Tier-2 ideas from the 2026-08-14 BMAD v6.11 review (self-development/brief/03_competitive_landscape.md,
local-only). Deferred, not rejected: good fits that lose the ROI contest against the five
BACKLOG items. Promote by moving an entry into BACKLOG.md unchanged (Scheme A: this backlog
uses the repo-wide monotonic BL-#### space for future items too, per
backlog/README.md, so promotion needs no renumbering).
All entries are methodology/-touching ⇒ T2 (maintainer-authored) like the rest of the epic.
Replace prose-only scope sections with three actionable tiers: Always (invariants),
Ask First (decisions that trigger halt-and-ask — wires directly into the existing
escalate-by-default stance), Never (non-goals and forbidden approaches). Sharper for
agents than “Out of scope” prose alone. Source: the Boundaries & Constraints block in
BMAD’s spec-template.md. Targets: templates/, 04_backlog_items.md, 03_epics.md.
Promoted and shipped in v1.26.0 — see ARCHIVE.md. Promoted out of order (ahead of BL-0020) at maintainer direction, after E06 had already closed; the entry stays here struck through so the deferral decision and its reversal are both legible.
For long epics, an optional one-page digest distilling strategy/pillar/charter constraints an
implementer actually needs (“by purpose, not by source”), with the mechanical staleness rule:
stale if any upstream planning doc is newer than the digest — regenerate before trusting.
Builds on BL-0019’s budgets. Defer until BL-0015/BL-0019 land (they may make this unnecessary
for most projects). Source: BMAD compile-epic-context.md (800–1500-token epic context
with mtime invalidation). Targets: 03_epics.md, 08_lessons_and_memory.md.
Extend brownfield adoption stages A–B with a reverse-engineering session: inventory the
existing conventions and ratify them rather than inventing new ones, and run every command
before writing it down (verified, not guessed, project facts). Output feeds the instruction
file and pillar backfill (stage E). Source: BMAD bmad-project-context discovery mode +
architecture “ratify the conventions already there.” Target: 00_README.md brownfield
section.
Second source (2026-08-19, E09). GitHub Spec Kit ships a convergence command that does the
same move from the other direction — assess the existing codebase against the spec/plan/task
artifacts and append the remaining work rather than assuming the artifacts describe reality. Two
independent tools converging on “reconcile the written record with what is actually there” is a
stronger case for this entry than the single BMAD source it was filed on. Noted rather than
promoted: E09 was deliberately scoped to one convention, and this one is still a 00_README.md
restructure.
Guidance for the moment work reaches the human final gate: instead of a raw diff, provide an
oriented walkthrough — what changed and why, where to look first, how to test it, path:line
references — then stop talking and let the human review (“front-load, then shut up”).
Cheap addition; the user gate is our binding constraint, so making review fast compounds.
Source: BMAD bmad-checkpoint-preview. Targets: 07_definition_of_done.md,
11_human_roles.md.