Cookbook

The cookbook is a library of prompts that work well with icuvisor. Each one is written so your AI assistant reaches for the right tools, grounds every number in your real intervals.icu data, and refuses to guess when data is missing.

Two kinds of content:

The prompt library — short, single-message prompts grouped by task. Copy one, adjust the dates, send.
Recipes — longer, reusable prompt templates that drive a multi-step job (a weekly review, a taper, a roster triage). Treat a recipe as a small agent skill: paste it, fill the blanks, and the assistant runs the whole workflow.

Prompt library

One-line prompts for everyday questions, grouped by task.

Weekly training review

Summarize load, intensity, and recovery risk for the last 7-14 days.

Readiness check

Decide whether to train hard today from wellness and form.

Activity retrospective

Break down a single ride, run, or race in detail.

FTP and zones review

Decide whether a threshold has moved and update zones.

Season and block plan

Build a periodized plan toward a goal event.

Build and schedule workouts

Create structured workouts and put them on the calendar.

Race-week taper

Plan the final days before a goal event.

Coach roster triage

Scan a roster of athletes for red flags (coach mode).

How to get reliable answers

icuvisor only helps if the assistant actually calls it instead of answering from memory. These habits — baked into every recipe below — make that happen.

Name your data. Phrases like “using my intervals.icu data” or “use icuvisor” make the assistant fetch real numbers instead of estimating. The recipes always include this.
Give a concrete window. “the last 14 days”, “since 1 April”, “race week” maps cleanly onto date-range tool arguments. Vague windows produce vague tool calls.
Say what not to do. “Do not invent metrics that aren’t available” is the single most effective line for cutting hallucination. icuvisor returns explicit “unavailable” signals — tell the assistant to honor them.
Ask for sources. “Tell me which tool each number came from” turns a plausible-sounding answer into an auditable one.
One job per message. Building a season plan and authoring ten workouts in one turn overruns the context window. The recipes split big jobs into stages on purpose.
Let it look things up. If a recipe needs an activity or event, let the assistant find the ID with get_activities or get_events — you should not have to paste raw data.

⚠️

Subjective scales are not 0-10. intervals.icu sleep quality is 1-4, athlete-reported feel is 1-5, and RPE is 1-10. icuvisor labels every scale in its responses. If the assistant restates one wrong (“a 3 is a terrible sleep”), correct it — a 3/4 is good sleep.

What icuvisor will not do

Knowing the limits keeps you from chasing answers icuvisor cannot give:

Strava-imported activities are returned with most fields blank — Strava’s API terms forbid third-party apps from reading them. icuvisor labels these so the assistant says “unavailable” instead of guessing. Connect your device directly to intervals.icu for full data.
Writes are gated. Creating or updating events, workouts, and wellness rows works only when the server runs with write mode enabled; deletes need delete mode. See safety modes. If a write recipe cannot run, ask for a preview of the change instead.
icuvisor does not compute what intervals.icu does not expose. If a metric (decoupling, VI, a custom field) is not in the data, the recipes ask the assistant to say so rather than calculate a substitute.

Cookbook recipes vs. MCP prompts

If your client supports MCP prompts, icuvisor ships five curated ones (training_analysis, recovery_check, weekly_planning, race_week_taper, coach_roster_triage) you can invoke directly — they carry the same guardrails server-side. The cookbook recipes are the portable equivalent: plain text that works in any client, including ones with no prompt support. Each recipe notes its matching MCP prompt where one exists.