An Old Problem for Managers

In my years of running technical reviews with teams, what troubles me most isn't the proposals themselves, but the pattern of "documents written painstakingly, decisions made hastily."

A typical scenario: an engineer spends a week writing a thirty-plus-page technical design document, complete with architecture diagrams, technology comparisons, and performance benchmark data. At the review meeting, the attending managers, product folks, and ops staff spend ten minutes each skimming it, and then the first question they ask is: "So which option do you recommend, and why?"

The thicker the document gets, the lower the decision efficiency. This isn't a problem with any particular team—it's a problem of role mismatch. The person writing the document wants to cover everything; the people reading it only want decision-ready information.

I once tried to enforce a "must read the full document before the review" rule, and everyone just started pretending they had read it. Then I flipped my thinking: instead of requiring everyone to read thirty pages, why not compress the thirty pages into one, and let AI handle that compression step?

What I Want Isn't a Summary—It's a One-Page Decision Document

Let's be clear about the goal first. What I need isn't an "abbreviated document," but a one-page summary that supports decision-making, which must include:

1. What's being decided: state in one sentence what needs to be settled at this review
2. Comparison of candidate options: three lines per option—core approach, advantages, risks
3. Key trade-offs: where the compromises lie in terms of cost, timeline, and maintainability
4. Recommended conclusion and preconditions: which option is recommended, and what assumptions it depends on
5. List of open questions: questions that need input from others before the meeting

Extracting these five blocks of information manually from a long document takes at least an hour each time, and it tends to carry bias—when proposal authors summarize their own work, they often downplay the weaknesses of their own options. AI's value here is precisely that it "has no agenda," and you can set a fixed extraction framework so the output structure is consistent every time.

The Actual Workflow: Three Steps, Twenty Minutes

Step 1: Document preprocessing. Break the proposal document (PDF or Markdown both work) into processable text. If the document contains lots of charts, I have the engineer verbally add a "chart explanation" to feed in as well—otherwise AI will miss key information in the figures.

Step 2: Structured extraction. Run it through a fixed prompt template to get a first draft of the one-page decision summary.

Step 3: Human proofreading and distribution. This step cannot be skipped. I verify that the key numbers and risk descriptions in the summary match the original document, especially any risk points that AI has softened. The proofread summary is sent out with the meeting invitation two days in advance, so attendees arrive with questions rather than confusion.

Throughout this workflow, AI handles the "information compression and structuring"—the most time-consuming part—while humans retain "accuracy gatekeeping" and "final decision-making."

A Reusable Prompt Template

你是一位技术管理者的决策分析助手。请阅读我提供的技术方案文档，
输出一页决策摘要，严格按以下结构：

【决策事项】用一句话说明本次需要拍板的内容。
【候选方案对比】列出文档中的所有候选方案，每个方案包含：
  - 核心思路（1句话）
  - 主要优势（最多2条）
  - 主要风险（最多2条，必须来自原文，不得美化）
【关键权衡】从成本、工期、可维护性三个维度，说明方案间的取舍。
【推荐结论】如果文档有明确推荐，复述其结论和依据；
  如果没有，客观指出各方案的支持论据强度，不替作者做决定。
【决策前提】列出该推荐成立所依赖的假设（如预算、人力、时间窗口）。
【开放问题】列出会前需要其他角色补充输入的问题，标注应回答的角色。

要求：
- 只使用原文信息，不得编造数据或补充原文没有的判断
- 每个部分的信息必须可追溯到原文对应章节
- 总输出不超过500字

I maintain this template openly within the team, and anyone can propose improvements. Using a fixed template rather than writing prompts from scratch each time has the benefit of producing summaries with a stable structure—after a few reviews, people get faster and faster at reading them.

Before and After

Before using AI:

• Before review meetings, attendees' actual average reading time was under ten minutes
• Much of the meeting was spent repeatedly confirming "what did page such-and-such of the document say"
• Open questions only surfaced during the meeting, causing decisions to be postponed to the next meeting
• Summaries depended on the proposal author writing them, making bias hard to avoid

After using AI:

• Summary production went from about an hour to a few minutes of generation plus ten minutes of human proofreading
• Review discussions now start directly from the "trade-offs," and meeting length has shrunk noticeably
• Open questions are distributed to the relevant roles before the meeting, so most issues arrive at the meeting already answered
• The fixed template produces a consistent structure, making review materials comparable across projects

There was an unexpected benefit for risk management too: after requiring AI to "not sugarcoat risks," the risk descriptions in the summaries came out more blunt than what authors wrote by hand, which actually led to several more candid discussions. Of course, AI occasionally over-generalizes and distorts details, so the human proofreading step can never be skipped—AI handles speed, humans handle accuracy.

A Few Lessons Learned

1. The template should be public and iterated on. The prompt template is a team asset, not a personal skill.
2. The summary cannot replace the original document. Use the summary as your entry point for decision-making, but return to the relevant sections of the original when deep details matter.
3. Watch out for hallucinated numbers. Any number that appears in the summary should be verified one by one against the original during proofreading.
4. For cost-sensitive small teams, you can choose different tiers of models to run this workflow depending on document length—refer to the official pricing page for specifics.

If you want to get this workflow running too, you'll need a stable large-model API service to support it. You can try this platform's registration portal: https://api.thistoken.ai/register

From a thirty-page document nobody finishes reading, to a one-page decision document everyone can discuss—this is something AI deserves to do for you.

---

Every example in this post runs with a single API key — get yours at https://api.thistoken.ai/register and start in minutes.