Before any execution, run the generated plan through an automated checker across multiple quality dimensions, revising until it passes — a gate that prevents flawed plans from reaching code. In GSD, gsd-plan-checker verifies plans across eight dimensions with up to three revision cycles inside gsd-plan-phase.
Applied by (backlinks)
GSD:
- gsd-plan-phase — hosts the verification loop before execution.
- gsd-plan-checker — verifies across eight dimensions, up to three revisions.
Spec Kit:
- speckit-analyze — cross-artifact consistency & coverage gate, checked against the artifact-constitution, before implementation.
BMAD:
- bmad-check-implementation-readiness — cross-document completeness gate over the whole planning set (PRD + UX + architecture + epics), returning PASS / CONCERNS / FAIL before any story is built.
A third framework promotes this to a three-framework pattern. The gates differ in what they check: GSD scores a single plan on eight quality dimensions, Spec Kit checks consistency and coverage across multiple artifacts (spec ↔ plan ↔ tasks) plus constitutional compliance, and BMAD checks completeness across the full planning document set before implementation.
Compound Engineering:
- ce-doc-review — gates the requirements/plan with a reviewer-persona fan-out before execution.
- ce-spec-flow-analyzer — surfaces missing flows / ambiguous requirements / edge cases before implementation.
gstack:
- gstack-plan-eng-review — locks architecture / edge-cases / tests, gating the plan before execution.
In Fowler’s harness engineering terms this is an inferential sensor run at the earliest useful point — a feedback control on the plan before any code exists (topic-harness-engineering).
Seeds — the cluster’s first gate that is not a model:
- seeds-plan-submit — an AJV schema generated from the plan template checks required sections,
min_length,min, step-index range, and self-references before any child seed is spawned. Failure is pre-write and returns a patchable partial-state diff (errors[].path/code/fixplus the plan as submitted), so the agent revises rather than restarts — the framework’s “one-shot with resume”. - seeds-plan-validate — re-runs the gate against the current template definition, the mechanism behind its claim that a failure should be fixed in the planning process rather than in code review.
The trade against the five model-based gates above is sharp in both directions: this one cannot be argued out of its verdict, and it checks shape rather than quality — it notices a missing risks array, never a wrong approach.
See Also
- pattern-spec-driven-development — the broader method this gates.
- topic-harness-engineering — the guides/sensors control system this is a sensor within.
- stage-plan — the stage where this applies.