Why Mondays stopped negotiating with me

The question

Why did another watch face and another planning app still leave me arguing with myself every Sunday, and what would it take for the plan to show up already decided for the next week?

Why I built it

I didn't need more workouts. I needed a week that updates when goals, progress, and fatigue change.

The failure looked the same each time. A race goal on paper (right now a September 5K, PB 22:34), then a week that stacked hard days because ego felt fresh, or went soft because I "felt" tired with no receipt. Strength and running collided: crush legs on Tuesday, then Wednesday quality feels like wet cement. Chat-style planning made it worse in a friendly way. Every Sunday I re-explained constraints the model had already "agreed" to last week, then lost the thread in scrollback.

So I built a coach for one athlete: me. Not a product, a closed loop. Garmin feeds Intervals.icu. The repo holds profile, principles, plans, and lift logs. Cursor Automations run the Sunday job, a daily sync, and a chat router. Slack #fitness is the phone UI. I train, I log: what I did, and sometimes I send intent: when life changes the path.

If the week is already on my phone with a "why," I stop spending morning willpower on planning theatre.

This is personal-ops engineering (git, automations, deterministic rules), not a training diary. Repo: github.com/AlexTouvras/fitness-coach. Diagrams and data ownership: /architecture/fitness-coach.

The science it actually encodes

Nothing here invents a physiology paper. The research/ pack is blunt about the lineage: Seiler-style easy/hard distribution (most volume stays conversational), recreational 5K shape from Daniels' Running Formula (one quality focus, one longer easy run, taper in the last week), and concurrent-training spacing so strength does not sabotage the key run. The non-negotiables live in research/principles.md and get cited in the weekly plan: overload with recovery, easy days stay easy, readiness beats ego, one primary focus per block, honest logs, cite why.

Running form comes from Intervals' Fitness chart: chronic load (CTL), acute load (ATL), form as TSB ≈ CTL − ATL. Zones are wired (fresh, productive, loading, high fatigue). If form is high_fatigue or TSB ≤ −20, AthleteState sets cut_quality and Wednesday stops pretending to be a hero session.

Lifts rarely land as WeightTraining in my Intervals pull, so strength does not get a free pass. The coach builds a Foster-style session-RPE load (RPE × minutes / 10) and a local Banister EWMA (strength CTL ~28d, ATL ~7d). High recent lift RPE or strength high-fatigue flips ease_lower_lift, so Tuesday softens before Wednesday. That is concurrent training in one rule: protect the session that pays the race.

Phases are calendar-dumb on purpose. More than 21 days to the event → build. Inside 21 → sharpen. Inside 7 → taper (keep Mon/Wed/Fri anchors, cut volume, keep a little sharpness). The race does not get a speech. It gets a different template.

Slack is not a diary

The useful upgrade since the first version is routing. Prefixes decide what git is allowed to change.

• log: — session writeback only (RPE debt, strength form). Example: log: lift | squat 3x5 @ RPE 7
• intent: — path tweak for this block: writes data/course.json, regenerates the week. Example: intent: 5K is a fitness test, not chasing sub-22; lean into strength
• goal: — mesocycle swap; proposal until I approve in a follow-up. Example: goal: after 5K switch primary to hypertrophy 8 weeks

intent: was the missing layer. Life does not always need a new block_id. Sometimes I need "keep sessions short" or "simplify Tue/Thu to one lift" without throwing away the 5K pack. coach-intent merges that into data/course.json, patches profile flags, and regenerates data/plans/YYYY-Www.json without bumping the lift week or pushing Intervals unless the plan actually changed.

log: never rewrites the program. Otherwise every tired Tuesday note becomes a program edit.

How a week gets generated

Sunday Automation runs coach-week, posts Slack, then coach-publish (commit and push data/ to master). Slack alone is not done. I learned that when a cloud agent posted a beautiful week that never landed in git.

The mechanical path is short. Sync Intervals activities, wellness, and calendar; rebuild AthleteState from that summary plus recent Slack logs and data/course.json. Profile + block_id (today: 5k-2026-09) lock run days Mon/Wed/Fri and lift days Tue/Thu. Readiness flags edit the skeleton: cut_quality turns Wednesday into easy + strides; ease_lower_lift annotates Tuesday. If Intervals already has a run on a locked day, reuse it. Garmin-native plans are not readable through the API, so when the calendar is empty the coach generates and pushes Intervals → Garmin. Strength comes from coach-owned templates (data/library/coach_wN_*.json) plus an HTML page I can print; form-check clips join from a Heimdall snapshot at render time.

The Slack post is the phone UI: kickoff (quote, one motivate clip, playlists), Mon–Sun sessions, lift bullets, and Why this week from the signals list. That last line is the point. If form says cut quality, Wednesday becomes easier and the post says so.

Daily 07:00 sync refreshes summary + AthleteState without rewriting the week. Midweek can re-check (coach-midweek). Sunday still owns the story. The step-by-step loop lives on /architecture/fitness-coach.

What this does to an ordinary week

Sunday morning a message lands in #fitness while coffee is still optional. Kickoff clip, quote, playlists, then Mon–Sun on the phone. No Notion tab. No "what should I do this week" thread with myself.

Gym days: open the lift bullets or the HTML export. Squats, RDLs, split squats, already sized for maintenance so Wednesday is not a crime scene. If last week's log: lift | … RPE 9 showed up, Tuesday may already be softer. That is the system paying me back for honesty.

Run days: the watch already has the session when the push worked. Easy stays easy. Quality is the one hard thing. If life shifts (sleep deprivation, travel, a race that became a fitness test), I send intent: once and the week rewrites with a receipt, not a vibes negotiation in my head.

Decision tax drops. Concurrent training stops being two sports fighting in the same body. After the 5K I can archive the block, flip lifting toward hypertrophy, and keep the same Automations: durable loop, replaceable goal pack.

What still hurts: skip lift logs and strength form is fog. Slack is uglier than a native training app. Cloud agents invent stale weeks unless they hard-reset to master before they speak. Sharp pain still means stop and see a clinician; the coach will not play doctor.

Research the pack draws on

Named methods only, not a literature review. Canonical list: research/principles.md. Block packs sit under research/blocks/.

• Seiler (2010) — most endurance volume stays easy; hard days stay hard. I borrow the distribution, not a polarized-vs-pyramidal argument.
• Foster session-RPE (2001) — internal load = RPE × duration; used when Intervals has no lift activities.
• Banister / PMC load models — EWMA fitness/fatigue on that load; same family as Intervals CTL/ATL, with separate τ for strength in form_metrics.py.
• Intervals.icu Fitness chart — running CTL/ATL/TSB; cut_quality at TSB ≤ −20.
• Schumann et al. (2022) — concurrent aerobic + strength; spacing beats same-session stacking.
• Daniels' Running Formula — recreational 5K phase shape. Book, not a URL; pace anchors in research/5k-pace-anchors.md.

Takeaway

I built this so the week arrives already argued (CTL/TSB, RPE debt, block research, and an explicit intent: when life moves), not by Monday motivation. The science is old and specific. The generation path is mechanical. The payoff is ordinary: less negotiation, more showing up, and a written reason when the plan tells ego to sit down.