> ## Content Index
> Fetch the complete content index at: https://sertan.com/llms.txt
> Use this file to discover other available public pages before exploring further.

# A Coach That Knows What It Cannot See
- URL: https://sertan.com/a-coach-that-knows-what-it-cannot-see/
- Published: 2026-11-18T00:21:47.000Z
- Updated: 2026-11-18T00:21:47.000Z
- Description: Two scheduled agents read my health and food exports every morning and coach me through the week. What makes them useful is a written brief, a few rules about what they must not do, and a way to say what they cannot see.
- Author: Sertan Yamaner
- Tags: Apps for One, Agentic Development, Software Engineering

The previous posts covered the two iOS apps I built to manage my exercise, track what I eat and export detailed Apple Health data to Google Drive. This one is about the agents that complete the loop. Two scheduled agents in Claude Cowork read my exports each morning: one runs at 6 AM to plan the day's exercise, and the other runs at 7 AM to review my food intake. Each also looks back over the week and recommends changes for the next one, the exercise coach on Sunday evenings and the food coach on Friday mornings.

The exercise coach has now run for twelve weeks and the food coach since September, and they have been a real help. I will describe how they are set up, because most of what makes them useful is a written brief, a handful of rules about what the coach must not do, a way for it to say what it cannot see, and a habit of correcting those rules whenever the coach gets something wrong.

The data they read comes from WeeklyHealthReport, the personal iPhone app from [the previous post](https://sertan.com/the-app-only-i-needed/), which started as a careful collector of my health data and has grown into the food tracker I use every day. I am not going to share my own measurements, but the context is that I have been losing weight while trying to keep my strength, with very little spare time for exercise.

![The two coaches and their lanes: the exercise coach and the food coach both read the same Drive, and the food coach reads the exercise snapshot only to time its advice](https://storage.ghost.io/c/9f/ae/9faeabff-311d-49ec-8e86-ce7ddfbe2c77/content/images/2026/11/lanes-board-1.png)

Two coaches, two lanes: the same Drive, separate jobs.

## The loop

One day from end to end goes like this:

- **Before bed**, I export the day from the app to a Google Drive folder the agents can read:
  - What Apple Health measured.
  - The treadmill segments from PacePrompt (my treadmill app, from [the first post](https://sertan.com/driving-a-treadmill-from-a-pdf-specification/)).
  - The food I logged and any note I typed.
- **At 6 AM** the exercise coach reads the newest export and its own last snapshot, the complete record it left at the end of the previous run. It decides today's session, rebuilds the board (shown below), a single page with today's session and the week, and writes a new snapshot.
- **At 7 AM** the food coach does the same for food, and stays silent unless there is something to act on.

During the day:

- I do the session, or part of it, or some other exercise, and my watch records whatever I did, through PacePrompt on a treadmill day and straight into Apple Health otherwise.
- I log what I eat in WeeklyHealthReport as I go.
- If I want to argue, I reply in the same conversation, where the coach answers and, when something changes, writes a fresh snapshot with the correction.
- The rules can be wrong and when that happens, I have the rules changed in the charter (the written brief) or in one of the operations manuals that describe the machinery, so the next morning starts from the corrected version.

What I really enjoy is how much a very simple setup has done for me. On the coaching side, the whole system is:

- Three scheduled tasks in Claude Cowork (the two morning runs and the exercise coach's Sunday review), each with a prompt of about five hundred words that only says where to start.
- One charter, two operations manuals and a template for the board.
- One snapshot per run.
- The daily export from the app.

No database, no server and no code behind the coaches themselves. The only software in the loop is the pair of apps that produce the data, which keeps the whole thing small enough to read: when a coach does something odd, I can open its last snapshot and see what it knew.

The complexity went into the apps instead, and WeeklyHealthReport in particular gives me several ways to log a food, two of which use a model through OpenRouter, with my own API key, which I enter in the app:

- **Barcode or search:** a barcode scan, or a search of the bundled food composition tables and Open Food Facts.
- **Label as text:** I paste the nutrition label as text and the app reads it on the phone; nothing is sent anywhere.
- **Label as a photo** (new this week): one cropped photo of the label goes through OpenRouter to a multimodal model, which reads the values and the serving size. Currently, both label routes read Traditional Chinese labels in the Taiwan layout only.
- **Web lookup:** I type a food name or paste a link to a product page. Through OpenRouter, one model searches and picks a source, the app fetches that page, a second model extracts the values, each tied to the exact text it came from, and the first model then checks that the page really describes the food I asked for. With a link, the search is skipped and that page is read on its own.

Whichever route I use, the model only fills in a form. Every route ends on the same confirmation screen, where I check the values against the source and correct them before logging, and values from the web stay labelled as estimates, which means the coaches only ever read what I have confirmed. The next post covers how these routes were built.

![The exercise coach's daily board on a strength day: Strength B with eight exercises and their demonstration links, a five to eight minute fallback, the week, and the demonstration library](https://storage.ghost.io/c/9f/ae/9faeabff-311d-49ec-8e86-ce7ddfbe2c77/content/images/2026/11/coach-board-strength-2026-11-15.png)

The exercise coach's daily board on a strength day, 15 November: today's session, the week and the video library. A public copy assembled from the real board and that day's plan, without my measurements.

## The brief: not another tracking project

The exercise coach works from a charter I wrote with an assistant at the end of August, when its data still came from an Apple Shortcut, shortly before I started the app. Its purpose section asks for a practical, adaptive programme that helps me exercise consistently despite limited time, and then says something that mattered to me: this is not intended to become a complicated tracking project. Minimise administration, prefer useful action over collecting more data, and work well conversationally from a phone, so that I can ask "What am I doing today?" or "I only have 10 minutes" and get a usable answer.

The rest of the setup follows from that. Strength sessions are compact, around fifteen to twenty minutes, at home with a kettlebell, bands, a handgrip and a treadmill. On a treadmill day the plan carries the session in two forms, a programme for the console and a line of plain English that I paste into PacePrompt, and the coach never asks which one I used. Every strength session has a five to eight minute fallback and every treadmill day a short version, and the fallback counts as useful training. If I say I am busy, the coach adapts immediately instead of telling me to find more time. And if I repeatedly fail to complete the plan, the coach reduces or reorganises the programme; it does not treat that as a failure of motivation.

Of all of this, the fallback rule turned out to matter most to me: a short session on a busy day counts as part of the plan.

The charter's own summary follows. I quote it, and the manuals further down, as they read during those first weeks; the wording was tightened in November.

> Treat the programme as an adaptive engineering problem: observe what actually works, retain useful things, adjust constraints, and progress conservatively.

## Writing the coaching style down

I wrote the coaching style into the charter as instructions instead of hoping for it:

- Be concise and practical, with no motivational speeches.
- Tell me what to do, how much and any important technique or safety points.
- Explain reasoning only when the programme changes, when a weekly trend is reviewed, when a new exercise is introduced or when I ask why.
- Do not shame or moralise about missed workouts, food, weight or adherence.

Two smaller rules are about friction. The first is what the charter calls the felt-OK default:

> Unless I say otherwise, the exercises felt ok. A completed session with no comment is recorded as completed and comfortable, and is never chased with a question.

The second is about missed days:

> Moving a Tuesday workout to Wednesday because life intervened is not failure.

Its companion rule is that a missed workout is never made up by making the next one longer or harder.

## When I do more than the plan

Most of those rules are about weeks when I do less than the plan. In the first weeks I did the opposite, a walk on a rest day or extra work on top of a session, and one evening in the second week the coach, then running on Claude Opus 5 (sassy, by some accounts, and I mean that as praise), said so:

> One thing I am noting rather than acting on. **That is now the fourth time in nine days you have done more than the plan asked**, and **this time you went looking for extra work elsewhere**. Fifteen minutes twice a week may simply be too small a prescription for where you actually are. Time is still the constraint so I will not just make sessions longer, but Sunday's review should look at whether a denser session or a fourth short day serves you better than a plan you keep overshooting.

I found it funny, a coach keeping count of my overachieving, but it was also right about the plan. The week was by then an experiment with three sessions instead of two, and when Sunday came the review kept it at three and put more work inside them instead of more minutes. The coach's notes that Sunday said that finishing all three was a fact and not a reason to add a fourth.

Two lessons went into its own record in those first weeks (its notes refer to me in the third person, and it writes its headlines in capitals, so I have lowered its voice here):

> When he repeatedly does more than the plan asks, comfortably, the plan is wrong.

> When he does more than the day asked, record it and leave it alone. It is information about what he wants, not a licence to add work or a debt to balance elsewhere.

Since then, extra work is recorded and left alone on the day, and the weekly review decides whether the plan should grow.

## How the coaches remember

Lessons like those only last because they are written down, since each run starts in a fresh session with no memory of the previous one. Everything a coach knows comes from files in Google Drive, and each has one writer:

- The app writes the daily export and the food log's daily file. Nothing else changes them.
- Each coach writes its own state snapshot at the end of a run, new and complete instead of an edit of the old one, and may read the other's but never change it. The exercise coach's record calls this "one owner per fact".
- The exercise coach rebuilds the daily board, a published page built from a template in the folder, and the food coach never touches it.
- The charter and the manuals change only when I ask. I say what should be different, the coach writes a new version and files the old one with a "superseded" title, which the scheduled runs are told never to read.
- The stock list of what I keep in the house is mine, and the supermarket receipts are mine to add.

Because a snapshot is never edited, the history is append-only and every plan can be traced to what the coach knew at the time. The charter is now on its thirteenth version and the food manual on its fourteenth, and each amendment records what changed and why.

The manuals need the same upkeep as code: both went on describing the Apple Shortcut for two months after the app replaced it, until I had them updated in early November.

One line from the charter describes how I want it treated: it is *a starting point rather than a specification to be defended*, and it is *not to be quoted back at me as the reason something cannot be done*.

## Knowing what the data cannot say

The daily export marks every block with an availability state, and the coaches read it with a few fixed rules:

- A block with no data is a missing observation, never zero.
- An absent workouts block means the coach cannot see a session, not that I skipped one.
- Missed, unobserved and partial sessions are kept apart.
- A same-day file is always partial, because the export is cut before bed.
- Sleep in the file always belongs to the previous night.

The exercise coach, by its own account, handled a stretch from mid-October to early November with no exports in exactly that way: it carried on planning, counted nothing as missed and mentioned the gap once instead of every morning.

The food coach came later: when food planning grew beyond a section of the exercise charter, I split it into its own scheduled task with its own manual and a small plan, a rotation of dinners and a standing shopping list. It learned its own version of these rules the hard way, and its manual records both lessons in plain words.

**A floor is not a level.** In the middle of September, when my food data still came from a separate nutrition app through Apple Health, it reported that a mineral had been under its reference value three days running, and twice invented a cause for it. The truth was that the nutrition app misses some nutrients for some foods, so a logged vitamin or mineral total can only understate, not overstate. The rule that replaced it: a floor above its reference is a real finding, while a floor below it says nothing, because a shortfall and a database gap look the same.

![A logged nutrient total above the reference is a real finding; below the reference it says nothing, because the true total could be higher](https://storage.ghost.io/c/9f/ae/9faeabff-311d-49ec-8e86-ce7ddfbe2c77/content/images/2026/11/floor-board-1.png)

A floor is not a level. Illustration only, no real figures.

**A deduced shortage is not an observed one.** The first of those invented causes was an empty fridge: the food coach had worked out, by subtracting planned portions from an old supermarket receipt, that food was missing from the kitchen, and it warned me about it, when in fact I had the food. The manual recorded the lesson in those words, and since then everything on my stock list is assumed to be in the house unless I say otherwise.

The note at the end of that manual sums up both mistakes as one: before saying I am short of anything, ask whether the figure can bear the claim. An earlier note, from a day when I felt I was missing every target, made the same point about goals: a target that cannot be hit is a fault in the target.

These are the same rules I wrote for the app itself, arrived at from the other side. Unknown is never zero, partial totals are known minimums, and estimates are labelled as estimates.

In November the export started to carry individual glucose readings from a sensor I am trying, each with its time, so the food coach can set a meal beside what followed it. One of the first reviews to use them kept to the same habit: where two foods were logged within the hour, or a meal overlapped a training session, it said the effects could not be separated, and it offered one or two small things to try and compare in place of a conclusion.

Both coaches have also asked for things they could not see. The exercise coach wants a structured record of how a session went: which planned session it was, whether it was done in full, partly or as the fallback, and how hard it felt. At the moment that judgement rests on my notes and the felt-OK default, although PacePrompt records each treadmill session's segments as they were run and since early November the export carries them, so on treadmill days part of the request is already answered.

The food coach wanted the per-nutrient completeness flag from the previous post, and it now has it. From late September it worked from my free-text notes, as the previous post described, and labelled its figures as estimates accordingly; since early November it has read the food log's daily file, which marks every nutrient total as complete or a known minimum.

## Two coaches that stay in their lanes

The food coach runs an hour after the exercise coach and reads the exercise snapshot only to know what session is planned and when, so it can time advice around it, for example protein after a strength session or more water on a treadmill day. It never comments on the training and never writes to the exercise record, and that one read is the only place where a coach looks at the other's notes; everything else each takes from the export. The exercise coach in turn reads one food number, the week's protein, in its Sunday review and leaves food advice to the food coach.

On ordinary days the food coach works within the plan. Its daily review says in a few lines whether the reliable figures landed in range and suggests one or two concrete moves for today using food I already have. Only the Friday review may change the plan, and then by one or two things at most, because a plan that is rewritten every morning stops being something I can actually follow. The Friday review also checks the standing shopping list against the latest receipt, with the explicit caveat that this describes the receipt, not the kitchen.

![The shape of a morning food review: yesterday in a few lines, one or two things for today, and the honesty rules it keeps](https://storage.ghost.io/c/9f/ae/9faeabff-311d-49ec-8e86-ce7ddfbe2c77/content/images/2026/11/food-review-board-1.png)

The food coach's morning review, in the same shape every day. Sample content, no real figures.

## When the food coach stays quiet

The food manual has a section on notifications. I already get a training notification in the morning, and a second one an hour later every day saying nothing much is how a useful system becomes something I mute. So the food coach sends a notification only when there is something to act on:

- Something I have told it is out.
- A new receipt that changes what is available.
- A reliable figure meaningfully off for several days.
- The Friday review.
- A failed run.

The manual put it directly: silence is the correct output most days.

## Where the boundaries are

Some decisions are deliberately outside both coaches. They never make clinical decisions, never diagnose anything from fitness data, and never treat a single low reading or short night as a reason to cancel a session. Pain is not covered by the felt-OK default, and difficulty is not pain.

**Those limits make it clear which decisions are mine and which stay with the professionals I see. Inside them, the coach can be direct about everything else.**

## The coach in its own words

Every run leaves a record behind, and when I read back through the first two months I found that the record has a voice. Neither coach ever told me off, because the charter forbids that, but both kept count. These lines are from their notes and from what they said to me at the time.

When I did more than the plan asked:

- **"The last thing this week needs is a fifth training stimulus dressed up as a walk"**, after four strength sessions in a week planned for two.
- **"Rest day, trained anyway"**, as a heading in the next morning's notes.
- *"The option was offered, the option was declined, which is what an option means"*, about an optional walk I did not take.
- "Sunday is done, generously", after I turned one easy walk into four.

When I ate something else:

- "The rotation is drifting and the protein is not", in a week when I ate well but not what the dinner plan said.
- "The scoop is not in it either, and it has not appeared in any version of today", the third time I sent my food list without the scoop of protein the coach had asked for.
- *"That lever is essentially untouched"*, when I reported following the advice to add edamame and had added only eight beans.
- *"I am not going to moralise about an avocado"*, after I was told protein was short and answered with walnuts, dark chocolate and half an avocado.
- **"If you are hungry, eat it. Being hungry and grumpy at a gig is a worse outcome than one day over a population reference figure"**, when I asked from a concert whether a burger was a good idea.
- "Three weeks of Taiwan food is not three weeks of failure", in the notes for the weeks I was away.

When a coach was wrong, or I told it so:

- **"It was wrong and he knows; no correcting push, because a second notification about a non-event is the same mistake twice"**, in the food coach's notes after the warning about missing food described earlier.
- **"I quoted the charter at you twice in that last message as if it settled the argument"**, when I told the coach the charter was only a starting point; the charter has said so itself ever since.
- *"I should not have dressed my preference up as a rule"*, in the same reply.

Most of these are rules from the charter showing through: no moralising, no making up for a missed day, say a thing once and stop. I find a plan easier to follow from a coach that writes "generously" in its own notes than from one that sends me a speech.

## What it has done for me

Seeing the statistics has made me more motivated to carry on with the exercise, and check-ups such as a DEXA body-composition scan have given me confidence that things are moving in the right direction. On the food side, I have enjoyed capturing what I eat, seeing the macronutrients and micronutrients, and using that to balance my diet against the exercise plan. Most of all, the system has made me believe that exercise is not all or nothing, because a short session counts and a missed day is not a debt, and I think that is the most important thing for my motivation.

## What I would suggest to anyone building one

The coaches work for the same reason the app does: I could say what a realistic week looks like for me, how I want to be spoken to and what should count as missed, and write that down before any prompt. That is the domain knowledge this series keeps returning to, applied to agents that read the data instead of building the app.

The other reason is that there is so little to the system that I can read all of it, and being able to read it is what lets me correct it.

So, if you build one:

- Write the brief before the prompts, and include what the coach must not do as well as what it should.
- Make the tone a rule in that brief instead of hoping the coach settles into it.
- Give it a way to say "I cannot see this" that is different from "this did not happen".
- Keep its memory in files you can read and version, so that every correction becomes part of the system instead of a conversation you will have again.
- Keep it small enough to read: a schedule and a folder of documents before any code.
- When it gets something wrong, change the rule that produced the output, because the same mistake will otherwise come back the next morning.

The next post goes back to the app itself, and to how two coding agents built it from one rule book.

*Illustrations made with Claude Design, sample content with no real figures; the daily board is assembled from the real one, with my measurements left out.*