| name | writing-gpt-prompts |
| description | Use when authoring or refining a prompt for GPT-5.5 / OpenAI models (developer message, user message, tool description, agent instructions) to make it follow OpenAI's official prompt-guidance â covers outcome-first design, the Role / Personality / Goal / Success / Constraints / Output / Stop scaffold, markdown vs XML mixing, few-shot examples, agentic preambles, reasoning-effort and verbosity knobs, tool-persistence rules, and the phase parameter for the Responses API. |
Writing GPT-5.5 Prompts
Overview
OpenAI å
¬åŒã® Prompt GuidanceïŒGPT-5.5 / Responses API åãïŒ+ Prompt Engineering Guide + Using GPT-5.5 ãçµ±åãã referenceãGPT-5 ç³»ïŒç¹ã« 5.5ïŒã¯ outcome-firstïŒdestination ã瀺ã path ã¯ã¢ãã«ã«ä»»ããïŒèšèšãåæã§ãClaude 4.x ã®ãå
·äœçã»åŸ¹åºçã«æžãããšã¯èšèšææ³ãç°ãªãã
Golden RuleïŒé»éåŸïŒ:
destination ãå®çŸ©ããå¹çç㪠path ã¯ã¢ãã«ã«éžã°ããã ããã»ã¹ã step-by-step ã§èŠå®ããã®ã§ã¯ãªããéæãã¹ã outcome / success criteria / constraints / stop rules ãæžãã
æ žãšãªãåå:
- Outcome-first: ãA â B â C ããããã§ã¯ãªãããããæºãããããæåããšæžã
- Markdown ãš XML ã®äœ¿ãåã: hierarchy / èŠåºã㯠markdownãã³ã³ãã³ãå¢çïŒreference docs / äŸïŒã¯ XML
- çºè©±é åº: Identity â Instructions â Examples â ContextïŒåå©çšéšã¯åæ¹ã«çœ®ã㊠prompt cache ãå¹ãããïŒ
- 絶察ã«ãŒã«ãæ¿«çšããªã:
ALWAYS / NEVER / must / only ã¯æ¬åœã® invariant ã«ã ã䜿ã
- stop rules ãå¿
ãæžã: ãretry / fallback / abstain / ask / stop ããã€éžã¶ãããæç€º
- GPT-5.5 ã®ããã©ã«ã㯠concise & direct: æž©ããããŒã³ãå¿
èŠãªãšãã ãæç€ºãã
å
¬åŒãªãã¡ã¬ã³ã¹:
When to Use
以äžã®å Žé¢ã§çºç«ãã:
- ãŠãŒã¶ãŒã GPT-5.5 / GPT-5 / OpenAI ã¢ãã«åãã® prompt ããæžããŠïŒäœã£ãŠïŒæ¹åããŠïŒã¬ãã¥ãŒããŠããšèšã£ããšã
- developer message / system çžåœ / user message / tool description / agent instructions ãæ°èŠäœæã»ä¿®æ£ãããšã
- OpenAI Responses API / Chat Completions API ã䜿ãã¢ããªã® prompt ãèšèšãããšã
- GPT-4.x â GPT-5 / GPT-5 â GPT-5.5 ãžã®ç§»è¡ã§ prompt ãæŽæ°ãããšã
- ãŠãŒã¶ãŒããOpenAI ã®ãã¹ããã©ã¯ãã£ã¹ã«åŸã£ãŠâŠããprompt-guidance ã«æ²¿ã£ãŠâŠããšæç€ºãããšã
- Codex CLI / Codex æ¡åŒµã§å
éšã® prompt ãçµã¿ç«ãŠããšãïŒãã ã
codex:gpt-5-4-prompting skill ãããå Žåã¯ãã¡ããåªå
ïŒ
䜿ããªãå Žé¢:
- Claude / Anthropic çšã® prompt â
writing-claude-prompts ã䜿ã
- Gemini ãªã©ä»ç€Ÿ LLM çš â ãã®ã¢ãã«åºæã® reference ã䜿ã
- åã«æç« ãæžãã¿ã¹ã¯ïŒprompt èšèšãå«ãŸãªãïŒ
Core Patterns
1. 7 èŠçŽ ã¹ãã£ãã©ãŒã«ã
OpenAI å
¬åŒã® Suggested Frameworkãé åºéãã«æžããš outcome-first ã® prompt ãèªç¶ã«çµã¿äžããã
| èŠçŽ | å
容 | äŸ |
|---|
| Role | 1ã2 æã§æ©èœãš context ãå®çŸ© | "You are a customer-support agent for a SaaS billing system." |
| Personality | ããŒã³ / æ¯ãèã / ååã¹ã¿ã€ã« | "Patient, respectful, practical. Assume the user is competent and acting in good faith." |
| Goal | ãŠãŒã¶ãŒããèŠããã¢ãŠãã«ã | "Resolve the customer's issue end to end." |
| Success criteria | æçµåçåã« true ã§ãªããã°ãªããªãæ¡ä»¶ | "Eligibility decision is made from policy and account data; allowed action is completed before responding." |
| Constraints | policy / safety / evidence / å¯äœçšã®å¶é | "No refunds over $100 without manager approval. Cite policy IDs." |
| Output | ã»ã¯ã·ã§ã³ / é·ã / ããŒã³ | "Single message under 200 words, plain paragraphs." |
| Stop rules | retry / fallback / abstain / ask / stop ã®å€æ | "If account data is unavailable, ask the user for the order ID instead of guessing." |
2. Outcome-first vs. Process-heavy
# â NG: process-heavyïŒGPT-5.5 ã窮å±ã«ãªãïŒ
First inspect A, then inspect B, then compare every field,
then think through all possible exceptions, then decide which tool to call.
# â
OK: outcome-first
Resolve the customer's issue end to end.
Success means: the eligibility decision is made from available policy and
account data; any allowed action is completed before responding.
step-by-step ãå¿
èŠãªã®ã¯ãpath èªäœããããã¯ãã®äžéšãïŒèŠå¶å¯Ÿå¿ / ç£æ»ãã° / æè²ã¢ããªïŒã«éãã
3. ã¡ãã»ãŒãžããŒã«ã®éå±€
| ããŒã« | çšé | åªå
床 |
|---|
developer | ã¢ããªéçºè
ã®æç€ºïŒæ§ systemïŒ | æé« |
user | ãšã³ããŠãŒã¶ãŒå
¥å | äž |
assistant | ã¢ãã«å¿ç | â |
instructions API param | æ¯ãèãã®äžäœã¬ã€ãã³ã¹ïŒinput prompt ããåªå
ïŒ | developer çžåœ |
developer ãã颿°å®çŸ©ããuser ãã颿°åŒæ°ããšæãããGPT-5.5 ã§ã¯ developer ãåæ¹ãuser ãåŸæ¹ã«çœ®ãã®ãåºæ¬ã
4. Markdown ãš XML ã®äœ¿ãåã
| ç®ç | 䜿ããã® |
|---|
| éå±€ / ã»ã¯ã·ã§ã³åã / äžèЧ | MarkdownïŒ## èŠåºãã- ãªã¹ããè¡šïŒ |
| ã³ã³ãã³ãå¢çïŒåç
§ããã¥ã¡ã³ã / äŸ / å
¥åããŒã¿ïŒ | XML ã¿ã°ïŒ<reference_docs>, <example>, <input>ïŒ |
| è€éãªæç€ºããã㯠| XML ã¿ã°ïŒ<output_contract>, <verbosity_controls> çïŒ |
| ã€ã³ã©ã€ã³ã³ãŒã / ãã¡ã€ã«ãã¹ / ã³ãã³ã | ããã¯ã¯ã©ãŒã |
XML ã¿ã°ã¯ äžè²«ããåœå ã䜿ãã屿§ã§ metadata ã衚çŸã§ããïŒäŸ: <document index="1" source="policy.md">ïŒã
5. æšå¥šã»ã¯ã·ã§ã³é
# 1. Identity â Role, purpose, style, high-level goals
# 2. Instructions â Rules, do's/don'ts, constraints, stop rules
# 3. Examples â Input/Output pairs (few-shot)
# 4. Context â Reference docs, proprietary data, retrieved chunks
# 5. User input â æåŸïŒããã«å®éã®ã¯ãšãªãæ¥ãïŒ
åå©çšãããéšåïŒIdentity / Instructions / ExamplesïŒãåæ¹ã«çœ®ããšãResponses API ã® prompt caching ãå¹ããŠã³ã¹ããšã¬ã€ãã³ã·ãäžããã
6. Few-Shot Examples
3ã5 åã®å€æ§ãª input/output ã developer ã¡ãã»ãŒãžå
ã«çœ®ããXML ã¿ã°ã§æç€ºçã«åºåãã
<examples>
<example>
<product_review>This blender is loud but powerful.</product_review>
<assistant_response>{"sentiment":"mixed","aspects":["noise","power"]}</assistant_response>
</example>
<example>
<product_review>Stopped working after a week.</product_review>
<assistant_response>{"sentiment":"negative","aspects":["reliability"]}</assistant_response>
</example>
</examples>
ãšããžã±ãŒã¹ïŒææ§ / çãã / å€èšèªïŒãå«ããã
7. Reasoning Effort
| effort | çšé | åè |
|---|
none | ã¬ã€ãã³ã·æåªå
/ æšè«äžèŠ | ããã©ã«ãåçŠæ¢ |
low | ã¬ã€ãã³ã·æåºŠé« + è€éãªæç€ºãã | çãæç€ºã»åé¡ã¿ã¹ã¯ |
medium | GPT-5.5 ã®ããã©ã«ã | å質 / ã¬ã€ãã³ã· / ã³ã¹ãã®ãã©ã³ã¹ |
high | eval ã§æž¬ããåªäœãããæ | 倿®µæšè« / é£å解決 |
xhigh | é·æé agentic ã¿ã¹ã¯ã§ç¥æ§ > é床 | ããã©ã«ãåçŠæ¢ãäºåã« <completeness_contract> <verification_loop> <tool_persistence_rules> ãå
¥ããããš |
response = client.responses.create(
model="gpt-5.5",
reasoning={"effort": "medium"},
input=[{"role": "developer", "content": developer_msg}, ...],
)
8. Verbosity Control
GPT-5.5 ã®ããã©ã«ãæäœã¯ concise & directã
text={"verbosity": "low"}
text={"verbosity": "medium"}
text={"verbosity": "high"}
ã«ã¹ã¿ããŒåãã§æž©ãããæ¬²ããå Žå㯠developer åŽã§ personality ãæç€ºïŒGPT-5.5 ã®ããã©ã«ãã¯äºåçã«ãªããã¡ïŒã
9. Tool PreamblesïŒã¹ããªãŒãã³ã°å¿çæ§ïŒ
è€æ°ã¹ãããã® tool åŒã³åºãã§ãæåã® tool call åã«çã user-visible ãªã¹ããŒã¿ã¹ãåºããããšã§ãã¹ããªãŒãã³ã° UI ã®äœæé床ãäžããã
Before your first tool call, send a short user-visible update that acknowledges
the request and states the first step you will take. After each notable tool
call, briefly state what you learned and what you will do next.
Do NOT add a preamble for trivial tool calls.
10. Phase ParameterïŒResponses APIïŒ
previous_response_id ã§ã¯ãªãåºåã¢ã€ãã ãæ¯ã¿ãŒã³æåã§æž¡ãçŽãéçšã§ã¯ãassistant åºåã«ä¹ã£ãŠãã phase ã ãã®ãŸãŸä¿æããŠéãè¿ãã
| phase | æå³ |
|---|
commentary | äžéã¢ããããŒãïŒãã㟠X ã調ã¹ãŠããŸããïŒ |
final_answer | æçµåç |
- user ã¡ãã»ãŒãžã«ã¯ phase ãä»ããªã
- reasoning effort / preamble / ç¹°ãè¿ã tool åŒã³åºããããå Žåã¯ç¹ã«éèŠ
11. Tool Use ã®èŠåŸ
<tool_persistence_rules>
- Decompose multi-step requests into sub-tasks before calling tools.
- After each tool call, reflect on the result before continuing.
- Confirm each sub-task is complete before yielding control to the user.
- When uncertain, prefer one targeted tool call over several speculative ones.
- Do not invent tool arguments. If a required parameter is missing, ask the user.
</tool_persistence_rules>
䞊ååŒã³åºã㯠äŸåããªããšã ã ãæç€ºãããGPT-5.5 ã¯ããã©ã«ãã§æ§ãããªã®ã§ã䞊åãåŒãåºãããå Žåã®ã¿ãããã¯ãå
¥ããã
12. Citation / Grounding
ããã¥ã¡ã³ãåç
§ããŒã¹ã§çããããå Žå㯠retrieval budget ãæç€ºããã
<citation_rules>
For ordinary Q&A, start with one broad search using short, discriminative
keywords. If the top results contain enough citable support for the core
request, answer from those results without searching again.
Cite using inline markers like [doc:NAME §SECTION]. Do not cite passages you
did not actually retrieve.
If the available evidence is insufficient, say so explicitly and ask the user
for the missing piece. Do not infer "no" from absence of evidence.
</citation_rules>
13. Coding Tasks
<output_contract>
- Implement the requested change as a minimal patch.
- Add or update unit tests covering the new behavior.
- Run the most relevant validation available: targeted unit tests, type checks,
lint checks, or minimal smoke tests.
- Report the validation command and its result.
- If validation fails, fix or roll back; do not paper over with broad excepts
or silent defaults.
</output_contract>
<default_follow_through_policy>
Prefer making progress over stopping for clarification when the request is
already clear enough to attempt. Use context and reasonable assumptions to
move forward, and surface assumptions in the final report.
</default_follow_through_policy>
14. Frontend Engineering
GPT-5.5 æšå¥šã¹ã¿ãã¯: Tailwind CSS / shadcn/ui / Radix Themes / Lucide icons / Motionã
zero-to-one ã¢ããªçææã¯ãAI çæã«ããããããã©ã«ãïŒçŽ«ã°ã©ããInterãcenter-aligned heroïŒãæç€ºçã«é¿ããããã
Avoid the following common generated-UI defaults:
- Purple gradients on white or dark backgrounds
- Inter / Roboto / system-ui as the only typeface
- Center-aligned hero with a single CTA and a placeholder illustration
- Card grids with no information density
Use a typographic system with at least one display face and one body face,
a constrained palette tied to the product, and component states (hover,
focus, disabled, loading, empty, error) explicitly defined.
15. Customer-Facing Workflows
Persona: friendly senior support agent.
Channel: in-app chat.
Register: warm, plain-spoken, no jargon.
Format: prose only â no headers, no bullet lists, no bold.
Length: max 120 words.
ããã©ãŒããããçŠæ¢ããããããŒã㪠length äžéãã æç€ºãGPT-5.5 ã¯ããã©ã«ãã§ markdown ãåºãåŸåãããã®ã§ãprose éå®ãªãæç€ºçã« banned formatting ãæžãã
16. Research / Synthesis
For research tasks, use the minimum evidence sufficient to answer correctly,
cite it precisely, then stop. If evidence is insufficient, say so and ask;
do not pad the answer with speculation.
17. Creative Drafting Guardrails
Use retrieved or provided facts for concrete product, customer, metric,
roadmap, date, capability, and competitive claims, and cite those claims.
Do not invent specific names, first-party data claims, metrics, or customer
outcomes to make the draft sound stronger.
18. Instruction Priority ãš Task Update
äŒè©±éäžã§ã¿ã¹ã¯ãåãæ¿ããã¢ããªã§ã¯ãdeveloper åŽã®åªå
é äœãæç€ºããã
<instruction_priority>
1. Safety and policy constraints in this developer message.
2. The most recent <task_update> block in this conversation.
3. Earlier developer instructions.
4. User instructions.
When in conflict, follow the higher-priority source and tell the user briefly.
</instruction_priority>
ã¿ã¹ã¯å€æŽæã¯æ°ãã <task_update> ãããã¯ã developer åŽã§éãã
Quick ReferenceïŒãã§ãã¯ãªã¹ãïŒ
GPT-5.5 çš prompt ãæžãçµãã£ãã確èª:
WorkflowïŒpromptãäœããšãã®æé ïŒ
- ç®ç確èª: äœãéæããã prompt ã / ã©ã®ã¢ãã« / ã©ã® effort / ã©ã® verbosity / å
¥åºåãã©ãŒãããã¯äœã ããŠãŒã¶ãŒã«èãåã
- outcome èšèš: ããããæã£ããæåãã 1 è¡ã§æžãããŸã§è©°ãã
- 7 èŠçŽ ã®éªšæ Œ: Role / Personality / Goal / Success / Constraints / Output / Stop ã®èŠåºãã§èçš¿
- æ§é å: Identity â Instructions â Examples â Context ã®é ã«äžŠã¹ã
- Examples: 3ã5 åã®
<example> ãçµã¿èŸŒãïŒå¿
èŠãªãïŒ
- Reference docs: 倧ããªåç
§ã¯
<reference_docs> ã§åæ¹é
眮ïŒcache å¹çåïŒ
- API knobs: reasoning effort / verbosity / phase / preamble ã®å¿
èŠæ§ãå€å®ããŠä»äž
- Anti-pattern ãã§ãã¯: 絶察ã«ãŒã«æ¿«çš / process-heavy / stop rule æ¬ åŠ ãåãé€ã
- ãã§ãã¯ãªã¹ã ã§èªå·±ã¬ãã¥ãŒ
- ã¢ãã«å¥åŸ®èª¿æŽ: GPT-5.5 ãªã concise ããã©ã«ãã«åããããGPT-5 / 5.1 äºæãèŠããªã effort ã®ããã©ã«ãå·®ãèæ
®
Common Mistakes
| ããããééã | ä¿®æ£ |
|---|
Step 1: ... Step 2: ... Step 3: ... ã§å
šå·¥çšãèŠå®ãã | outcome / success criteria / stop rules ã«çœ®ãæããpath ã¯ã¢ãã«ã«ä»»ãã |
ALWAYS, NEVER, MUST ãå€çš | çã® invariant ã«ã ã䜿ãããã以å€ã¯éåžžã® declarative ã«æžã |
| stop rules ãæžããªã | ãretry / fallback / abstain / ask / stopãã®å€æãå¿
ãæç€º |
| markdown ãå
šé¢çã«çŠæ¢ããåçºæç€º | banned formatting ãç®æ¡æžãã§åæïŒ length äžéãäœµèš |
system: ããŒã«åã§æžã | GPT-5.5 / Responses API ã§ã¯ developer ããŒã«ã«å€æŽ |
reasoning effort ãæžãã xhigh ãæåžžäœ¿çš | ããã©ã«ã mediumãxhigh 㯠agentic é·æéã¿ã¹ã¯éå® |
previous_response_id äžäœ¿çšã〠phase ãè¿ããªã | æå state 管çæã¯ assistant ã® phase ãä¿æããŠåé |
| evidence äžè¶³æã«ãnoããšçããŠããŸã | citation_rules ã§ãmissing â noããäžè¶³ããã askããšæç€º |
| 䞊åããŒã«åŒã³åºããèµ·ããªã | äŸåãªãã®å Žåã®ã¿äžŠåãèš±å¯ãããããã¯ã developer ã«è¿œå |
| åµäœã§åºæåè©ã»ææšãæé ãã | source-backed ãš creative wording ãåé¢ããã¬ãŒããæžã |
| customer-facing ã§çŽ ã£æ°ãªã | personality ãæç€ºãGPT-5.5 ããã©ã«ãã¯äºåç |
<example> ã©ãããªãã§äŸã貌ã | <example> / <examples> ã§æ§é åãã |
| é·æããã¥ã¡ã³ããæ«å°Ÿã«çœ®ã | åæ¹é
眮ïŒIdentity / Instructions ã®çŽåŸïŒãprompt cache ãå¹ã |
Reference Snippets
ãã®ãŸãŸåã蟌ããå®çšãããã¯éãã³ããããŠäœ¿ãã
A. Output ContractïŒæ±çšïŒ
<output_contract>
- Format: <plain prose | markdown sections | JSON matching schema>.
- Length: <hard limit, e.g. †200 words>.
- Sections: <ordered list of required sections, or "single message">.
- Tone: <e.g. patient, plain-spoken, no jargon>.
- Banned: <e.g. headers, bullet lists, emojis> (when prose-only is required).
</output_contract>
B. Verbosity ControlsïŒdeveloper åŽã§åºå®ãããå ŽåïŒ
<verbosity_controls>
Default to concise, direct prose. Do not restate the user's question.
Do not add closing pleasantries. Use bullets or tables only when the content
is genuinely list-like or tabular. Each paragraph should end when the point
is made â no padding sentences.
</verbosity_controls>
C. Default Follow-Through Policy
<default_follow_through_policy>
Prefer making progress over stopping for clarification when the request is
already clear enough to attempt. Use context and reasonable assumptions to
move forward, and surface those assumptions in the final report. Ask for
clarification only when the missing information would change the outcome.
</default_follow_through_policy>
D. Tool Persistence Rules
<tool_persistence_rules>
- Decompose multi-step requests into sub-tasks before calling tools.
- After each tool call, reflect on the result before continuing.
- Confirm each sub-task is complete before yielding control to the user.
- When uncertain, prefer one targeted tool call over several speculative ones.
- Do not invent tool arguments. If a required parameter is missing, ask.
- Avoid speculative changes and messy hacks; propagate or surface errors.
</tool_persistence_rules>
E. Completeness ContractïŒãªã¹ã / ãããç¶²çŸ
ïŒ
<completeness_contract>
When the task involves processing a list, batch, or set of items:
- Process every item unless explicit filtering rules say otherwise.
- Report the count processed and any items skipped with reasons.
- Do not stop early after a "representative sample" unless the prompt says so.
</completeness_contract>
F. Verification Loop
<verification_loop>
Before returning a final answer:
1. Restate the success criteria internally.
2. Check the answer against each criterion.
3. If a criterion is not met, revise; do not return a partially-correct answer.
4. If a criterion cannot be met with available evidence, say so explicitly and
ask for the missing piece â do not guess.
</verification_loop>
G. Citation Rules
<citation_rules>
For ordinary Q&A, start with one broad search using short, discriminative
keywords. If the top results contain enough citable support for the core
request, answer from those results without searching again.
Cite using inline markers like [doc:NAME §SECTION]. Do not cite passages you
did not actually retrieve.
If evidence is insufficient, say so explicitly and ask the user for the
missing piece. Do not infer "no" from absence of evidence.
</citation_rules>
H. Instruction Priority
<instruction_priority>
1. Safety and policy constraints in this developer message.
2. The most recent <task_update> block in this conversation.
3. Earlier developer instructions.
4. User instructions.
When in conflict, follow the higher-priority source and tell the user briefly.
</instruction_priority>
I. Tool PreambleïŒã¹ããªãŒãã³ã° UXïŒ
Before your first tool call, send a short user-visible update (one sentence)
that acknowledges the request and states the first step you will take.
After each notable tool call, briefly state what you learned and what you
will do next. Do NOT add a preamble for trivial tool calls or for the final
answer itself.
J. Frontend AestheticsïŒAI slop åé¿ïŒ
<frontend_aesthetics>
Avoid common generated-UI defaults:
- Purple gradients on white or dark backgrounds.
- Inter / Roboto / system-ui as the only typeface.
- Center-aligned hero with a single CTA and a placeholder illustration.
- Low-density card grids with no real information.
Use a typographic system with at least one display face and one body face,
a constrained palette tied to the product, and explicit component states
(hover, focus, disabled, loading, empty, error).
</frontend_aesthetics>
K. Customer-Facing Persona
Persona: <role, e.g. senior support agent>.
Channel: <e.g. in-app chat / email / SMS>.
Register: <e.g. warm, plain-spoken, no jargon>.
Format: prose only â no headers, no bullet lists, no bold.
Length: max <N> words.
Personality: assume the user is competent and acting in good faith;
respond with patience, respect, and practical helpfulness.
L. Coding Agent Header
<role>You are a software engineering agent working in a real codebase.</role>
<output_contract>
- Implement the requested change as a minimal patch.
- Add or update unit tests covering the new behavior.
- Run the most relevant validation: unit tests, type checks, lint, smoke tests.
- Report the validation command and its result.
- If validation fails, fix or roll back; do not paper over with broad excepts.
</output_contract>
<default_follow_through_policy>
Prefer making progress over stopping for clarification when the request is
already clear enough to attempt. Surface assumptions in the final report.
</default_follow_through_policy>
M. Long Reference DocsïŒåæ¹é
眮ïŒ
<reference_docs>
<document index="1" source="policy-2026-Q1.md">
{{POLICY_TEXT}}
</document>
<document index="2" source="account-record.json">
{{ACCOUNT_JSON}}
</document>
</reference_docs>
åç
§ããã¥ã¡ã³ã㯠Identity / Instructions ã®çŽåŸïŒuser input ããåïŒã«çœ®ããåå©çšãããéçéšåãåæ¹ã«åºãããš prompt caching ãå¹ãã
Anti-Patterns
- process-heavy 㪠step-by-step: ãFirst A, then B, then Cããšæžããš GPT-5.5 ã窮å±ã«ãªããoutcome ãæžã
- 絶察ã«ãŒã«ã®æ¿«çš:
ALWAYS / NEVER / MUST / ONLY ãè£
食çã«äœ¿ããšãæ¬åœã«å®ãã¹ã invariant ãåããã
- stop rules ã®æ¬ åŠ: ãã〠retry ããããã〠ask ããããã〠stop ããããç¡ããšãagentic èšå®ã§æŽèµ°ãããã
- silent default ã§èª€éåã: äŸå€ãæ¡ã朰ããããevidence ãªãã§ãnoããšçããã®ã¯çŠå
system: ããŒã«å: Responses API / GPT-5.5 ã§ã¯ developer ããŒã«
xhigh ãããã©ã«ãå: ã³ã¹ã / ã¬ã€ãã³ã·ãèšãããé·æé agentic ã¿ã¹ã¯éå®
- ãã³ã°ããã¥ã¡ã³ããæ«å°Ÿã«çœ®ã: cache å¹çãèœãšããåæ¹ã«çœ®ã
- markdown å
šé¢çŠæ¢ã ãæžã: ãçŠãã formattingããåæããlength äžéã䜵èš
- åµäœã§åºæåè© / ææšãæé : source-backed ãš creative wording ãåé¢ããã¬ãŒããæžã
<example> ç¡ãã§è²Œã: æ§é åãããªããš few-shot ãå¹ãã«ãã
- phase ã user ã«ä»ãã: assistant åŽã®ã¿ã«ä¿æãuser ã¡ãã»ãŒãžã«è¶³ããŠã¯ãªããªã
GPT-5.5 vs GPT-5 / 5.1 ã®å·®åïŒç§»è¡æã®æ³šæïŒ
| é
ç® | GPT-5 / 5.1 | GPT-5.5 |
|---|
| reasoning effort ã®ããã©ã«ã | æç€ºãç¡ããšäžå® | medium ãããã©ã«ã |
| æäœã®ããã©ã«ã | éåžž | concise & direct |
| phase parameter | ä»»æ | Responses API ã§éèŠ |
| outcome-first æšå¥šåºŠ | æšå¥š | åŒ·ãæšå¥šïŒprocess-heavy ã§æ§èœäœäžïŒ |
| markdown ã®åºæ¹ | å€ã | æ§ãã |
| customer-facing ã®æž©ãã | èªç¶ã«åºã | personality æç€ºãèŠã |
GPT-5 / 5.1 ã§åããŠãã process-heavy 㪠prompt ã¯ãGPT-5.5 ã§ å¿
ã outcome-first ã«ãªãã¡ã¯ã¿ããããstep 1 ã§ããããã»step 2 ã§ãµãããã®ãããªåããåºãããã
Bottom Line
ãGPT-5.5 ã¯ã·ãã¢ã®ååãdestination ãæž¡ãã° path ã¯èªåã§éžã¶ãjunior åãã®çްããæé æžãæž¡ããšãããã£ãŠé
ã / æµ
ããªããã
outcome ãå®çŸ©ãããconstraints ãçµããstop rules ãå¿
ãæžãã ããã GPT-5.5 ãæå€§éåŒãåºãæ¹æ³ã