Skip to main content

​
Learning

The gated-taskset prompt pattern (originally used for single-repo feature work) scales cleanly to cross-app polish passes that touch many submodules in one session. The discipline that makes it work:
  1. Outline first, halt before each set. Before any edit, present numbered TASKSETS and wait for GO TASKSET N. This converts a sprawling “fix several things” request into a sequence of independently-verifiable units.
  2. Order by dependency, not by complexity. Brand assets first (TS1) because every later set consumes them. Auth fix second (TS2) because everything else needs a usable login to test. UI reorder third (TS3) because it’s trivial but easier to validate with auth working. The plan inverts the natural temptation to do “easy stuff first”.
  3. Atomic submodule commits, conventional bodies. Each submodule got one commit per logical change set with a multi-line body documenting why, not what. This made parent superproject ref bumps a single batched chore commit at the end.
  4. Surface architectural surprises mid-flight. TS7 uncovered that atlas, ares, manual, landing are standalone repos rather than registered submodules. Rather than power through, the agent re-scoped: standalone-repo commits stay in their own histories, parent commit only bumps the four registered submodules, and the gap is documented as a follow-up. The gating made the surprise containable.
  5. Vercel deploy failures fold back into the same workflow. When two builds went red post-push, the recovery wasn’t a panicked hotfix — it was a continuation of the gated workflow: reproduce locally → identify root cause → split into doctrine-grade fix → recommit → push.
The key insight: gating doesn’t slow work down; it eliminates the cost of re-scoping mid-execution. A polish pass that touches 8 repos has too many moving parts to hold in working memory. Numbered tasksets externalise the structure and let the agent move fast within each set without losing track of the whole.

​
Cost / Impact

  • 7 tasksets, ~3 hours of focused execution
  • 9 commits across 8 repos, all atomic and reviewable
  • 3 doctrines + 3 findings extracted into atlas as durable knowledge (this learning is one of them)
  • Zero rework, zero scope creep, zero pre-existing dirty state disturbed in the parent
  • Two latent build bugs (validators leak, theme provider drift) caught and codified before they could bite a future deploy
The pattern is now reusable for any “fix several rough edges across the ecosystem” request.