From prompting AI
to orchestrating the work.

Before Many Hats, I drafted plans manually, polished them with a generalist agent, and delegated tasks to generalist sub-agents. I still supplied the specialist review and revision that brought the pieces together.

I built Many Hats around my own delegation patterns. Liza, the product-lead role, helps organize complex work and coordinate specialists with relevant methods, templates, and context. The Design Dash supplied stage-oriented methods; ongoing work also needed stable responsibilities between stages.

I reviewed public agent collections for role boundaries and verification patterns, then wrote contracts defining each role’s job, evidence expectations, and approval limits.

I separated stable roles, reusable skills, shared facts, and project deliverables. A project correction belongs with the project; a reusable practice change needs a motivating case, an evaluation, and human approval.

The alternative was to let every project rewrite the team’s core instructions. Keeping the layers distinct makes corrections easier to route and prevents local assumptions from becoming universal rules. It also adds coordination: maintaining context and selecting roles takes work, so specialization has to earn its cost.

The shared object library belongs in Knowledge. A domain specialist stewards it; every role can cite it. Projects may reference or fork an object without silently rewriting it for everyone.

The independent critic reviews direction without inheriting responsibility for delivering it. If a reviewer also has to protect the proposed plan, agreement becomes easier to confuse with evaluation.

I kept the challenge as a citable artifact alongside the direction and response. The human still owns the tradeoffs and external actions. Specialists can prepare recommendations; role agreement does not grant approval.

Library Holds is a fictional holds-and-pickup demonstration. The tempting scope included branch transfers, waitlist incentives, and a notification preference center. The core need was simpler: patrons needed to know when a book was ready and when to collect it.

The product lead proposed a Hold MVP. The domain specialist defined Hold, Patron, Item, and Pickup Window; flow and content roles worked from that model. The independent critic challenged the additional features. The accepted direction cited that challenge and left them out.