# cycling-training

A modular **Claude skill** that provides periodization logic, workout design, research-backed
protocols, and interpretation guidance for endurance and off-road cycling training — designed to
**companion an Intervals.icu MCP server**, not replace it.

## The companion-to-MCP architecture

This skill deliberately holds **no data and no calculators**. It is the *reasoning* half of a
two-part system:

| Layer | Owns | Examples |
|-------|------|----------|
| **Intervals.icu MCP server** (data + computed metrics) | Fetching and computing everything numeric | activities, wellness, power curves, CTL / ATL / TSB, training-load ramp, zones, TSS, best efforts, streams |
| **`cycling-training` skill** (knowledge + reasoning) | Deciding what the numbers *mean* and what to do next | periodization logic, workout design, research-backed protocols, interpretation of trends, confound-aware caveats |

**The split is a hard rule.** If a script would only reproduce math the MCP API already returns
(zones, TSS, load metrics, CTL/ATL/TSB), it is intentionally absent. Ask the MCP server for the
number; ask this skill what to do about it.

### How they work together (typical flow)

1. **MCP** returns the data — e.g. `get_athlete_summary` (CTL/ATL/TSB), `get_wellness_data`
   (HRV, RHR, sleep), `get_athlete_power_curves`, `get_activities`.
2. **Skill** supplies the reasoning — e.g. "TSB is −25 eight days out from an A-race; the taper
   doc targets +5 to +15 by race day, so shed fatigue," or "this VO2 session banked little time
   ≥90% VO2max; the interval doc says check whether rep length / work:rest kept power in the band."
3. **You (Claude)** combine them into a recommendation, always citing which layer supplied what.

## Repository layout

```
cycling-training/
├── SKILL.md                      # Router — loads the right reference doc by need
├── README.md                     # This file
├── references/
│   ├── durability.md             # Fatigue resistance as a trainable quality + field test
│   ├── mtb-xco-demands.md        # MTB/XCO demand profile; power governs, HR caveat
│   ├── strength-for-cyclists.md  # Heavy strength; masters rationale; from-zero ramp
│   ├── vo2max-intervals.md       # Maximizing time ≥90% VO2max; work:rest; rep length
│   ├── periodization.md          # Distribution philosophy; base/build/peak; CTL ramp; taper
│   ├── data-confounds.md         # HR reliability; power-governs; decoupling/RHR caveats
│   └── citations.md              # Master citation list with PMID/DOI — single source of truth
└── assets/
    ├── plan-template-base.md         # Backward-mapped base mesocycle skeleton (3:1 / 2:1)
    ├── plan-template-build.md        # Build block: VO2 + durability + strength integration
    └── field-test-durability.md      # The ~1,500 kJ → 1-min hill durability field test
```

## Installing as a Claude skill

A Claude skill is a directory containing a `SKILL.md` with YAML frontmatter (`name` +
`description`). To install:

**Claude Code (project or personal):**
```bash
# Personal (available in every project):
git clone <this-repo> ~/.claude/skills/cycling-training

# Or project-scoped (checked in with a repo):
git clone <this-repo> .claude/skills/cycling-training
```

Claude auto-discovers any `SKILL.md` under a `skills/` directory. The `description` field in the
frontmatter is what Claude matches against to decide when to load the skill, so keep it intact.

**Claude Desktop / claude.ai (import a zip):** the skill ships as a packaged `.zip` whose single
top-level `cycling-training/` folder holds `SKILL.md` and the docs — the layout Claude expects. Get
the zip one of two ways, then import it in **Settings → Skills**:

- **From a release:** download `cycling-training-<version>.zip` from the repo's Releases (produced
  automatically — see *Packaging & releases* below).
- **From the packages registry:** the same zip is published to the Gitea generic packages registry:
  ```
  https://git.farh.net/api/packages/farhoodlabs/generic/cycling-training-skill/<version>/cycling-training-<version>.zip
  ```
- **Build it yourself:**
  ```bash
  scripts/build-skill-zip.sh v1.0.0   # -> dist/cycling-training-v1.0.0.zip
  ```

Once installed, this skill loads automatically when you ask cycling-training questions, and it
pairs with your connected Intervals.icu MCP server for the live data.

## Packaging & releases

`scripts/build-skill-zip.sh [version]` stages the installable files and produces
`dist/cycling-training-<version>.zip` (validating that `SKILL.md` carries `name`/`description`
frontmatter). A Gitea Actions workflow, `.gitea/workflows/release-skill.yml`, runs it:

- **Push a tag `v*`** (e.g. `git tag v1.0.0 && git push origin v1.0.0`) → builds the zip,
  **attaches it to a Gitea release** for that tag, and **publishes it to the packages registry**.
- **Run the workflow manually** (workflow_dispatch) → builds the zip and uploads it as a run
  artifact you can download.

The workflow uses two Actions secrets: `RELEASE_TOKEN` (repo write — creates the release and
uploads the asset) and `REGISTRY_TOKEN` (package write — publishes to the generic packages
registry). A registered Actions runner and Actions-enabled repo are also required.

## How to read the docs

- **`SKILL.md` is a router, not a manual.** It points to the one reference doc that answers the
  question at hand. Load docs on demand; don't read them all up front.
- **Every quantitative claim carries a citation** (author/year, with PMID/DOI in `citations.md`).
  This is deliberate: it lets future-you *verify* rather than trust.
- **Where the mechanism is established but the optimal dose is not, the docs say so** — marked
  `⚠️ under-researched — track individual response`. Treat those as hypotheses to test on
  yourself, not settled protocol.

## Scope & honesty notes

- Content is written for **balanced road/TT and MTB** use; discipline-specific numbers are
  labelled and not over-transferred (small-n elite XCO data ≠ masters marathon reality).
- This skill gives **general training-science reasoning, not individualized medical advice.**
  Nothing here overrides a physician, and readiness/overtraining calls should use the
  confound-aware logic in `references/data-confounds.md`, never naive single-number rules.