Why this exists: the frozen table only works if it pulls the work the whole way. Left in a drawer until review, it does nothing. And a sub-agent dispatched without the bar will invent its own — usually a lower one.
Do
- Re-read the frozen acceptance table and the relevant quality bar before starting. Aim each stage at its expected value — don't just run to "it executes" and wait for Step ④ to catch problems.
- Pick the superpowers execution skill that fits: executing-plans (batch with checkpoints), subagent-driven-development (fresh sub-agent per task), test-driven- development, or systematic-debugging.
- Fractal hand-off: when you dispatch a sub-agent, write that stage's expected value × how to verify into its prompt. The sub-agent runs the same 0→⑤ loop inside its scope and returns its own self-check. When you dispatch a spec reviewer, the "spec" you feed it is the frozen acceptance stage, not a vague plan.
- Don't defer judgment-dense work. If it can be finished this session, don't split into "do the mechanical part now, review the hard part later." Schedule only for real future triggers (an external dependency, an observation window), not to push today's hard call into tomorrow.
Note
The sub-agent's self-check is Step ③, not Step ④. Execution verifying itself is not independent verification — that still has to happen in Step ④.
Relationship to superpowers
This skill is thin on purpose. Execution is entirely superpowers' job; this only enforces that the frozen spec travels with the work down to every sub-agent.