在上一篇文章《让 Agent 读懂你的身体：数据获取篇》中，我花了很大的篇幅讲清楚了第一个问题：Agent 怎样拿到数据。如果照着实践完，你大概率已经收获了一份存储了运动健康数据的 SQLite、Markdown 或 JSON 文件。
此时，大多数人大概就忍不住让 AI Agent 来帮忙分析一下文件了——
「Claude，这是我的健康运动数据，帮我做一下详细的分析。」
我自己起初也是这么做的。在简短的提问之后，Claude 就能生成一个精美的 HTML。上面有每个动作的训练趋势、恢复状态的图标以及按照日历排列的训练日历图等。信息排得很满，但研究一下就会发现，它除了好看，并没有什么用。
这就是问题所在——AI Agent 擅长描述，而如果想让它分析实际问题并给出建议，比如下一次训练该加多少重量，则需要用户自己对需求的判断与适当引导。
所以，在这篇文章中，我会根据我的实际经验，来详细回答后面的两个问题：拿到数据以后，怎样让它正确地处理？它给出的结论能信几分？
注：目前我这套方法只适用于辅助日常的运动训练，不建议用于排查身体的实际症状，也不应用于应对医疗情况。
怎样让 Agent 正确处理数据
我的整体方案
以我自己为例，我用：
- Apple Watch 记录运动训练数据，包括有氧和抗阻训练；
- Oura Ring 记录睡眠、夜间恢复以及日常活动数据；
- 训记 App 专门记录抗阻训练的容量和组数（同时使用训记和 Apple Watch 不会重复计入运动数据）
其中，Apple Health 和 Oura 的数据都接进第三方服务 Freddy。它通过 Oura 的官方接口授权读取，所以 Oura 自己算出的准备指数（Readiness）等算法数据也能拿到。
接着，通过 Agent 预先写好的 Python 脚本，经 Freddy 和训记的 API 将数据拉取到本地，存进 SQLite 数据库。而体脂、体重这类低频指标和训练感受等，我会用文字或图片的形式直接发给 Agent，由它整理后录入数据库。
上述项目搭建在家里的 NAS 上，Python 脚本定时拉取最新的生理和训练数据；网页看板设置好端口，通过局域网地址就可以直接访问。
去健身房之前，我用手机或电脑通过局域网连上 NAS 触发任务。常规任务在网页看板上留有按钮，可以一键启动；复杂的任务则通过终端里的 CLI 工具对话进行。触发后，AI Agent 读取最新的身体数据和训练记录，先由两个「教练」给出动作、组数和重量建议，再由「循证教练」核查训练量和强度。最终由我自己确认后，将计划作为一条训练记录，通过 API 写进训记。
到了健身房，我只要打开训记照着计划练，并给每组动作备注余量和训练感受。这些内容最终会通过 API 回传给 AI，完成一次循环。
把这些环节串起来并不复杂。真正需要花心思的，是给 Agent 建立一套「行为规范」，让它稳定地分析已有的数据并输出建议。方法很简单，搞个 Skill 就行。
为什么不用现成 Skill
「不要重复发明轮子」是软件开发领域的至理名言。一般来说，我们应该先看看社区里已有的 Skill，如果它们可以覆盖需求，直接套用即可。
其中，Coach Paddy 相对最接近我个人的需求。它内置了数据接口，能够读取 Garmin、COROS、WHOOP 和 Apple Watch 的数据，力量训练数据可以选配训记的 API；记忆存在用户的 Obsidian 里，包括运动员档案、训练计划和每日生理数据等。它会给出准备度评分，用来决定当天全力、降级还是休息，也能提供训练计划、进行每周复盘，算是一套相对完整的解决方案。
Coach Paddy 还有个值得借鉴的机制：出现数据缺失时，准备度评分就会降级，替换为 1 到 5 分的主观精力评分。
我最终没有选择 Coach Paddy。首先，我不需要它的数据接口模块。而且它虽然自称支持多项运动，分析的重心却在耐力训练和赛事备战，有比赛信心分、比赛倒计时等功能，和我的应用场景不吻合。此外，项目的 Star 和 Fork 不多，说明实际效果可能验证不足。
其实，现有的 Skill 多多少少都有类似的问题。因为训练是一件很个人的事情：训练目标、关注的指标、可用的器械、作息习惯、运动体系等，人人都不一样，现成的 Skill 很难完美适配。同时，健康和训练类 Skill 的关注度普遍较低，我目前找到的最多也不过 200 多个 Star，使用反馈大都不足。
注：
- 统计时间为 2026.10.16；
- Stars 一栏中的「-」表示该 Skill 为项目的子文档，没有单独的 Star；
- 类型一栏是我的主观分类，仅供参考。
所以，如果你能接触到顶级模型（例如 Claude Fable 或 ChatGPT Astra 等），不如让 Agent 吸取现成 Skill 的可取之处，在此之上定制一个适配个人习惯的 Skill。
定制 Skill，最难也是最关键的一步就是把需求梳理清楚，然后制定详细的实施方案。我会先用 grill-me 技能把需求问清楚。它来自 Matt Pocock 的开源 Skill 仓库，作用是让 Agent 反过来向你提问：从粗略的想法出发，一直追问到形成确定而详细的结论。坏处就是有时候你会觉得，它怎么还有这么多问题。
统计数字交给代码
在我与 AI 的第一轮对话中，AI 就问出了一个非常关键的问题：计算和决策分别由谁来做？
本文由 AI宝库自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
In the last article, "Agent reads your body: Data Access," I spent a lot of time making the first question clear: How did Agent get the data? If you're done, you probably have a SQLite, Markdown or JSON file that stores physical health data.
At this point, most people probably couldn't help but ask AI Agent to help analyze the file.
"Claude, this is my health-motion data. Help me with a detailed analysis."
I did it myself at first. After a short question, Claude can produce a fine HTML. It contains training trends for each action, an icon for recovery and a calendar-based training calendar. The information was very full, but a study would show that it was of little use but good-looking.
This is the problem -- AI Agent is good at describing it, and if it is to analyse the actual problems and give advice, for example, on how much weight to add to the next training, it will require the user ' s own judgement of needs and proper guidance.
So in this article, I will answer the following two questions in detail, based on my practical experience: When you get the data, how do you make it work? How many points does it give?
Note: For the time being, my methodology applies only to the day-to-day exercise training and is not recommended for the examination of physical symptoms or for medical treatment.
How to get Agent to process the data correctly
My overall plan.
Take my own example. I use:
- Apple Watch records sports training data, including aerobics and resistance training;
- Oura Ring records sleep, night restoration and daily activity data;
- Training App dedicated to recording the capacity and the number of resistance trainings (the use of both the training and Apple Watch will not be duplicated in campaign data)
Among them, Apple Health and Oura data are fed into third-party services, Freddy. It is authorized to read through the official interface of Oura, so that algorithms such as Reardines can be obtained from Oura’s own calculations.
Then, by prewritten Python scripts prepared by Agent, the data is drawn locally through Freddy and the API of the SQLite database. In the case of lipids, low-frequency indicators such as body weight and training feelings, I will send them directly to Agent in text or pictures, which will then be compiled and recorded in the database.
The above-mentioned project is built on a home-based NAS, where Python scripts regularly extract up-to-date physiological and training data; the web board is set up with a port to be accessed directly through the LAN address.
Before I went to the gym, I used a cell phone or a computer to connect to NAS. Regular tasks leave a button on the web board and can be started with a single key; complex tasks are performed through the CLI tool dialogue in the terminal. After the trigger, AI Agent reads up-to-date body data and training records, with two " coaches" first giving suggestions for action, group numbers and weight, and then " evidence-based coaches" verifying the number and intensity of training. I eventually confirmed the plan as a training record, through API.
When I get to the gym, I'm just going to open up the training to follow the plan, and I'm going to fill out the extras and the training feeling for each group. These will eventually be sent back through API to AI to complete a cycle.
It's not complicated to tie them together. What really needs to be thought-out is to create a set of "behaviours" for Agent, so that it can analyse the available data and output recommendations in a stable manner. It's a simple way to get a Skill.
Why don't you use the ready-made Skill?
"Don't repeat the invention wheel" is the best word for software development. In general, we should look at the Skills that already exist in the community, if they can cover the needs and just apply them directly.
Of which Coach Paddy is relatively closest to my personal needs. It incorporates data interfaces that allow access to Garmin, COROS, WHOOP and Apple Watch data, power-training data to select the API with typologies; and Obsidian, where memory exists, including athletes ' archives, training programmes and daily physiological data. It gives a level of preparedness, which is used to determine whether the day is full, downgraded or rest, and provides training programmes and weekly reviews, which is a relatively complete solution.
Coach Paddy also has an interesting mechanism: when data is missing, the preparation rating is downgraded and replaced with a subjective energy rating of 1 to 5.
I finally didn't choose Coach Patty. First, I don't need its data interface module. And even though it claims to support a number of campaigns, the focus of analysis is on endurance training and preparation for events, with skills such as game confidence points, game countdowns, etc., which do not match my application. In addition, the Star and Fork projects are limited, suggesting that the actual effects may not be sufficiently validated.
Actually, the existing Skill has similar problems. Because training is a very personal thing: training goals, indicators of interest, available devices, pacing habits, sports systems, etc., are not the same for everyone, and ready-made Skills can hardly fit perfectly. At the same time, health and training Skill is generally of low interest, and I now have just over 200 Stars, most of which are under-used.
Notes:
- Count time is 2026.10.16;
- The "-" in the Stars column indicates that Skill is a sub-document of the project, without a separate Star;
- The type column is my subjective classification for reference purposes only.
So if you have access to top-level models (e.g. Claude Fable or ChatGPT Astra, etc.), then why don't you let Agent learn what's right for Skill and customize a Skill that suits his personal habits.
Customizing Skill, the most difficult and critical step is to sharpen demand and then develop detailed implementation programmes. I'll first use the Grill-me skills to clarify the need. It's from Matt Pocock's open source, Skill's warehouse, and its role is to get Agent to ask you the other way around: from a cursory point of view, to a definitive and detailed conclusion. The bad thing is sometimes you think it has so many questions.
The statistics are to the code.
In the first round of my conversation with AI, AI asked a very critical question: who does the calculations and decision-making?
Auto-fetched and translated by AI Gem Box. All rights belong to the original author.