| name | vzt-design |
| description | Plan the given task on Opus 5 at max effort — the `opus@max` rung. Use for architecture, technical specs, roadmaps, migration plans, and breaking down a feature when the planning needs full conversation context. This is the DEFAULT planning turn; use /vzt-plan (Fable) only when the design is genuinely novel. Usage: /vzt-design <task>. |
| model | opus |
| effort | max |
VZT Design — planning turn on the opus@max rung (Opus 5)
This turn runs Opus 5 at max effort (the skill's model + effort override). It is
the default planning turn of the protocol. Opus 5 is a step change on deep
reasoning at half Fable's cost, so /vzt-plan (Fable) is reserved for planning
with no prior art — a novel or greenfield architecture, or a one-way-door
distributed-systems call. Everything else lands here.
This is technical design — architecture, schemas, APIs, migration plans.
Visual design (look and feel, tokens, palette, brand) is a different lane:
/vzt-ui. The word is overloaded; if the ask is about how something looks rather
than how it is built, use that instead.
Produce an execution-ready plan for the requested task:
-
Read the actual code/config the plan touches before deciding anything. A plan
built on a remembered API is a guess with steps in it.
-
Pick one approach; defend it in two sentences. Mention an alternative only if
the trade-off is genuinely close — give a recommendation, not a survey.
-
Plan at the scope asked. Don't widen the brief into a redesign or quietly
narrow it; if the ask looks wrong, say so in a sentence and plan it anyway
with the concern flagged.
-
End with a step-routing table so each step runs on the cheapest sufficient
tier next turn:
Tier values: haiku (mechanical/recon), sonnet (default execution),
opus (dense/tightly-coupled steps). Never tag a step opus@max or fable —
those are planning rungs, not execution tiers.
-
Give every step a machine-checkable verification (command + expected result),
chosen before the step is built.
-
Flag the load-bearing seam — the one or two steps that earn an Opus review.
Escalate once, honestly. If the design turns out to have no prior art to
reason from, stop and say so: /vzt-plan takes it to Fable with full context.
Don't do frontier work at this rung silently, and don't escalate to look
thorough — most planning does not need Fable.
Do not implement anything this turn. The session model returns next prompt;
execution belongs to vzt-builder/vzt-mechanic per the table (or /vzt-build
to execute in-context on Sonnet). If the work is big enough that the plan is
really a program of work, point at /vzt-ship — it puts the spec on disk where
compaction can't eat it.
