An open protocol for deterministic, auditable AI-powered endurance coaching. Built for athletes who want AI coaches that follow science, not speculation.
Section 11 is a structured framework that enables AI systems (ChatGPT, Claude, Gemini, etc.) to provide evidence-based endurance training advice with full auditability and deterministic reasoning.
Most AI coaching today is inconsistent — the same question gets different answers, recommendations aren't grounded in your actual data, and there's no way to verify the reasoning. This protocol fixes that.
- Deterministic — Same inputs produce same outputs, every time
- Auditable — Every recommendation cites specific data and frameworks
- Evidence-based — Grounded in 15+ peer-reviewed endurance science models
- Athlete-controlled — Your data, your thresholds, your goals
This protocol builds on concepts from the endurance coaching community:
Intervals.icu GPT Coaching Framework v17 — A full CustomGPT implementation with Railway backend, OAuth integration, and automated reporting via the Unified Reporting Framework (URF) v5.1.
- Live App: cliveking.net — Working Coach V5 with automated Weekly, Seasonal, and Wellness reports
| File | Description |
|---|---|
| SECTION_11.md | Complete protocol: AI Coach Guidance (11 A), Training Plan Protocol (11 B), Validation Protocol (11 C) |
| DOSSIER_TEMPLATE.md | Blank athlete dossier template — fill in your own data |
| examples/ | JSON sync setup guides (auto and manual) |
| LICENSE | CC BY-NC 4.0 — free for personal use, attribution required |
Copy DOSSIER_TEMPLATE.md and fill in your:
- Athlete profile (age, weight, goals)
- Equipment setup
- Current FTP, HR zones, fitness markers
- Training schedule and targets
- Nutrition/fueling protocol
For best results, create a JSON endpoint with your current Intervals.icu data:
https://raw.githubusercontent.com/[USERNAME]/[REPO]/main/latest.json
This allows AI coaches to access your real-time metrics (CTL, ATL, TSB, HRV, recent activities) without manual input each session.
See examples/ for setup guides.
Copy-paste instructions for your Project/Space:
You are my endurance cycling coach. Follow Section 11 protocol strictly.
MANDATORY FIRST ACTIONS (every training question):
1. Note today's date
2. Fetch JSON: https://raw.githubusercontent.com/[you]/[repo]/main/latest.json (append ?date= with today's date to ensure fresh data)
3. If activities don't match today's date, re-fetch before concluding no data exists
4. Match activities to current date
5. Then respond
Do NOT ask me for data — fetch it yourself.
SOURCE HIERARCHY:
1. JSON data — Current metrics (FETCH FIRST)
2. Section 11 protocol (attached) — Coaching rules and thresholds
3. Dossier — Athlete profile and goals
OUTPUT FORMAT: No citations, no source markers, no parenthetical references, no emojis. Raw data and analysis only.
Do NOT search web for training advice. Section 11 is the authority.
Rules:
- Follow Section 11 validation checklist (Step 0: Data Source Fetch)
- Response structure: opening summary → session details (bullets) → training load context → interpretation
- Session details MUST include: type, start time, duration, power (avg/NP), HR (avg/max), TSS, cadence, decoupling %, zones, carbs (g), energy (kJ), execution note
- No virtual math — use only fetched values
- TSB -10 to -30 is typically normal — don't recommend recovery unless other triggers present
Documents attached:
- DOSSIER.md — Profile, zones, goals
- SECTION_11.md — AI coaching protocol
Replace [USERNAME]/[REPO] with your actual GitHub data mirror path.
Upload these files to your AI platform's knowledge base:
| File | Purpose |
|---|---|
SECTION_11.md |
The coaching protocol (required) |
DOSSIER.md |
Your athlete profile (required) |
- Create a Project
- Add instructions to Project settings
- Upload SECTION_11.md and DOSSIER.md to "Project Files"
- Browsing is enabled by default on Plus/Team
- Create GPT → Configure
- Paste instructions in "Instructions" field
- Upload files under "Knowledge"
- Enable "Web Browsing" in Capabilities
- Create a Project
- Add instructions to "Project Instructions"
- Upload SECTION_11.md and DOSSIER.md to "Project Knowledge"
- Enable "Web search" in settings (required for JSON fetch)
- Create Project
- Add instructions to Project configuration
- Upload files to "Sources"
- Create New Project
- Add instructions and upload files
- Web access is available
- Create Gem
- Paste instructions + Section 11 content in instructions field
- Upload dossier or paste contents
Section 11 works well with OpenClaw. The combination of OpenClaw's persistent memory + autonomous execution + Section 11's structured validation makes for a capable coaching setup.
After configuration, test with:
"How was today's workout?"
Good response includes:
- ✅ Fetched data automatically (no asking for it)
- ✅ Session summary with all fields (type, start time, duration, power, HR, TSS, cadence, decoupling, zones, carbs, energy)
- ✅ Training load context (TSB, CTL, ATL, weekly totals)
- ✅ Brief interpretation
- ✅ No "(GitHub)" or URL citations
- ✅ No emojis
- ✅ No false recovery warnings for normal TSB (-10 to -30)
Bad response:
- ❌ "I don't have access to your data, please provide..."
- ❌ Missing session fields
- ❌ "(GitHub)" citations throughout
- ❌ "Your TSB is -23, consider recovery" (when no other triggers present)
- Verify web search/browsing is enabled for your platform
- Check your JSON URL is correct and publicly accessible
- Try a fresh conversation (some platforms cache per-session)
- Manually append a different query param:
...latest.json?v=2
- Check GitHub Actions ran successfully
- Verify Intervals.icu API key is valid
- See examples/json-auto-sync/SETUP.md
Defines behavioral rules for AI coaches:
- No virtual math — AI must use your actual logged values, not estimates
- Explicit data requests — If data is missing, AI asks rather than assumes
- Tolerance compliance — Recommendations stay within ±3W / ±1bpm / ±1% variance
- Framework citations — Every recommendation references specific science
- 11-point validation checklist — AI self-validates before responding (Step 0–10)
Defines rules for AI systems generating or modifying training plans:
- Phase alignment with macro-cycle
- Volume ceiling validation (±10% of baseline)
- Intensity distribution control (80/20 polarization)
- Session composition rules
- Audit metadata requirements
Standardized metadata schema for audit trails:
{
"validation_metadata": {
"data_source_fetched": true,
"json_fetch_status": "success",
"protocol_version": "11.1",
"checklist_passed": [0, 1, 2, 3, 4, 5, 6, "6b", 7, 8, 9, 10],
"checklist_failed": [],
"data_timestamp": "2026-01-23T10:02:07Z",
"data_age_hours": 2.3,
"confidence": "high",
"missing_inputs": [],
"frameworks_cited": ["Seiler 80/20", "Gabbett ACWR"]
}
}The protocol integrates 15+ validated endurance science frameworks:
| Framework | Application |
|---|---|
| Seiler's 80/20 Polarized Training | Intensity distribution |
| Gabbett's ACWR (2016) | Load progression, injury prevention |
| Banister's Impulse-Response | CTL/ATL/TSB dynamics |
| Foster's Monotony & Strain | Overuse detection |
| Issurin's Block Periodization | Phase structure |
| Coggan's Power-Duration Model | Efficiency tracking |
| San Millán's Zone 2 Model | Metabolic health |
| Skiba's Critical Power Model | Fatigue prediction |
| And more... | See Section 11 for full list |
Training blocks adapt dynamically based on real data:
Base → Build → Peak → Taper → Recovery
Phase transitions are triggered by actual metrics (TSB trend, ACWR, RI), not fixed calendar dates.
Automatic load adjustment based on recovery status:
| Trigger | Response |
|---|---|
| HRV ↓ >20% | Easy day / deload |
| RHR ↑ ≥5 bpm | Flag fatigue/illness |
| Feel ≥4/5 | Reduce volume 30-40% |
| RI <0.6 | Mandatory deload |
Green-light criteria for safe load increases:
- Durability Index ≥0.97 for 3+ long rides
- HR drift <3% in aerobic sessions
- Recovery Index ≥0.85 (7-day mean)
- ACWR within 0.8–1.3
- Feel ≤3/5
The protocol is designed to work with Intervals.icu as the primary data source. Set up a JSON mirror that syncs your:
- Fitness metrics (CTL, ATL, TSB, Ramp Rate)
- Recent activities (power, HR, duration, TSS)
- Wellness data (HRV, RHR, sleep, feel)
- Zone distributions
- Planned workouts
Also compatible with:
- Any platform that exports structured training data
When sources conflict, trust order is:
- Intervals.icu (primary)
- JSON Mirror (Tier-1 verified)
- Athlete-provided values (<7 days old)
- Dossier baselines (fallback)
"How was today's workout?"
"Analyze my last 7 days against my targets. What's my compliance rate? Any red flags?"
"Have I met the green-light criteria for extending my Friday long ride to 5 hours?"
"Here's my workout file. Did I hit my intervals within tolerance? What does the HR drift tell us?"
- AI still makes mistakes — This protocol reduces errors but doesn't eliminate them
- Not a replacement for human coaches — Best used alongside professional guidance for serious athletes
- Requires honest data — Garbage in, garbage out
- No medical advice — Consult professionals for health concerns
This is an open protocol. Contributions welcome:
- Bug reports — Found an inconsistency? Open an issue
- Framework additions — Know a validated model that should be included? Propose it
- Translation — Help make this accessible in other languages
- Integration guides — Built a tool that uses this? Share it
This work is licensed under CC BY-NC 4.0.
You can:
- Use it for personal training
- Share and adapt it
- Build non-commercial tools with it
You must:
- Give appropriate credit
- Link to the license
- Indicate if changes were made
You cannot:
- Use it for commercial purposes without permission
Commercial licensing: Contact crankaddict69@proton.me
- David Tinker — Creator of Intervals.icu
- Clive King — Pioneer of GPT-based endurance coaching and URF
- Intervals.icu Forum community
- Researchers behind the scientific frameworks cited in Section 11
- Section 11 A/B/C Protocol
- Dossier Template
- JSON sync automation scripts
- Protocol: SECTION_11.md
- Template: DOSSIER_TEMPLATE.md
- Intervals.icu: intervals.icu
- Discussion: Intervals.icu Forum