Building the Build Process
The Coaching Animator project is maturing — not just in features, but in how I build. SpecKit mastery, design audit discipline, model delegation, and a tentative step toward self-hosted LLMs with LM Studio.
Since the April restart, the Coaching Animator project has shifted. It's no longer just about shipping features — it's about engineering how I ship them. The workflow itself became the deliverable.
SpecKit Mastery
The build pipeline has evolved through three generations: raw Spec-Kit cold-starts in January, CLEO task orchestration in February, and now a full subagent-driven chain. The current flow:
/speckit.specify → /speckit.clarify → /speckit.plan → /speckit.task → /speckit.implement
With /impeccable (design audit) and /audit (architecture assessment) layered on top. Work gets distributed across specialized agents rather than crammed into one marathon session. This matches where the project is now: well-defined tasks, clear interfaces, mechanical execution.
The Audit Discipline
Design audits are powerful — and dangerous without a gate. A full /impeccable run generates 20–40 recommendations. My first instinct was to implement them all. That's a trap. The correct pattern is triage: identify the 20% that deliver 80% of value, implement only those, and mark the rest as deferred (not closed). Audit → everything → implement is scope creep. Audit → triage → implement is discipline.
Model Delegation That Actually Works
I've settled into a delegation strategy that matches models to their strengths:
- Opus — planning, architecture decisions, spec writing. Anything ambiguous or high-stakes.
- Sonnet — complex spatial reasoning, new component work, anything requiring judgment.
- Haiku — mechanical pattern-following: repetitive refactors, file moves, test scaffolding. Handled 17 of 21 tasks in the last sprint.
The catch: Haiku fails silently on ambiguous types. The TypeScript gate (npx tsc --noEmit) catches this before it compounds. Token rationing has become a workflow skill, not a sign of project failure — I switch to Sonnet for execution and reserve Opus for the next session's planning.
New Frontier: Self-Hosted LLMs
On the side, I've started tentatively exploring local models. Gemma 4 running through LM Studio, paired with n8n for agentic workflow automation. It's early — the models aren't yet matching cloud-hosted performance for the complex reasoning my project demands. But the direction is clear: models that run on my machine, not someone else's. The economics and privacy story get more compelling every month.