Learn By Doing... But Efficiently

After every AI enablement workshop series I run, there are typically two competing feelings from executives.

(+) "My team is now way more capable and knows how to use this technology ~effectively~."

(-) "Now they know what AI can do... are they about to obliterate our token budgets?"

The challenge becomes keeping ROI-focused experimentation going, without capping people's abilities to continue learning through necessary trial and error. Because in reality, no "course" alone is going to get you there. This stuff takes time, energy, and tokens to get the hang of.

However, being effective with AI doesn't mean you can't be efficient with it either.

To remedy this, I've built a skill that you can use with the likes of Claude Cowork, Code, and other agent frameworks called /route.

Its purpose is to:

> Take an idea of what someone is looking to have an agent conduct (e.g. "I want to analyze 40 customer interview transcripts from Granola and organize them into a themes report I can present to a client."

> Present back the ideal model, effort, and orchestration technique to use - along with alternatives (e.g. "Opus 5 - High Effort - One Coordinating Main Agent with 10 Sonnet 5 Subagents on Low Effort - Additional Opus low-effort verification pass)

> Provide a teaching moment on the why behind that decision, based on model purpose and capabilities from announcement pages, public benchmarks and leaderboards, and cost and context metrics.

After using it myself for the last week, I've found myself:

- Spending far less on tokens while maintaining expected standards in results

- Seeing more consistent output from models and being frustrated way less

- Using orchestration techniques I would not have thought about otherwise

- Ultimately, using /route less and less

And that last point is kind of the point? This skill should be an on-ramp, given the teaching moment should help build better organization, strategies, and use of AI over time.

You can start working with the skill using the steps in the Github repo in the comments. And DM with feedback!