AI Agents & Workflows

A daily coach for my running and CrossFit training. Each day it turns activity data into one decision: keep the planned session or adjust it.

I configured it as a Claude project with the Strava connector as its main data source, Todoist for follow-ups, and working documents: season plan, training zones, logs. Sleep data has no connector, so a daily screenshot fills the gap. It acts as an active coach that argues for its advice and fits the plan to the week I actually have, but it never makes medical calls or fills in missing data.

• Reads the morning’s sleep data and decides whether today’s session stands.
• Summarises yesterday’s training with my own effort and pain notes.
• Reviews the week and plans the next one day by day.
• Follows open injury topics in every summary.

The first version was too cautious and flagged risk constantly. I changed the default to load-ready: it pulls back only when I report a problem or the data shows a concrete signal.

A daily assistant for task and life management. It keeps one Todoist system consistent, so everything I capture is filed, prioritised and closed properly.

I configured it as a Claude project connected only to Todoist, with a written standard for projects, labels and priorities. Its scope is deliberately narrow: it manages and plans, and when something needs real content work it writes a starting prompt for another project. It never deletes tasks or opens a new section without asking.

• Files new tasks and ideas in the right place with a priority.
• Adds a decision note before closing a task and opens a new one for the next step.
• Moves ideas through a simple lifecycle until each becomes a task or is dropped.
• Handles reminders in conversation, since my Todoist plan cannot set them through the API.

Early on it acted on an ambiguous request and closed the wrong task. Since then the rule is simple: if a request could mean two things, it asks.