## I

I. First, Let Me Tell You About the Version We Screwed Up

Our team isn't big—five or six projects running in parallel. Every Friday afternoon, collecting progress from each project and piecing together a weekly report was the two hours I dreaded most. When I heard AI could do this, I eagerly built the first version: I wrote a script that dumped all projects' chat logs, task board exports, and commit logs into a large language model, with a single instruction: "Help me summarize this into this week's team weekly report."

The result? The output looked presentable at first glance, but on closer inspection, it was full of problems:

• It included a stale task from three weeks ago as "this week's progress";
• One project was clearly delayed, yet the report said "overall steady progress"—the model was fabricating transition sentences;
• Each project's description style was completely inconsistent—some got a 200-word laundry list, others just one sentence;
• When leadership asked "what are the risks for Project B," the answer was nowhere to be found in the report.

The moment the first version got shot down, I realized: the problem wasn't the model—it was that I treated it as "a person who can write articles" rather than "a processing pipeline that requires constrained inputs." It's the same principle as dumping raw logs into an AI—garbage in, garbage out.

II. Three Common Failure Patterns

Looking back, there are three common pitfalls when using AI to summarize weekly reports:

Pitfall 1: Dirty inputs. Throwing raw chat logs and messy task titles at the model means it can't determine "what belongs to this week" or "what the key information is," so it just improvises based on linguistic plausibility. The more fluent the fabricated output, the more dangerous it is.

Pitfall 2: No structural constraints. Just saying "summarize this for me" without specifying how many items per project, what format each item should follow, or what priority ordering to use. The more freedom the model has to improvise, the less controllable the output.

Pitfall 3: One big batch. Summarizing all projects in a single request means the model starts losing details as the context grows longer, and smaller projects' progress often gets "averaged" out.

III. The Approach That Worked: AI as "Sorter + Formatter," Humans Do the Final Gatekeeping

For the second version, I changed my approach: humans move their work to the front, AI's work moves to the back, with structure in the middle.

Step 1: Define the Weekly Report "Skeleton" First

Agree on a weekly report template with the team: each project gets four fixed sections—Completed This Week, Planned for Next Week, Risks & Blockers, and Decisions Needed. Each section has at most three items, one sentence each. This skeleton is defined by humans, not delegated to AI.

Step 2: Bucket by Project, Process One at a Time

Instead of swallowing everything in one request, run a separate pass for each project. The inputs are also cleaned: only this week's task changes, milestone nodes, and discussions with designated tags are included, with chitchat removed.

Step 3: Run AI with the Same Prompt Template

This is the key. The template is below—copy and adapt it directly:

你是一个项目周报整理助手。请严格按以下规则处理我提供的项目原始记录：

【背景信息】
项目名称：{project_name}
统计周期：{start_date} 至 {end_date}
项目当前阶段：{project_phase}

【原始记录】
{raw_materials}

【输出要求】
1. 按以下四个栏目输出，每栏最多3条，每条一句话（不超过40字）：
   - 本周完成
   - 下周计划
   - 风险与阻塞
   - 需要决策的事项
2. 只使用原始记录中明确出现的信息，禁止推测、补全或美化。
3. 无法归类的信息单独放在"未归类信息"一栏，不要强行塞进上述四栏。
4. 涉及时间的信息，标注原始记录中的日期。
5. 某一栏目没有内容时，写"无"，不要编造。
6. 输出格式为Markdown表格。

The most useful parts are rules 3 and 5—giving AI permission to say "I don't know" is the single most effective trick for eliminating hallucinated fabrication. Since adding these two rules, fabricated transition sentences like "overall steady progress" have never appeared again.

Step 4: Humans Only Review the "Uncategorized" and "Risks" Sections

Each project's AI output flows back to me, and I focus on two things: whether any major items were missed in the uncategorized information, and whether the risks section was misled by the wording of the original records. Reviewing each project takes about two minutes.

IV. Before-and-After Comparison

Total time dropped from over two hours to under twenty minutes, but that's not the biggest gain. The biggest change is that the report became interrogatable—because every item comes from structured inputs, when leadership asks "where does this conclusion come from," I can trace it directly back to the original records. This was impossible with the first version's pure "AI free improvisation" model.

As for cost, the weekly processing volume isn't large, so even with token-based commercial models, the expense is minimal (check each provider's official pricing page for specifics). If your team is already managing calls through a unified API gateway, these periodic tasks can be set up as scheduled jobs directly—saving even the manual trigger.

V. A Few Tips for Adapting This Approach

This method isn't limited to weekly reports—it applies to any scenario involving "periodically consolidating information from multiple sources": weekly customer feedback summaries, monthly bug trend reports, multi-channel operations data digests. The core principles are just three:

1. Define the structure first, then feed the data—the output format is a human-defined contract, not AI's improvisation;
2. Process in buckets, don't throw everything together—one independent call per subject yields far higher quality than mixing them all up;
3. Give AI an escape hatch to say "none"—forcing every section to be filled only encourages the model to fabricate.

If you don't yet have a convenient model-calling environment, or want to manage these periodic consolidation tasks through a unified entry point, give https://api.thistoken.ai/register a try to get periodic AI tasks like a weekly report bot running reliably.

---

Tired of juggling provider integrations? Register at https://api.thistoken.ai/register and call every model through one base_url.