01 — INTENT
A blurry request becomes a confident wrong answer.
"Make this better" is a vibe, not a spec
I extract a typed goal with constraints, success criteria, and alternative interpretations before a single token of work is generated.
02 — AMBIGUITY
A polished artifact nobody actually asked for.
Under-specified scope, audience, format
An epistemic-humility layer audits each request along fixed dimensions and flags the known unknowns by severity — a decision gate, not a gut feel.
03 — CLARIFIER
The agent turns into an annoying questionnaire.
Seven questions, user abandons
The clarifier searches the web first, auto-resolves what it can, and spends user attention only on what's genuinely unknowable. One question, not seven.
04 — PLANNER
A brittle monolith you can't resume or verify.
Five paragraphs into the wrong answer
Intent becomes a dependency-aware DAG of numbered, independently verifiable steps — auditable and editable before anything runs.
05 — EXECUTOR
Execution loses the thread and can't prove it hit the goal.
Technically correct, goal-irrelevant
Steps run in dependency order with streamed output and a final trace-to-goal pass that maps the deliverable back to the original intent.
∞ — RESILIENCE
One missing dependency takes the whole system down.
No search? No cache? No RAG?
Systems degrade, never die: missing keys fall back to in-memory and local paths. Graceful degradation is a feature, not an error.
Open to new engagements