📖 英文と日本語訳（一文ずつ）

🎧 通し読み（全文）

リスニング・シャドーイング用の全文です。

OpenAI has released a prompt engineering guide for its latest model, GPT-5.5. Compared to the previous GPT-5.4 version, short "outcome-first" prompts that concisely state the desired result are reported to yield better results than traditional, lengthy prompts that specify processes in detail. Additionally, because inference efficiency has improved, it is recommended to first re-evaluate performance at low and medium levels before increasing the reasoning effort level. The guide highlights that GPT-5.5 can independently select the optimal path to a solution as long as a "destination" is provided. Reusing existing prompts designed for older models without modification is said to risk introducing noise through unnecessary instructions, which may narrow the model's search range and lead to mechanical responses. Therefore, it is stated that absolute imperatives such as "ALWAYS," "NEVER," and "must" should be used only for true invariants, such as safety rules and mandatory output items, while decision rules should be employed in situations requiring judgment. From the perspective of persona design, it is recommended to distinguish between the "personality" that defines a model's way of speaking and the "collaboration style" that defines its way of working, providing concise instructions for each. GPT-5.5’s default style is efficient, direct, and task-oriented. For interactive products such as customer assistants, it is adjusted to explicitly incorporate warmth and proactivity. A technique has been introduced to improve the perceived speed of streaming responses for long tasks involving tool calls, which involves instructing the system to first output a short "preamble" to acknowledge the request and communicate the initial step. This will improve the user's perceived experience regarding the time until the first visible token is displayed. The output format can be adjusted through the text.verbosity parameter, with the API's default value set to "medium." In ordinary conversation and explanations, it is considered desirable to use plain paragraph format as the primary structure, while limiting the use of headings and bulleted lists to situations where comparison or ranking is required. Regarding citations and the provision of evidence, it is recommended to explicitly state the "retrieval budget," a stopping rule for searches. Guidelines state that an initial broad search should be conducted using concise keywords, and a response provided without re-searching if sufficient evidence is found. Conditions for permitting a re-search include cases where essential facts are missing or a reference to specific documents is required. For documents containing creative elements, such as slides and customer-facing copy, it is required to distinguish between factual descriptions and creative expressions and to avoid fabricating concrete facts, such as product capabilities and customer outcomes. In verifiable tasks, it is also important to have a mechanism that allows the model to verify its own work. Coding agents are instructed to execute targeted unit tests, type checks, and build checks, and to inspect visual deliverables for layout issues or missing elements after rendering. The phase parameter introduced in GPT-5.4 remains effective. In long-running workflows, intermediate updates can be distinguished as "commentary" and the completed response as "final_answer." While the API automatically maintains state when using previous_response_id, the original phase value must be carried over when manually resending assistant output.

📝 学習のヒント

1. 1まず英文を読む — 知らない単語にあたりをつけてから音声へ。
2. 2一文ずつ確認 — 日本語訳と照合し、構文を理解する。
3. 3通し読み Normal で耳を作る — 内容を追いながらリピート。
4. 4Fast でシャドーイング — 口を慣らし、リスニング速度を上げる。
5. 5翌日に復習 — 1日空けて再聴すると長期記憶に定着しやすい。