Gemini 3.6 Flash：PromptsRush 的长上下文、多模态与 Agent 提示词

一句话结论

这份 42 条提示词清单把 Gemini 3.6 Flash 的优势落成可复制的任务契约：把完整语料放入 1M 上下文、要求逐条引用、明确输出 schema，并用独立验证和分级路由控制成本。

适用场景

• 适合的任务：全量文档检索、合同/政策核对、视频音频分析、浏览器自动化、代码库审计、结构化批处理和有来源的研究简报。

• 不适合的任务：需要模型自行决定权限、自动提交不可逆操作、没有可核验输入的开放式事实判断。

• 适用的模型版本：Gemini 3.6 Flash；页面数据以 2026 年 7 月为准。

• 适用的客户端、Agent 或 API：Gemini API、Google AI Studio、具备文件/屏幕工具的 Agent harness；浏览器提示词假设 harness 能提供屏幕访问。

• 推荐的推理档位和参数：来源建议使用输出契约和引用约束；没有给出统一 API 参数。对代码或 Agent 任务可先用 thinking_level=medium，困难任务再升到 high，并记录额外 token。

可直接使用的内容

以下为原文中可直接复制的代表性完整提示词；原文另有 42 条模板，选择这些是为了覆盖模型的主要输入模态和 Agent 工具类型。

1. 全语料问题（Whole-Corpus Question）

I am attaching [N] complete documents totaling roughly [SIZE]: [ATTACH THE FULL SET].
Question: [WHAT YOU NEED TO KNOW]
Answer from the attached material only. For every claim, cite the document name and the section or page it came from. If the answer requires combining facts from multiple documents, show the chain. If the corpus does not contain the answer, say so plainly rather than reasoning from general knowledge.

2. 矛盾查找（Contradiction Finder）

Read this complete document set: [ATTACH ALL].
Find every place where two documents state incompatible things: different numbers for the same metric, conflicting policies, dates that cannot both be true, procedures that contradict each other.
Output a table: claim A (with source), claim B (with source), type of conflict, and which one is more likely authoritative based on recency or document status. Do not report differences that are just different levels of detail — only genuine contradictions.

3. 视频摘要（Video Digest）

Watch this video and produce a working digest: [ATTACH VIDEO].
Deliver: a one-paragraph summary, a timestamped outline of the major sections, every claim or number stated on screen or in narration (with timestamp), any on-screen text worth capturing, and the three moments most worth watching directly.
Timestamps must be accurate — I will spot-check them.

4. 有边界的浏览器任务（Bounded Browser Task）

Complete this task in the browser: [DESCRIBE THE TASK].
Done means: [MEASURABLE END STATE — e.g. the form is submitted and the confirmation number is captured].
Rules: take a screenshot before any irreversible action and describe what you are about to do. Never enter payment details, delete data, or send anything without explicit confirmation. If a page differs from what you expected, stop and describe the difference rather than improvising.
Report the confirmation evidence at the end, not just a claim of success.

5. 完整仓库 Bug 排查（Full-Repo Bug Hunt）

Here is the complete repository: [ATTACH]. Symptom: [WHAT HAPPENS]. Expected: [WHAT SHOULD HAPPEN].
Trace the actual execution path through the real code — not a plausible one. Identify the exact file and line where behavior diverges from intent, explain why, and give the minimal fix.
If multiple causes are consistent with the symptom, rank them and give the one-line check that distinguishes them. Do not propose rewriting modules that are not implicated.

6. 结构化提取（Structured Extractor）

Extract structured data from the following item: [PASTE OR ATTACH ITEM].
Return JSON matching exactly this shape: [PASTE YOUR SCHEMA].
Rules: never invent a value — use null for anything not present in the source. For every extracted field, the value must be traceable to specific text in the input. If the input appears to be a different document type than expected, return {"error": "unexpected_type"} and nothing else.

7. 有来源的研究简报（Sourced Research Brief）

Research this question and build a brief: [THE QUESTION]. Decision it informs: [WHY IT MATTERS].
Structure: what is established (with source type), what is contested (both positions, one line each), what is unknown, and your synthesis with a confidence level.
Every number gets a source. Separate facts from your inference explicitly. Where you are unsure, write "unsure" — do not fill the gap with plausible filler. I would rather have six solid facts than twenty vibes.

测试/工作流步骤

原文给出的三条可直接采用的链路：

1. 文档智能：Whole-Corpus Question → Contradiction Finder → Needle Audit；先全量读，再对矛盾和关键术语做穷举核验。

2. 多模态到行动：Recording-to-Actions → Structured Extractor → Bulk Quality Gate；把音视频决策转成带 schema 的任务记录，再做质量门禁。

3. 批量内容：Voice-Matched Draft → Ruthless Edit（新会话）→ Repurposing Fan-Out；把生成、编辑和多平台改写拆开，避免首轮输出直接发布。

原始证据与观察

• 原文明确建议把完整 corpus 放入上下文，而不是只贴预先筛选的片段。

• 每个抽取模板都要求来源指针、null 或 unsure，把“没有证据”与“模型不知道”显式区分。

• 浏览器模板要求不可逆操作前截图、最终提交前确认，并在页面偏离预期时停止。

• 原文把 Gemini 3.6 Flash 定位为 volume tier，并建议 hardest 20% 任务升级到更强模型；这是路由建议，不是官方保证。

适用边界

• 这些是可复用提示词，不是 PromptsRush 对每个任务的受控实验结果；应在自己的文件、权限和工具上回归。

• 浏览器提示词不能替代客户端的权限控制；“停止并报告”必须由 harness 强制执行。

• 1M 上下文和视频/音频能力并不保证每个文件类型、时长或大小都能成功处理，需按 API 当前限制测试。

• 缓存节省和输出 token 结论依赖 provider 计费，不能把文章中的价格直接套到其他路由。

来源摘录或观察（仅做合规短引）

原文把核心规则写成 “Cite or say unsure”；这应与完整来源、字段 schema 和人工核验一起使用。