A complete project context engineered from scratch. We extract your domain, identity, constraints, and decision history into a layered specification — pillar stack, identity layer, session protocols, notation. The deliverable is a project you can paste into any model and have it operate as if it's been working with you for months. Buildable when: you're starting a new initiative or your existing context has accumulated enough drift that a rewrite costs less than a patch.

Diagnostic pass on an existing context. We read your current setup against the framework — looking for unspecified layers, contradictory instructions, scope bleed, drift between stated identity and operational behavior, and the failure modes that produce generic output. The deliverable is a written audit with prioritized fixes, plus the patches themselves where the fix is mechanical. Buildable when: you have a context that mostly works but produces output that feels off, drifts across sessions, or doesn't match the position you're trying to occupy.

Engagement-grade work on a multi-project system. Not a single context — an architecture: pillar stack across projects, brand-level canon, cross-project routing, session handoff machinery, signature stack, the whole operating layer. We work in defined sprints with explicit deliverables and decision write-back to your constitution. Buildable when: you're running multiple projects that need to share canon without contaminating each other, or you're standing up a practice that depends on AI as part of the production stack.

Form the destination before the model sees a word. Hold the target clearly enough that every downstream constraint can be measured against it. Envisioning is the work you do before prompting — the step the tutorials skip and the step that determines whether anything that follows has a chance of landing.

Convert imagined destination into AI-legible structure. Translate intent into constraints, parameters, format requirements, examples, and negative space. Specification depth matters more than count — depth is how far into structure a constraint reaches. A single deep constraint displaces output further than ten surface ones.

Issue the structured request and govern the loop. Hand the model the specification and steer the generation in real time — correcting drift, narrowing scope, and adding constraints mid-flight when output reveals an unspecified dimension. The role is director, not author. The model is a construction crew, not a collaborator.

Falsify before accepting. Run NHE-gated checks on output against the original specification. Did displacement actually occur, or did the model converge back to default? Does the artifact satisfy the constraint envelope? What did it invent that wasn't asked for, and is the invention load-bearing or noise? Evaluation is the only step that catches sycophantic agreement disguised as quality.

Deliver the artifact and capture the operation. Output goes out; the specification, prompt, decisions, and evaluation get logged into the project's persistent state so the next iteration starts from organized memory, not from zero. Shipping without capture is the fastest way to lose everything you just built.