convez.ai
SkillAutomaçãoAvançado
fengjh97

Treinador de Perda de Gordura Baseado em Evidências

fengjh97/claude-skill-fat-loss-coach

Coach de perda de gordura objetivo com red lines baseadas em ciência

Ver no GitHub
InicialPluginScore · de 0 a 100

Skill Claude que integra dados do Garmin Forerunner 570 para fornecer decisões diárias de treinamento em 4 tiers (REST/LIGHT/MODERATE/GREEN_LIGHT) e revisões semanais de 8 KPIs, com base em 206 citações científicas. Implementa um árbitro anti-overtraining com red lines determinísticas que recusam negociações quando o protocolo detecta risco de overreach.

Instalação

$npx convez-ai fengjh97/claude-skill-fat-loss-coach

1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.

Métricas

Estrelas
0
Forks
0
Acompanhando
0
Saúde
42%
Último commit
há 6 meses
Versão
—
Issues abertas
0
Contribuidores
1

Casos de uso

  • Decisões diárias de intensidade de treinamento guiadas por HRV e Body Battery
  • Revisão semanal de 8 KPIs com direcionamento de deload/manutenção/progressão
  • Recusa automática de overtraining quando dados indicam risco de REDs (relative energy deficiency)

Veredito

Skill altamente especializada e rigorosa: ideal para atletas com tendência a overtraining que precisam de decisões objetivas amarradas a literatura científica. Exige calibração individual e leitura do SOP antes do uso; implementação Python determinística garante reprodutibilidade, mas o disclaimer médico é crítico.

README

A tradução para o português deste README ainda está na fila.

claude-skill-fat-loss-coach

An evidence-based Claude Code skill that turns a Garmin Forerunner 570 into an objective fat-loss coach + "anti-overreach referee" — every recommendation traces to a numbered citation in a 206-reference, 1731-line scientific SOP.

⚠️ Disclaimer first

  • Not medical advice. This skill encodes a personal protocol calibrated to one user (adult, ~80 kg, 2800 kcal TDEE, 800 kcal deficit, Garmin Forerunner 570). The decision thresholds reference peer-reviewed literature, but the parameters are tuned to one body. If you use it, calibrate to yours.
  • Hard floor is 2000 kcal/day intake. Going lower than this without medical supervision crosses into the REDs / low-energy-availability zone documented in the SOP §7.2. If you weigh less than ~75 kg, the protocol automatically downgrades the deficit. Read research/SOP_AI_Fat_Loss_Coach.md §2.1 before adopting.
  • Six hard-STOP red lines. If the data shows one of them (SOP §7.5), the protocol terminates the cycle. The skill is designed to refuse you when this happens — that's the point.
  • See a doctor for any of the eight medical red flags. Syncope, RHR <40 with symptoms, amenorrhea ≥3 months, BMI <17, etc. The skill defers, it does not diagnose.

What this is

A Claude skill that runs in three modes:

  1. Morning briefing — paste your Garmin data, get one of four training tiers (REST / LIGHT / MODERATE / GREEN_LIGHT) plus a specific prescription. The decision tree is implemented in scripts/compute_decision_tier.py as deterministic Python — Claude never eyeballs thresholds.
  2. Weekly review (Sundays) — eight KPIs from SOP §6.4 computed by scripts/weekly_kpi_calculator.py, color-coded, with an auto-generated deload / maintain / progress directive for next week.
  3. Ad-hoc questions and pushback — if you try to negotiate a tier upward ("I feel fine, can I add a HIIT session?"), Claude cites the triggered rule and refuses using the verbatim templates in references/coach_persona.md. This is the deliberate "anti-overreach referee" behavior.

The skill's core thesis: for a user with overtraining tendencies ("卷"), objective data must outrank subjective feelings. All decisions resolve to a numbered SOP rule (e.g., DAILY_INTENSITY_LIGHT, POST_GREEN_LIGHT_RECOVERY_GATE, MANDATORY_DELOAD_ON_RED), and the rule is non-negotiable.

Quick start

Install as a Claude Code skill

# Clone directly into your skills directory
git clone https://github.com/fengjh97/claude-skill-fat-loss-coach.git \
  ~/.claude/skills/fat-loss-coach

# Verify by listing
ls ~/.claude/skills/fat-loss-coach/
# Should show SKILL.md, references/, prompts/, scripts/, assets/, research/

The skill auto-registers next time you launch claude. It triggers on Chinese phrases (减脂, HRV, 复盘, 加练, 静息心率, Body Battery, 训练决策, 反卷, 红线) and equivalent English terms (fat loss coach, training tier, HRV-guided training, deload, overreach, RED-S).

Try it without Claude (just the decision script)

cd ~/.claude/skills/fat-loss-coach
# Fill in your morning's Garmin data in assets/example_input.json, then:
python3 scripts/compute_decision_tier.py --input assets/example_input.json --pretty

Sample output:

{
  "tier": "LIGHT",
  "duration_min": 45,
  "intensity_cap_pct_max_hr": 70,
  "zone_cap": "Z2",
  "note": "Aerobic base only, conversational pace. No HIIT, no threshold work.",
  "triggered_rules": ["LIGHT::HRV<baseline-0.5SD"],
  "flags": {"sleep_debt_override": false, "infection_pre_flag": false}
}

Architecture

fat-loss-coach/
├── SKILL.md                       # Entry point (~150 lines, triggers + 3 modes)
├── references/                    # Progressive-disclosure deep docs
│   ├── sop_full.md                #   Full 1731-line SOP, 206 citations
│   ├── decision_tree.md           #   §5 — 4-tier daily logic
│   ├── red_lines.md               #   §7 — STOP conditions + medical red flags
│   ├── garmin_metrics.md          #   §3 — 10 metrics, black-box vs validated
│   ├── training_plan.md           #   §4 + §8 + §9 — programming + hike + cycle
│   ├── nutrition.md               #   §2 — 800 kcal deficit science + protein targets
│   ├── progress_tracking.md       #   §6 — 8 weekly KPIs + EMA algorithm
│   └── coach_persona.md           #   §10 — refusal templates (use verbatim)
├── prompts/
│   ├── morning_briefing.md        #   Mode 1 output template (3 examples)
│   └── weekly_review.md           #   Mode 2 output template (10 sections, ~1500 字)
├── scripts/
│   ├── compute_decision_tier.py   #   Deterministic 4-tier tree (source of truth)
│   └── weekly_kpi_calculator.py   #   8 KPIs → next week directive
├── assets/
│   └── example_input.json         #   19-field morning data schema
└── research/                      # Standalone research artifacts (not loaded by Claude)
    ├── SOP_AI_Fat_Loss_Coach.md   #   Full SOP (also in references/sop_full.md)
    └── findings/                  #   6 evidence modules, 250 findings, JSON
        ├── module_a_findings.json #     Fat-loss science (41 findings)
        ├── module_b_findings.json #     Garmin metrics (51 findings)
        ├── module_c_findings.json #     Training programming (48 findings)
        ├── module_d_findings.json #     Decision rules (31 findings)
        ├── module_e_findings.json #     Progress tracking (38 findings)
        └── module_f_findings.json #     Risk signals (41 findings)

The 4-tier decision tree (quick reference)

TierTrigger (any one for REST/LIGHT, all-of for GREEN_LIGHT)Output
RESTHRV < baseline − 1 SD, or RHR ≥ +5 bpm 7d, or skin temp ≥ +0.6°C, or TR < 25, or BB wake < 250 min; optional ≤30 min Z1 walk
LIGHTHRV < baseline − 0.5 SD, or TR 25–49, or BB 25–49, or deep sleep < 45 min, or RPE yesterday ≥8, or sleep debt 7d ≥5h, or infection pre-flag45 min, Z2 cap (≤70% HRmax)
MODERATEHRV within ±0.5 SD AND TR 50–74 AND BB 50–74 AND RPE yesterday <875 min, Z3 cap (≤85% HRmax)
GREEN_LIGHTAll 7 gates pass: HRV ≥ baseline + 0.5 SD, TR ≥75, BB ≥75, RHR ≤ baseline, deep sleep ≥60, skin temp dev <0.3°C, RPE yesterday ≤6120 min, full Z5 (HIIT authorized)

Cross-tier override: If yesterday was GREEN_LIGHT with TSS ≥150, today caps at LIGHT unless triple-confirmed (HRV ≥ baseline AND BB ≥70 AND RPE ≤7). Prevents 2-day HIIT stacking.

The six hard-STOP red lines (SOP §7.5)

Any of these terminates the cycle — the skill refuses to issue daily tiers until resolved:

  1. HRV 7d mean < baseline − 3% × 14 consecutive days with no recovery.
  2. RHR 7d mean ≥ baseline + 5 bpm × 14 consecutive days.
  3. Weight EMA dropping > 1.5% BW/wk × 3 consecutive weeks.
  4. Sleep efficiency drop ≥10% or deep-sleep drop ≥30% × 1 week.
  5. Key compound lift dropping >10% at same RPE × 2 weeks.
  6. POMS Depression-Dejection ↑ / Vigor ↓ × 2 weeks + subjective collapse.

Scientific basis

The full SOP at research/SOP_AI_Fat_Loss_Coach.md cites 206 numbered references, including:

  • Fat-loss safety bounds: Helms et al. 2014 (J Int Soc Sports Nutr), Aragon/ISSN 2017 position stand, Garthe et al. 2011 (slow vs fast loss in athletes), Trexler et al. 2014 (metabolic adaptation review)
  • Protein during deficit: Longland et al. 2016 (Am J Clin Nutr, 2.4 g/kg/day RCT), Phillips & Van Loon 2011 (J Sports Sci), Schoenfeld & Aragon 2018 per-meal protein
  • HRV-guided training: Plews et al. 2013–2017 series (Sports Med), Buchheit 2014 (Frontiers), Vesterinen et al. 2016, Düking et al. 2021 meta-analysis
  • Overtraining & RED-S: Meeusen et al. 2013 (Eur J Sport Sci ECSS/ACSM consensus), Mountjoy et al. 2023 IOC RED-S consensus (Br J Sports Med), Bellenger et al. 2016 RHR meta
  • Sleep & fat loss: Nedeltcheva et al. 2010 (Ann Intern Med — the 5.5h vs 8.5h paper), Wang et al. 2018
  • Infection pre-detection: Mason et al. 2022 TemPredict (Sci Rep — 2.75-day-ahead Oura signal), Grzesiak et al. 2021 multimodal early warning
  • Training load: Gabbett 2016 ACWR sweet spot 0.8–1.3 (Br J Sports Med), Williams et al. 2017 EWMA, Coggan TSS framework

The deep-research process (6 parallel evidence-gathering agents → synthesis) is documented in research/findings/.

Customizing for your own setup

The skill works for you if you:

  1. Have a Garmin watch that exposes HRV Status, Body Battery, Training Readiness, RHR, Skin Temperature (FR570/955/965/Fenix 7+, Epix Pro, Venu 3, etc.).
  2. Adjust two numbers in your morning input: replace the 800 kcal deficit / 2000 kcal intake assumptions with yours. The 75 kg auto-downgrade threshold also needs adjusting if your weight is materially different.
  3. Establish baseline first. Wear the watch 21 consecutive nights before trusting any decision — HRV baseline needs the time per Plews 2013. The script forces LIGHT until baseline is built.

The 4-tier decision tree's thresholds are percent-of-personal-baseline, not absolute — so they generalize across body sizes. Same for the red lines.

What does NOT generalize:

  • The 800 kcal deficit recommendation (it's calibrated to ~2800 TDEE — adjust for your TDEE)
  • The protein target table (it's in g/day for specific weights — recompute for yours)
  • The sample weekly training plan (it's 10 h/wk — scale to your training availability)

Modes of contribution

Issues and discussions welcome on:

  • Threshold tuning based on new literature.
  • Garmin algorithm changes (Firstbeat occasionally updates, breaking assumptions).
  • Adapting to other wearables (Whoop, Oura, Polar Vantage). The decision logic is wearable-agnostic if the metrics are computed equivalently.
  • Bug fixes in the Python scripts.

What this project intentionally won't accept:

  • Recommendations to relax the red lines.
  • Personalized advice requests (this isn't a coaching service).
  • Anti-science alternative medicine.

License

MIT — see LICENSE.

Credits

  • Research synthesized by Claude (Opus 4.7) over 6 parallel agents, 250 findings, ~30 minutes wall-clock.
  • Built on top of the Claude Code skills progressive-disclosure architecture.
  • The 1731-line SOP and 206 numbered citations are the substance — the code is the wrapper.
Ver o original no GitHub

Bom para

Relacionados