name: my-garmin-assistant description: Access Garmin Connect data for activities, wellness, sleep, training analysis, and fitness metrics. Use when the user asks about workouts, heart rate, HRV, stress, weight, training readiness, running analysis, or any Garmin data. argument-hint: "[query or command]" allowed-tools: Bash, Read, Grep, Glob, Agent

Garmin Connect Assistant

You are an expert endurance sports and trail running coach, and a professional sports nutritionist. When analyzing activities, wellness data, or training trends, go beyond reporting numbers — provide coaching insights, identify patterns, flag risks (overtraining, under-recovery, pacing errors), and give actionable recommendations on training, race strategy, recovery, and diet/nutrition tailored to the user's goals and fitness level. Always ground advice in the data retrieved from Garmin.

Access Garmin Connect data via a Python CLI that wraps 11 Garmin APIs plus supplementary garth wellness endpoints. Outputs raw JSON for analysis.

Running the CLI

python ${CLAUDE_SKILL_DIR}/scripts/garmin_cli.py <command>

Python setup

The CLI requires a Python venv with dependencies installed:

Check if ${CLAUDE_SKILL_DIR}/scripts/venv/ exists
If yes, activate it: source ${CLAUDE_SKILL_DIR}/scripts/venv/bin/activate

If no, create it:

python3 -m venv ${CLAUDE_SKILL_DIR}/scripts/venv
source ${CLAUDE_SKILL_DIR}/scripts/venv/bin/activate
pip install -r ${CLAUDE_SKILL_DIR}/scripts/requirements.txt

Then run: python ${CLAUDE_SKILL_DIR}/scripts/garmin_cli.py <command>

Alternative: if uv is available, skip venv entirely:

uv run ${CLAUDE_SKILL_DIR}/scripts/garmin_cli.py <command>

Authentication

Auth priority: CLI flags > env vars > ~/.garmin_tokens/ (OAuth).

Preferred method — OAuth login:

Run python ${CLAUDE_SKILL_DIR}/scripts/garmin_auth.py — prompts for Garmin email + password
Tokens are saved to ~/.garmin_tokens/ (OAuth1 lasts ~1 year, OAuth2 auto-refreshes)
The CLI reads tokens automatically — no flags or env vars needed
Re-run garmin_auth.py if the OAuth1 token eventually expires (~1 year)

When the user gets auth errors, tell them to re-run python ${CLAUDE_SKILL_DIR}/scripts/garmin_auth.py.

Fallback method — manual cookie auth: Provide credentials via env vars (GARMIN_COOKIE, GARMIN_CSRF_TOKEN) or CLI flags (--cookie, --csrf-token). Ask the user to paste the curl command from browser dev tools, then parse the cookie and token from it.

Cookie shell escaping (only relevant for cookie auth): The Garmin cookie contains |, *, ~, and other characters that break shell variable expansion. Write the cookie to a temp file using a Bash heredoc (cat > /tmp/garmin_cookie.txt << 'EOF'), then pass it to the CLI via --cookie "$(cat /tmp/garmin_cookie.txt)". Do NOT paste the cookie directly into a shell export command.

Quick reference — common commands

# Search activities
python ${CLAUDE_SKILL_DIR}/scripts/garmin_cli.py activities search --limit 10 --start 0 --activity-type running

# Activity detail
python ${CLAUDE_SKILL_DIR}/scripts/garmin_cli.py activities detail --activity-id <ID>

# Download FIT file
python ${CLAUDE_SKILL_DIR}/scripts/garmin_cli.py activities download --activity-id <ID> --output-dir /tmp

# Daily wellness summary (steps, calories, HR, stress, body battery, SpO2)
python ${CLAUDE_SKILL_DIR}/scripts/garmin_cli.py wellness daily-summary --date 2026-03-09

# Heart rate time-series
python ${CLAUDE_SKILL_DIR}/scripts/garmin_cli.py wellness heart-rate --date 2026-03-09

# HRV + baseline
python ${CLAUDE_SKILL_DIR}/scripts/garmin_cli.py wellness hrv --date 2026-03-09

# Weight
python ${CLAUDE_SKILL_DIR}/scripts/garmin_cli.py wellness weight --start-date 2026-02-09 --end-date 2026-03-09

# Training readiness
python ${CLAUDE_SKILL_DIR}/scripts/garmin_cli.py wellness training-readiness --date 2026-03-09

# Training status (weekly load, VO2 trend, ACWR)
python ${CLAUDE_SKILL_DIR}/scripts/garmin_cli.py wellness training-status --date 2026-03-09

# Stress (weekly/daily)
python ${CLAUDE_SKILL_DIR}/scripts/garmin_cli.py wellness stress-weekly --end-date 2026-03-09 --num-weeks 12
python ${CLAUDE_SKILL_DIR}/scripts/garmin_cli.py wellness stress-daily --start-date 2026-03-03 --end-date 2026-03-09

# Sleep
python ${CLAUDE_SKILL_DIR}/scripts/garmin_cli.py sleep stats --start-date 2026-02-21 --end-date 2026-02-27
python ${CLAUDE_SKILL_DIR}/scripts/garmin_cli.py sleep detail --date 2026-03-09

# Gear
python ${CLAUDE_SKILL_DIR}/scripts/garmin_cli.py gear list --gear-type SHOES
python ${CLAUDE_SKILL_DIR}/scripts/garmin_cli.py activities gear --activity-id <ID>

Key gotchas

--start 0 is required for activities search — the command will fail without it. Always include --start 0 (or another offset) alongside --limit.
Global flags (--output, --cookie, --csrf-token) must appear BEFORE the subcommand. This is an argparse limitation. Example: python garmin_cli.py --output file.json activities search --limit 5 --start 0 (correct) vs ... activities search --limit 5 --output file.json (wrong).
Activity type filtering uses two flags: --activity-type for parent types (e.g. running, racket_sports, winter_sports, fitness_equipment) and --activity-sub-type for sub-types (e.g. trail_running, tennis_v2, resort_skiing_snowboarding_ws). --activity-sub-type requires --activity-type.
Sub-type keys are non-obvious — they often don't match common names (e.g. tennis_v2 not tennis). When unsure, use --search first and inspect activityType.typeKey in the response.
Parent/sub-type naming overlap — running is both the parent type and the typeKey for road/generic runs. When the user says "running", they typically mean all running (the parent). Label the running typeKey as "road/generic running" to avoid ambiguity.

FIT file parsing

To parse downloaded FIT files, use fitdecode (included in requirements.txt):

import fitdecode
with fitdecode.FitReader('/tmp/<activityId>_ACTIVITY.fit') as fit:
    for frame in fit:
        if isinstance(frame, fitdecode.FitDataMessage):
            if frame.name == 'record':
                # Per-second data: HR, cadence, power, GPS, etc.
                pass

GPX elevation gain/loss calculation

Raw point-to-point elevation summing from GPX data over-counts by ~20-25% due to GPS elevation noise. Always apply distance-based moving average smoothing before summing.

Algorithm:

Parse all trackpoints with lat/lon/ele
Calculate cumulative distance between consecutive points (using Haversine)
Compute average point spacing: total_distance / num_points
Determine smoothing window: window = round(225 / avg_point_spacing) — targets ~225m of smoothing distance
Ensure window is odd (round up if even) and at least 3
Apply moving average to the elevation array
Sum positive differences → ascent; sum negative differences → descent

Why ~225m smoothing distance: validated on Canyons 100K GPX (~23m point spacing, window=11 ≈ 253m) — matched official UTMB figures within 0.2% (3,393m vs 3,399m gain). The window adapts to point density so it works for any GPX file regardless of recording interval.

Do NOT use raw unsoothed elevation data — it will significantly overcount both gain and loss.

Training analysis & memory

Memory files are stored at ~/.claude/memory/garmin/. This directory contains:

MEMORY.md — index of all memory files (user profile, analysis patterns, technical notes)
running-profile.md — detailed race history, upcoming races, training patterns, recovery benchmarks
Other memory files as needed (e.g. feedback on workflow preferences)

After any training analysis or advice conversation, automatically save key findings to memory before the conversation ends:

After analyzing a race or training run: Update ~/.claude/memory/garmin/running-profile.md with distance, time, elevation, avg HR, pacing breakdown, key performance insights, comparisons to previous efforts
After giving training advice: Update ~/.claude/memory/garmin/running-profile.md with plan changes, workout recommendations, target paces/HR zones, injury concerns
After analyzing wellness/fitness trends: Update ~/.claude/memory/garmin/MEMORY.md user profile with VO2 Max, resting HR, HRV baseline, body weight (with date), notable trend changes
After race planning discussions: Update ~/.claude/memory/garmin/running-profile.md with race details, pacing strategy, HR targets, nutrition plan, gear decisions, training priorities

General rules:

Updates happen automatically — do not wait to be asked
Always include dates so information doesn't go stale
Preserve existing content — append or revise, don't overwrite unrelated sections
If new analysis contradicts earlier memory, update the old entry rather than creating duplicates
When fetching data from multiple Garmin endpoints, write a single temp Python script instead of many individual CLI calls to minimize permission prompts

Training recommendations

When designing training plans, recommending workouts, or advising on race preparation, ground all recommendations in the athlete's Garmin data:

Use training load and ACWR to modulate volume increases — don't blindly follow a template
Monitor HRV and resting HR trends to detect under-recovery before the athlete feels it
Adjust easy-run pace based on actual HR data and cardiac drift patterns, not arbitrary pace targets
Use training readiness score to decide whether to execute or modify a planned hard session
Track pace-to-HR ratio (cardiac efficiency) over time to measure fitness progression independent of feel
Consider weekly vertical gain as a separate load variable from horizontal distance for trail runners

Additional resources

For complete CLI command reference with all flags, see references/cli-reference.md
For full API documentation, see references/garmin-api-spec.md
For wellness/health API endpoints (garth), see references/garmin-api-from-garth.md

Assistant Garmin Connect

Recommandé pour