When revising a set of related documents where one is the canonical/upstream document and the others depend on it, do not dispatch revisions in parallel. Do them sequentially:
- Revise the upstream document first (e.g. the v4.0.0 architecture spec).
- Read the REVISED text.
- Review each downstream document one at a time, with the revised upstream as reference.
- Stop between downstream reviews so Rich can redirect if needed.
Why: When downstream docs reference upstream sections, a parallel revision means each downstream subagent reads v1.0 of its own doc and a FRAME BRIEF I wrote, not the actual revised upstream text. Cross-references go stale. Framing may be locally consistent but alignment to the real upstream is unverified.
When parallel IS correct: Independent research where each subagent has a distinct subject and there’s no cross-doc referencing. Phase 1 component evaluation across 35 subjects is the canonical case. Phase 2 SCR renderings across 8 SCRs also works because each SCR is independent.
How to tell the difference:
- Independent research subjects → parallel
- Refinement / revision / re-alignment of existing docs that cite each other → sequential
- Anything where the “first” output changes the shape of the “second” → sequential
The mistake to avoid: pattern-matching on an earlier parallel dispatch that worked well and carrying that instinct into a refinement pass where dependencies exist. Efficiency is not the goal; alignment across the doc set is.
How to apply: When Rich asks for a revision pass across multiple docs, think first: do these docs reference each other, is one of them the source of truth the others rely on? If yes, sequential. If no, parallel. Default to sequential when unsure; the cost of one-at-a-time is small, the cost of misalignment is not.