” ํŠน์„ฑ:

  • ๋ถˆํˆฌ๋ช…ํ•˜๊ณ  ์•”ํ˜ธํ™”๋œ ์•„์ดํ…œ์„ ์ƒ์„ฑํ•˜๋ฏ€๋กœ ๋‚ด๋ถ€ ๋กœ์ง์€ ์ง„ํ™”ํ•  ์ˆ˜ ์žˆ์Œ
  • ๊ฒ€์‚ฌ(inspection)๊ฐ€ ์•„๋‹Œ ๊ณ„์†(continuation)์„ ์œ„ํ•ด ์„ค๊ณ„๋จ
  • GPT-5.2 ๋ฐ Responses API์™€ ํ˜ธํ™˜
  • ๊ธด ์„ธ์…˜์—์„œ ๋ฐ˜๋ณต ์‹คํ–‰ํ•ด๋„ ์•ˆ์ „ํ•จ

๋ฒ ์ŠคํŠธ ํ”„๋ž™ํ‹ฐ์Šค:

  • ์ปจํ…์ŠคํŠธ ์‚ฌ์šฉ๋Ÿ‰์„ ๋ชจ๋‹ˆํ„ฐ๋งํ•˜๊ณ  ์ปจํ…์ŠคํŠธ ์œˆ๋„์šฐ ์ œํ•œ์— ๋„๋‹ฌํ•˜๊ธฐ ์ „์— ๋ฏธ๋ฆฌ ๊ณ„ํš
  • ๋งค ํ„ด๋งˆ๋‹ค๊ฐ€ ์•„๋‹Œ ์ฃผ์š” ๋งˆ์ผ์Šคํ†ค(์˜ˆ: Tool ์ง‘์•ฝ์  ๋‹จ๊ณ„) ํ›„์— Compact ์ˆ˜ํ–‰
  • ์žฌ๊ฐœ ์‹œ ํ–‰๋™ ์ดํƒˆ์„ ํ”ผํ•˜๊ธฐ ์œ„ํ•ด ํ”„๋กฌํ”„ํŠธ๋ฅผ ๊ธฐ๋Šฅ์ ์œผ๋กœ ๋™์ผํ•˜๊ฒŒ ์œ ์ง€
  • Compacted ์•„์ดํ…œ์€ ๋ถˆํˆฌ๋ช…ํ•˜๊ฒŒ ์ทจ๊ธ‰ํ•˜๊ณ  ๋‚ด๋ถ€์— ์˜์กดํ•˜๊ฑฐ๋‚˜ ํŒŒ์‹ฑํ•˜์ง€ ์•Š์Œ

์—์ด์ „ํŠธ ์กฐ์ข…์„ฑ๊ณผ ์‚ฌ์šฉ์ž ์—…๋ฐ์ดํŠธ

GPT-5.2๋Š” ์ ์ ˆํ•œ ํ”„๋กฌํ”„ํŒ…์ด ์ฃผ์–ด์กŒ์„ ๋•Œ ์—์ด์ „ํŠธ ์Šค์บํด๋”ฉ๊ณผ ๋ฉ€ํ‹ฐ์Šคํ… ์‹คํ–‰์— ๊ฐ•ํ•ฉ๋‹ˆ๋‹ค. ๊ธฐ์กด GPT-5.1์˜ <user_updates_spec>๊ณผ <solution_persistence> ๋ธ”๋ก์„ ์žฌ์‚ฌ์šฉํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.

GPT-5.2์˜ ์„ฑ๋Šฅ์„ ๋”์šฑ ํ–ฅ์ƒ์‹œํ‚ค๊ธฐ ์œ„ํ•œ ๋‘ ๊ฐ€์ง€ ํ•ต์‹ฌ ์กฐ์ •์ด ์žˆ์Šต๋‹ˆ๋‹ค. ์ฒซ์งธ, ์—…๋ฐ์ดํŠธ์˜ ์ƒ์„ธ๋„๋ฅผ ์ œํ•œํ•˜์—ฌ ๋” ์งง๊ณ  ์ง‘์ค‘์ ์œผ๋กœ ๋งŒ๋“ญ๋‹ˆ๋‹ค. ๋‘˜์งธ, ๋ฒ”์œ„ ๊ทœ์œจ์„ ๋ช…์‹œ์ ์œผ๋กœ ์ง€์ •ํ•˜์—ฌ ๋ฌธ์ œ ํ‘œ๋ฉด์ ์„ ํ™•์žฅํ•˜์ง€ ์•Š๋„๋ก ํ•ฉ๋‹ˆ๋‹ค.

1
2
3
4
5
6
7
8
9
10
<user_updates_spec>
- Send brief updates (1โ€“2 sentences) only when:
  - You start a new major phase of work, or
  - You discover something that changes the plan.
- Avoid narrating routine tool calls ("reading fileโ€ฆ", "running testsโ€ฆ").
- Each update must include at least one concrete outcome 
  ("Found X", "Confirmed Y", "Updated Z").
- Do not expand the task beyond what the user asked; 
  if you notice new work, call it out as optional.
</user_updates_spec>

Tool Calling๊ณผ ๋ณ‘๋ ฌํ™”

GPT-5.2๋Š” Tool ์‹ ๋ขฐ์„ฑ๊ณผ ์Šค์บํด๋”ฉ์—์„œ 5.1์„ ๊ฐœ์„ ํ–ˆ์œผ๋ฉฐ, ํŠนํžˆ MCP/Atlas ์Šคํƒ€์ผ ํ™˜๊ฒฝ์—์„œ ๊ทธ๋ ‡์Šต๋‹ˆ๋‹ค.

๋ฒ ์ŠคํŠธ ํ”„๋ž™ํ‹ฐ์Šค:

  • Tool์„ ๊ฐ„๊ฒฐํ•˜๊ฒŒ ์„ค๋ช…ํ•ฉ๋‹ˆ๋‹ค. ๋ฌด์—‡์„ ํ•˜๊ณ  ์–ธ์ œ ์‚ฌ์šฉํ•˜๋Š”์ง€์— ๋Œ€ํ•ด 1-2๋ฌธ์žฅ์œผ๋กœ ๊ธฐ์ˆ ํ•ฉ๋‹ˆ๋‹ค.
  • ์ฝ”๋“œ๋ฒ ์ด์Šค ์Šค์บ”, ๋ฒกํ„ฐ ์Šคํ† ์–ด, ๋˜๋Š” ๋‹ค์ค‘ ์—”ํ‹ฐํ‹ฐ ์ž‘์—…์— ๋Œ€ํ•ด ๋ณ‘๋ ฌํ™”๋ฅผ ๋ช…์‹œ์ ์œผ๋กœ ๊ถŒ์žฅํ•ฉ๋‹ˆ๋‹ค.
  • ๊ณ ์˜ํ–ฅ ์ž‘์—…(์ฃผ๋ฌธ, ์ฒญ๊ตฌ, ์ธํ”„๋ผ ๋ณ€๊ฒฝ)์— ๋Œ€ํ•ด์„œ๋Š” ๊ฒ€์ฆ ๋‹จ๊ณ„๋ฅผ ์š”๊ตฌํ•ฉ๋‹ˆ๋‹ค.
1
2
3
4
5
6
7
8
9
10
11
<tool_usage_rules>
- Prefer tools over internal knowledge whenever:
  - You need fresh or user-specific data (tickets, orders, configs, logs).
  - You reference specific IDs, URLs, or document titles.
- Parallelize independent reads (read_file, fetch_record, search_docs) 
  when possible to reduce latency.
- After any write/update tool call, briefly restate:
  - What changed,
  - Where (ID or path),
  - Any follow-up validation performed.
</tool_usage_rules>

๊ตฌ์กฐํ™”๋œ ์ถ”์ถœ, PDF, Office ์›Œํฌํ”Œ๋กœ์šฐ

์ด ์˜์—ญ์€ GPT-5.2๊ฐ€ ๋ช…ํ™•ํ•œ ๊ฐœ์„ ์„ ๋ณด์ด๋Š” ๋ถ„์•ผ์ž…๋‹ˆ๋‹ค. ์ตœ๋Œ€ํ•œ ํ™œ์šฉํ•˜๊ธฐ ์œ„ํ•œ ํ•ต์‹ฌ ์›์น™์€ ๋‹ค์Œ๊ณผ ๊ฐ™์Šต๋‹ˆ๋‹ค.

ํ•ญ์ƒ ์Šคํ‚ค๋งˆ ๋˜๋Š” JSON ํ˜•ํƒœ๋ฅผ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค. ์—„๊ฒฉํ•œ ์Šคํ‚ค๋งˆ ์ค€์ˆ˜๋ฅผ ์œ„ํ•ด Structured Outputs๋ฅผ ์‚ฌ์šฉํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ํ•„์ˆ˜ ํ•„๋“œ์™€ ์„ ํƒ ํ•„๋“œ๋ฅผ ๊ตฌ๋ถ„ํ•˜๊ณ , โ€œ์ถ”์ถœ ์™„์ „์„ฑ(extraction completeness)โ€์„ ์š”์ฒญํ•˜๋ฉฐ ๋ˆ„๋ฝ๋œ ํ•„๋“œ๋ฅผ ๋ช…์‹œ์ ์œผ๋กœ ์ฒ˜๋ฆฌํ•ฉ๋‹ˆ๋‹ค.

1
2
3
4
5
6
7
8
9
10
11
12
13
<extraction_spec>
You will extract structured data from tables/PDFs/emails into JSON.

- Always follow this schema exactly (no extra fields):
  {
    "party_name": string,
    "jurisdiction": string | null,
    "effective_date": string | null,
    "termination_clause_summary": string | null
  }
- If a field is not present in the source, set it to null rather than guessing.
- Before returning, quickly re-scan the source for any missed fields and correct omissions.
</extraction_spec>

๋‹ค์ค‘ ํ…Œ์ด๋ธ”/๋‹ค์ค‘ ํŒŒ์ผ ์ถ”์ถœ์˜ ๊ฒฝ์šฐ, ๋ฌธ์„œ๋ณ„ ๊ฒฐ๊ณผ๋ฅผ ๋ณ„๋„๋กœ ์ง๋ ฌํ™”ํ•˜๊ณ  ์•ˆ์ •์ ์ธ ID(ํŒŒ์ผ๋ช…, ๊ณ„์•ฝ ์ œ๋ชฉ, ํŽ˜์ด์ง€ ๋ฒ”์œ„)๋ฅผ ํฌํ•จํ•˜๋„๋ก ๊ฐ€์ด๋“œ๋ฅผ ์ถ”๊ฐ€ํ•ฉ๋‹ˆ๋‹ค.


GPT-5.2๋กœ์˜ ํ”„๋กฌํ”„ํŠธ ๋งˆ์ด๊ทธ๋ ˆ์ด์…˜ ๊ฐ€์ด๋“œ

GPT-5 ๊ณ„์—ด ๋ชจ๋ธ์€ reasoning_effort ํŒŒ๋ผ๋ฏธํ„ฐ(none, minimal, low, medium, high, xhigh)๋ฅผ ์ง€์›ํ•˜๋ฉฐ, ์ด๋ฅผ ํ†ตํ•ด ์†๋„/๋น„์šฉ๊ณผ ๊นŠ์€ ์ถ”๋ก  ์‚ฌ์ด์˜ Trade-off๋ฅผ ์กฐ์ ˆํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.

๋งˆ์ด๊ทธ๋ ˆ์ด์…˜ ๋งคํ•‘:

ํ˜„์žฌ ๋ชจ๋ธ ๋Œ€์ƒ ๋ชจ๋ธ ๋Œ€์ƒ reasoning_effort ๋น„๊ณ 
GPT-4o GPT-5.2 none ๋น ๋ฅธ/๋‚ฎ์€ ์ˆ™๊ณ  ํ–‰๋™ ์œ ์ง€; ํ‰๊ฐ€ ๊ฒฐ๊ณผ๊ฐ€ ํ‡ด๋ณดํ•  ๋•Œ๋งŒ effort ์ฆ๊ฐ€
GPT-4.1 GPT-5.2 none GPT-4o์™€ ๋™์ผ
GPT-5 GPT-5.2 ๋™์ผ ๊ฐ’ (minimal โ†’ none ์ œ์™ธ) ์ง€์—ฐ/ํ’ˆ์งˆ ํ”„๋กœํŒŒ์ผ ์ผ๊ด€์„ฑ ์œ ์ง€
GPT-5.1 GPT-5.2 ๋™์ผ ๊ฐ’ ํ‰๊ฐ€ ์‹คํ–‰ ํ›„์—๋งŒ ์กฐ์ •

GPT-5์˜ ๊ธฐ๋ณธ reasoning level์€ medium์ด๊ณ , GPT-5.1๊ณผ GPT-5.2์˜ ๊ธฐ๋ณธ๊ฐ’์€ none์ž…๋‹ˆ๋‹ค.

๋งˆ์ด๊ทธ๋ ˆ์ด์…˜ ๋‹จ๊ณ„:

  1. ๋ชจ๋ธ๋งŒ ์ „ํ™˜, ํ”„๋กฌํ”„ํŠธ๋Š” ์œ ์ง€: ํ”„๋กฌํ”„ํŠธ ํŽธ์ง‘์ด ์•„๋‹Œ ๋ชจ๋ธ ๋ณ€๊ฒฝ๋งŒ ํ…Œ์ŠคํŠธํ•˜๋„๋ก ํ”„๋กฌํ”„ํŠธ๋ฅผ ๊ธฐ๋Šฅ์ ์œผ๋กœ ๋™์ผํ•˜๊ฒŒ ์œ ์ง€ํ•ฉ๋‹ˆ๋‹ค. ํ•œ ๋ฒˆ์— ํ•˜๋‚˜์˜ ๋ณ€๊ฒฝ๋งŒ ์ ์šฉํ•ฉ๋‹ˆ๋‹ค.
  2. reasoning_effort ๊ณ ์ •: GPT-5.2์˜ reasoning_effort๋ฅผ ์ด์ „ ๋ชจ๋ธ์˜ ์ง€์—ฐ/๊นŠ์ด ํ”„๋กœํŒŒ์ผ์— ๋งž๊ฒŒ ๋ช…์‹œ์ ์œผ๋กœ ์„ค์ •ํ•ฉ๋‹ˆ๋‹ค.
  3. ๊ธฐ์ค€์„  ํ‰๊ฐ€ ์‹คํ–‰: ๋ชจ๋ธ๊ณผ effort๊ฐ€ ์ •๋ ฌ๋œ ํ›„ ํ‰๊ฐ€ ์Šค์œ„ํŠธ๋ฅผ ์‹คํ–‰ํ•ฉ๋‹ˆ๋‹ค. ๊ฒฐ๊ณผ๊ฐ€ ์ข‹์œผ๋ฉด ๋ฐฐํฌ ์ค€๋น„ ์™„๋ฃŒ์ž…๋‹ˆ๋‹ค.
  4. ํ‡ด๋ณด ์‹œ ํ”„๋กฌํ”„ํŠธ ํŠœ๋‹: Prompt Optimizer์™€ ๋Œ€์ƒ ์ œ์•ฝ(์ƒ์„ธ๋„/ํฌ๋งท/์Šคํ‚ค๋งˆ, ๋ฒ”์œ„ ๊ทœ์œจ)์„ ์‚ฌ์šฉํ•˜์—ฌ ๋™๋“ฑ์„ฑ์„ ๋ณต์›ํ•˜๊ฑฐ๋‚˜ ๊ฐœ์„ ํ•ฉ๋‹ˆ๋‹ค.
  5. ์ž‘์€ ๋ณ€๊ฒฝ๋งˆ๋‹ค ์žฌํ‰๊ฐ€: reasoning_effort๋ฅผ ํ•œ ๋‹จ๊ณ„ ์˜ฌ๋ฆฌ๊ฑฐ๋‚˜ ์ ์ง„์  ํ”„๋กฌํ”„ํŠธ ์กฐ์ • ํ›„ ๋‹ค์‹œ ์ธก์ •ํ•˜๋ฉฐ ๋ฐ˜๋ณตํ•ฉ๋‹ˆ๋‹ค.

์›น ๊ฒ€์ƒ‰๊ณผ ๋ฆฌ์„œ์น˜

GPT-5.2๋Š” ๋‹ค์ˆ˜์˜ ์†Œ์Šค์—์„œ ์ •๋ณด๋ฅผ ์ข…ํ•ฉํ•˜๋Š” ๋Šฅ๋ ฅ์ด ํ–ฅ์ƒ๋˜์–ด ๋” ์กฐ์ข… ๊ฐ€๋Šฅ(steerable)ํ•ฉ๋‹ˆ๋‹ค.

๋ฒ ์ŠคํŠธ ํ”„๋ž™ํ‹ฐ์Šค:

๋ฆฌ์„œ์น˜ ๊ธฐ์ค€์„ ์‚ฌ์ „์— ๋ช…์‹œํ•ฉ๋‹ˆ๋‹ค. ๋ชจ๋ธ์—๊ฒŒ ๊ฒ€์ƒ‰ ์ˆ˜ํ–‰ ๋ฐฉ์‹์„ ์•Œ๋ ค์ค๋‹ˆ๋‹ค. 2์ฐจ ๋ฆฌ๋“œ๋ฅผ ๋”ฐ๋ฅผ์ง€, ๋ชจ์ˆœ์„ ํ•ด๊ฒฐํ• ์ง€, ์ธ์šฉ์„ ํฌํ•จํ• ์ง€ ๋ช…์‹œํ•ฉ๋‹ˆ๋‹ค. ํ•œ๊ณ„ ๊ฐ€์น˜๊ฐ€ ๋–จ์–ด์งˆ ๋•Œ๊นŒ์ง€ ์ถ”๊ฐ€ ๋ฆฌ์„œ์น˜๋ฅผ ๊ณ„์†ํ•ด์•ผ ํ•œ๋‹ค๋Š” ์‹์œผ๋กœ ์–ผ๋งˆ๋‚˜ ๋ฉ€๋ฆฌ ๊ฐˆ์ง€ ๋ช…์‹œ์ ์œผ๋กœ ์ง„์ˆ ํ•ฉ๋‹ˆ๋‹ค.

์งˆ๋ฌธ์ด ์•„๋‹Œ ์ง€์‹œ๋กœ ๋ชจํ˜ธ์„ฑ์„ ์ œ์•ฝํ•ฉ๋‹ˆ๋‹ค. ๋ชจ๋ธ์—๊ฒŒ ๋ชจ๋“  ๊ทธ๋Ÿด๋“ฏํ•œ ์˜๋„๋ฅผ ํฌ๊ด„์ ์œผ๋กœ ๋‹ค๋ฃจ๊ณ  ๋ช…ํ™•ํ™” ์งˆ๋ฌธ์„ ํ•˜์ง€ ๋ง๋ผ๊ณ  ์ง€์‹œํ•ฉ๋‹ˆ๋‹ค. ๋ถˆํ™•์‹ค์„ฑ์ด ์กด์žฌํ•  ๋•Œ ํญ๊ณผ ๊นŠ์ด๋ฅผ ์š”๊ตฌํ•ฉ๋‹ˆ๋‹ค.

์ถœ๋ ฅ ํ˜•ํƒœ์™€ ํ†ค์„ ์ง€์ •ํ•ฉ๋‹ˆ๋‹ค. ๊ตฌ์กฐ(Markdown, ํ—ค๋”, ๋น„๊ต๋ฅผ ์œ„ํ•œ ํ…Œ์ด๋ธ”), ๋ช…ํ™•์„ฑ(์•ฝ์–ด ์ •์˜, ๊ตฌ์ฒด์  ์˜ˆ์‹œ), ์–ด์กฐ(๋Œ€ํ™”์ฒด, ํŽ˜๋ฅด์†Œ๋‚˜ ์ ์‘ํ˜•, ๋น„์•„์ฒจํ˜•)์— ๋Œ€ํ•œ ๊ธฐ๋Œ€๋ฅผ ์„ค์ •ํ•ฉ๋‹ˆ๋‹ค.

1
2
3
4
5
6
7
8
9
10
11
<web_search_rules>
- Act as an expert research assistant; default to comprehensive, well-structured answers.
- Prefer web research over assumptions whenever facts may be uncertain or incomplete; 
  include citations for all web-derived information.
- Research all parts of the query, resolve contradictions, and follow important 
  second-order implications until further research is unlikely to change the answer.
- Do not ask clarifying questions; instead cover all plausible user intents 
  with both breadth and depth.
- Write clearly and directly using Markdown; define acronyms, use concrete examples, 
  and keep a natural, conversational tone.
</web_search_rules>

๊ฒฐ๋ก 

GPT-5.2๋Š” ์ •ํ™•์„ฑ, ์‹ ๋ขฐ์„ฑ, ๊ทœ์œจ ์žˆ๋Š” ์‹คํ–‰์„ ์šฐ์„ ์‹œํ•˜๋Š” ํ”„๋กœ๋•์…˜๊ธ‰ ์—์ด์ „ํŠธ๋ฅผ ๊ตฌ์ถ•ํ•˜๋Š” ํŒ€์—๊ฒŒ ์˜๋ฏธ ์žˆ๋Š” ์ง„์ „์„ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค. ๋” ๊ฐ•๋ ฅํ•œ ์ง€์‹œ ์ค€์ˆ˜, ๋” ๊น”๋”ํ•œ ์ถœ๋ ฅ, ๋ณต์žกํ•œ Tool ์ง‘์•ฝ์  ์›Œํฌํ”Œ๋กœ์šฐ์—์„œ ๋” ์ผ๊ด€๋œ ํ–‰๋™์„ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค.

๋Œ€๋ถ€๋ถ„์˜ ๊ธฐ์กด ํ”„๋กฌํ”„ํŠธ๋Š” ์ดˆ๊ธฐ ์ „ํ™˜ ์‹œ reasoning effort, ์ƒ์„ธ๋„, ๋ฒ”์œ„ ์ œ์•ฝ์ด ๋ณด์กด๋  ๋•Œ ๊น”๋”ํ•˜๊ฒŒ ๋งˆ์ด๊ทธ๋ ˆ์ด์…˜๋ฉ๋‹ˆ๋‹ค. ํŒ€์€ ํ”„๋กฌํ”„ํŠธ ๋ณ€๊ฒฝ ์ „์— ํ‰๊ฐ€๋ฅผ ํ†ตํ•ด ํ–‰๋™์„ ๊ฒ€์ฆํ•˜๊ณ , ํ‡ด๋ณด๊ฐ€ ๋‚˜ํƒ€๋‚  ๋•Œ๋งŒ reasoning effort๋‚˜ ์ œ์•ฝ์„ ์กฐ์ •ํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. ๋ช…์‹œ์  ํ”„๋กฌํ”„ํŒ…๊ณผ ์ธก์ •๋œ ๋ฐ˜๋ณต์„ ํ†ตํ•ด GPT-5.2๋Š” ์˜ˆ์ธก ๊ฐ€๋Šฅํ•œ ๋น„์šฉ๊ณผ ์ง€์—ฐ ํ”„๋กœํŒŒ์ผ์„ ์œ ์ง€ํ•˜๋ฉด์„œ ๋” ๋†’์€ ํ’ˆ์งˆ์˜ ๊ฒฐ๊ณผ๋ฅผ ๋‹ฌ์„ฑํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.

์ฝ์–ด์ฃผ์…”์„œ ๊ฐ์‚ฌํ•ฉ๋‹ˆ๋‹ค ๐Ÿพ



-->