Skip to content

npapatheodorou/personal-trainer-skill

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

1 Commit
 
 
 
 
 
 
 
 

Repository files navigation

Coach — a portable Personal Trainer + Nutritionist for AI agents

A single, portable AI coaching system that turns your agent into a complete, evidence-based personal trainer and nutritionist for one athlete. It programmes your training with RPE-based progressive overload, logs every session, tracks progress and busts plateaus, calculates your calorie and macro targets, builds practical meal plans, and remembers all of it between conversations by keeping your profile and logs in files.

It is packaged two ways from the same knowledge base so it runs anywhere:

Every recommendation is grounded in the strength-and-conditioning and nutrition literature — the Coach is a scientific lifter and nutritionist, not bro-science.


What's in this folder

.
├── README.md                       ← you are here
├── AGENTS.md                       ← cross-agent instructions (Codex, opencode, Cursor, Gemini CLI…)
├── .opencode/
│   └── agent/
│       └── coach.md                ← dedicated opencode agent definition
└── personal-trainer/               ← the shared knowledge base + Claude skill
    ├── SKILL.md                    ← Claude Code / claude.ai entry point
    ├── references/
    │   ├── evidence.md             ← scientific evidence base + citations + confidence grading
    │   ├── training.md             ← programming, logging, progression, periodisation
    │   ├── nutrition.md            ← TDEE/macros, meal planning, supplements, stalls
    │   └── prompts.md              ← ready-made prompt stack
    └── assets/                     ← blank fill-in templates (copied into coach-data/)
        ├── athlete-profile.md
        ├── exercise-library.md
        ├── lift-log.md
        ├── programme-template.md
        └── nutrition.md

How it's designed

  • AGENTS.md and personal-trainer/SKILL.md are two entry points into the same brain. Each defines the Coach's identity, the persistent-data rules, and the onboarding flow, then routes to the detailed reference files. AGENTS.md is the universal one (most agents load it automatically); SKILL.md adds Claude's skill frontmatter so Claude auto-activates it.
  • personal-trainer/references/ holds the detailed coaching logic and the science. The agent reads only the file relevant to the request (training vs nutrition vs evidence vs prompts) — progressive disclosure keeps the always-loaded context small while deep knowledge stays available on demand.
  • personal-trainer/assets/ holds blank templates. On first use the agent copies them into a working coach-data/ folder, where your actual data lives and grows.

Persistence — the important bit

What makes this a coach and not a chatbot is memory. Whenever it runs, it looks for a coach-data/ folder in the working directory and reads:

File Holds
coach-data/athlete-profile.md Stats, goals, schedule, equipment, injuries, dietary prefs
coach-data/exercise-library.md Which exercises are YES / SUB / NO
coach-data/lift-log.md Working weights + every logged session (append-only)
coach-data/programme.md The active programme + block/week tracking
coach-data/nutrition.md Calculated targets, macros, meal-plan history

After every session log, target calculation, or programme change it writes the update back — so the next conversation picks up where you left off. The lift log is append-only: old entries are never deleted, because they are the source of truth for progress.


Installation & use

Any agent with file read/write gets full persistence. Pick your tool.

Codex (OpenAI Codex CLI)

AGENTS.md is read automatically from the working directory.

  1. Put this folder where you want your coach-data/ to live (or copy AGENTS.md + personal-trainer/ into an existing project). Codex also merges ~/.codex/AGENTS.md if you want the Coach available globally.
  2. Run codex in that folder and say: "Be my coach — set me up."
  3. Codex reads AGENTS.md, runs onboarding, creates coach-data/, and reads/writes your files directly from then on.

opencode

Two ways:

  • Automatic rules: opencode reads AGENTS.md from the project root — just open opencode in this folder and talk to it.
  • Dedicated agent: .opencode/agent/coach.md defines a primary agent named coach (read/write/edit/bash enabled). Switch to it with the agent picker (Tab) or @coach. To make it global, copy that file to ~/.config/opencode/agent/coach.md.

Cursor / Gemini CLI / Jules / other AGENTS.md agents

Open the folder; these tools pick up AGENTS.md as project instructions. Anything with file access maintains coach-data/ automatically. Then use natural language or the prompt stack in personal-trainer/references/prompts.md.

Claude Code

  1. Copy personal-trainer/ into a skills directory:
    • Personal (all projects): ~/.claude/skills/personal-trainer/ (Windows: C:\Users\<you>\.claude\skills\personal-trainer\)
    • Project-scoped: <your-project>/.claude/skills/personal-trainer/
  2. Open Claude Code in the folder you want coach-data/ to live in.
  3. Say "Be my coach. Let's set up my profile." — Claude detects the skill, onboards you, and creates coach-data/. Confirm install with /skills.

(Claude Code will also honour the root AGENTS.md if you prefer not to install the skill.)

claude.ai Project (no install, no file writes)

  1. Projects → New Project, name it (e.g. "Coach").
  2. Paste the body of personal-trainer/SKILL.md (skip the YAML frontmatter) into the project Instructions, then append the references/ files so the full logic is available (Projects can't lazy-load).
  3. Upload your filled-in assets/ templates into the project Files.
  4. ⚠️ Projects can't write back to files — after each session, ask for the updated lift-log.md block and paste it into your uploaded file yourself.

Any LLM API / Agent SDK

Load AGENTS.md (or SKILL.md) + the reference files as the system prompt and give the model a filesystem tool scoped to a coach-data/ directory. The data rules then drive persistence exactly as above.


A typical first session

  1. You: "Be my coach — set me up."
  2. Coach: creates coach-data/, interviews you for the essentials, walks the exercise library, saves everything, confirms what it knows and flags gaps.
  3. You: "Build me a 4-day programme" → full Day 1 + structure for Days 2–4, saved to programme.md.
  4. You: "Now calculate my fat-loss calories and protein" → shows the Mifflin-St Jeor working, sets targets, saves to nutrition.md.
  5. Day to day: paste your sessions to log them, ask for the next workout, request meal plans, swap exercises, run a deload, or troubleshoot a stall.

Evidence-based by design

The Coach is built as a scientific lifter and nutritionist — every protocol, number, and claim is grounded in the strength-and-conditioning and nutrition literature, not bro-science or fads. It carries its own evidence library in personal-trainer/references/evidence.md, which:

  • maps each coaching rule back to its research basis (e.g. volume as the primary hypertrophy driver, proximity-to-failure governing the stimulus, energy balance for fat loss, protein at 1.6–2.2 g/kg);
  • carries a short-form reference list (Morton et al. 2018, Schoenfeld et al. 2017, ACSM 2009, Kreider et al. 2017, ISSN position stands, etc.) the Coach cites from;
  • tells the Coach to grade its confidence (strong / moderate / emerging), be honest when evidence is thin, never fabricate a citation, and reject pseudoscience (detoxes, fat burners, spot reduction, "metabolic damage" myths) outright.

Ask it "why?" or "what's the source?" on any recommendation and it pulls the rationale and citation rather than hand-waving.

Coaching principles baked in

Training — RPE-based loading (7–8 hypertrophy, 8–9 strength), never percentages off a 1RM; progressive overload after 2 consecutive top-of-range sessions at target RPE (+2.5 kg upper compounds, +5 kg lower compounds, reps-before-weight on accessories); compounds first, accessories for weak points; Week 1 is calibration; deload every 4–6 weeks.

Nutrition — TDEE via Mifflin-St Jeor × activity factor, 300–500 kcal deficit for steady fat loss; macro priority protein 0.8–1 g/lb (≈1.6–2.2 g/kg, non-negotiable), fat ≥0.35 g/lb, carbs fill the rest; whole-food, practical meal plans that respect every restriction; evidence-backed supplements only.

Style — direct, concise, real numbers, tables for programmes and meal plans, metric units, no moralising.


Customising it

  • Edit AGENTS.md / personal-trainer/SKILL.md to change the coach's personality, units (e.g. switch to lb/imperial), or default split. Keep the two in sync since they share a brain.
  • Edit the references/ files to adjust programming or nutrition rules (progression increments, deficit size, periodisation, evidence).
  • Edit the assets/ templates to change what data the coach collects.

Disclaimer

The Coach is a tool, not a doctor or a registered dietitian. Train sensibly, and if you have a diagnosed health condition or complex dietary needs, consult a qualified professional. The system is built to flag this and defer on medical questions.

Knowledge distilled from guides by Hawks (@Hawks0x), grounded in the literature cited in personal-trainer/references/evidence.md.

About

A portable, evidence-based AI personal trainer & nutritionist. Turns any file-capable agent (Claude Code, Codex, Cursor, opencode, Gemini CLI…) into a coach that programmes your training, logs sessions, calculates macros, builds meal plans — and remembers it all between conversations.

Topics

Resources

Stars

Watchers

Forks

Contributors