Case study

Context engineering: a doctrine every writing agent loads

Five competing schools of copywriting, graded on their evidence and resolved into one set of rules an agent can follow.

evidence-graded research briefs: 5evidence-graded research briefs
axioms the agents apply: 10axioms the agents apply
tensions resolved with a written stance: 8tensions resolved with a written stance

Figures measured 2026-09-13.

The problem

Prompting an agent to "write good copy" produces the average of everything it has read. Copywriting advice also contradicts itself: urgency against trust, narrative against front-loading, voice against clarity.

What was built

Wrote five research briefs, one per school (direct response, conversion copywriting, positioning, jobs-to-be-done, UX writing), each graded on the strength of its evidence. Only message-level changes carry measured effect sizes in A/B research, and the brief says so.

Resolved each conflict with a written house stance and a reason: manufactured urgency is banned, prices are fully transparent, clarity wins when it collides with voice.

Packaged it as a skill the agents load: the doctrine, a context map (persona, awareness stage, page family, entry point), and step-by-step workflows for each writing job.

What changed

Every page rewrite since runs against the same written standard, so a reviewer checks output against rules rather than taste.

Inside your team

Any team using LLMs for customer-facing text needs exactly this: the house standard written down where the model can load it, with the conflicts already decided.

Skills shown

Talk about a role All case studies

The Multi-Location AI Swarm Blueprint

Get the deployment order for a marketing swarm across multiple locations — which agent goes first, what it depends on, and what has to be true before agent number two pays for itself.

The dependency map: which four foundation agents every other agent reads from, and why deploying a surface agent first strands it.

The 90-day sequence, week by week, with the acceptance test that closes each week.

The four loops that have to close — capture, decide, act, emit — and the failure mode when any one of them stays open.