A training cycle planning agent

July 5, 2026

An agent that turns office days, a workout split, and menstrual cycle phase into a scheduled week - exportable straight to a calendar. Built for ladies who take their climbing sends as seriously as their sprint commitments.

How it works

graph LR A[CLI prompts] --> B[WeekInput] B --> C[Pydantic AI Agent] C --> D[WeekPlan] D --> E[CLI output] D --> F[.ics export]

You answer a few prompts - office days, long office days, how many strength/climbing/running/yoga sessions you want, rest days, cycle phase, and add any notes. Then the agent returns a day-by-day plan with intensity levels and a load summary, optionally exported to your calendar.

The components

WeekInput bundles the week's constraints:

office_days,
long_days (workdays that run until 7pm),
a WorkoutSplit across strength, climbing, running, yoga and "random",
the current CyclePhase,
and free-text notes.

The system prompt (prompts.py) is where the actual planning logic lives:

describes training load guidance per cycle phase,
defines commute and time constraints on office days,
sets placement rules, e.g. never two high-intensity days back to back, or recovery instructions.

WeekPlan is the structured output, and is enforced with Pydantic AI's native Output model - it contains ScheduledActivity entries (activity, intensity, optional start time, notes) plus a summary and an overall load assessment.

gemma4:e2b runs locally through Ollama, the same setup as the meal planner agent: OllamaProvider pointed at the OpenAI-compatible /v1 endpoint.

calendar_sync.py exports the plan to .ics via the icalendar package, so the week goes straight from terminal to calendar app.

The experience

Compared to the meal planner, the output here is genuinely good, I couldn't have planned better myself. It also never hit the malformed-output retries the meal planner occasionally ran into, the Output came back valid every time. That's likely just because these instructions are much more precise than the meal planner's.

What I learned

It is interesting that classic software design pushes toward abstraction: generalize early, keep things broadly applicable. Agent design pulls the other way: the more specific the instructions, the better the agent performs.

It's also encouraging that a rather highly specialized agent like this runs fine on modest hardware - I use the same 16 GB unified-memory machine from the qwen2.5-coder setup. Watching resource usage while it ran: memory peaked around 7.5 GB and all 8 CPU cores were busy.

You can find the repo and usage instructions on GitHub.