SDLC Agent-Skill Wiki

Pattern: Scale-adaptive planning

4 min read

Size the ceremony to the work: a one-line fix skips straight to code, an enterprise system runs the full spec → architecture → epics → stories pipeline, and everything in between takes a proportionate middle path. Planning depth is a dial, not a fixed gate — “Choose your track based on planning needs, not story math.” This is bmad’s answer to the tension every spec-driven framework faces: rigour that helps big projects becomes dead weight on small ones.

BMAD expresses it two ways across its v6 line:

  • Numeric Levels 0–4 (classic v6) — Single Atomic Change → Small Feature → Medium Project → Complex System → Enterprise Scale, each prescribing which artifacts exist (tech-spec only → PRD → PRD + architecture + just-in-time tech specs).
  • Stakes-calibrated tracks (current v6) — Quick Flow (skip planning) / BMad Method (PRD + architecture + UX) / Enterprise (adds Security + DevOps). Solutioning is skipped, optional, or required per track.

A close relative is just-in-time design: the architecture spine fixes only invariants and defers concrete structure to the code, so detail is produced only when needed.

BMAD:

  • bmad-quick-dev — collapses the whole pipeline to a few turns for small, well-understood work.
  • bmad-spec — the lightweight SPEC kernel used instead of a full PRD on the light track.
  • bmad-architecture — the invariants-only spine (just-in-time design).

gstack:

  • gstack-autoplan — smart review routing sizes the review ceremony to the change (CEO skips infra, design skips backend).

nano-spec:

  • nano-spec-create — applies the pattern at the methodology level: rather than a dial inside one framework, nano-spec is a deliberate fixed middle tier on the no-spec → nano → full-spec (Kiro) ceremony spectrum, chosen because most work lives there. The dial is exercised by choosing nano-spec over the neighbours, not within it.

Agent OS:

  • agent-os-plan-product · agent-os-shape-spec — lightweight-by-default docs, sized to the project; v3’s whole design is subtraction — it deleted its own scaffolding and delegates the heavy lifting to the harness’s Plan Mode (another methodology-level point on the ceremony spectrum, like nano-spec).

Seeds:

  • seeds-create · seeds-plan-prompt — a deliberate two-tier dial with the rule stated in the docs: “use it when work is large or ambiguous enough that an LLM benefits from decomposing it — for small, well-scoped tasks just sd create directly.”
  • seeds-issue-workflow — the skill that encodes the choice as two named procedures (create-then-ready vs. plan-decomposition) with the routing rule between them.

pstack — applied to rigor rather than to ceremony, which is the variation worth noting:

  • pstack-figure-it-out — the rigor level is chosen up front and biased high, with one-way doors and high blast radius getting more and reversible low-stakes steps getting less. “Rigor is gates and artifacts, not ‘try harder’.”
  • pstack-how — a simple question skips exploration entirely and runs one explainer; a complex one fans out two to four explorers. “When in doubt, take the simple path.”
  • pstack-tdd — skips the test when the path would need brittle mocks, slow infrastructure, or large fixture churn. “Prefer no new test over a bad test.”
  • pstack-poteto-mode — the router itself: a large or cross-cutting effort escalates to pstack-figure-it-out even when a narrower playbook fits, and a standing program escalates again to the Orchestrate playbook.

Tessl — the same instinct applied to the review budget:

  • tessl-change-risk — assesses a diff and exits non-zero only when human review is required, allocating reviewer attention by measured blast radius rather than spending it uniformly. The organizational counterpart to pstack-figure-it-out’s rigor-sized-to-blast-radius rule.

See Also

Graph View

Backlinks