Stage: Align

Canonical lifecycle stage: align on what to build and why before implementation — close the human↔agent gap, capture and lock decisions. Framework-neutral name (was stage-discuss, GSD’s term). GSD calls it “Discuss”, Matt Pocock “grilling”, Addy “interview-me”, OpenSpec “explore”, Spec Kit “clarify”. Sits at the front of the loop, ahead of stage-specify and stage-plan. In the pre-spec arc (explore → shape → execute), this is the shape step.

Alignment flavors now show up in the evidence: GSD, Matt, Addy, and Spec Kit align by interrogating the human (pattern-grilling), whereas OpenSpec’s openspec-explore aligns by investigating the codebase and comparing approaches before committing — the explore sub-step of the arc (the counterpart to Addy’s addy-idea-refine ideation). Same canonical stage; complementary techniques. Spec Kit’s speckit-clarify grills after the spec is drafted (resolving [NEEDS CLARIFICATION] markers) rather than before — a sequencing variation, not a different stage. BMAD (sixth framework) is the first to populate all three flavors at once: it grills via the facilitator stance (bmad-analyst), investigates via bmad-document-project, ideates via bmad-brainstorming, and adds a pronounced adversarial pressure-testing sub-flavor (bmad-forge-idea, bmad-prfaq) that hardens an idea before it is specified. A further flavor — establishing governing principles (speckit-constitution) — appears in only one framework and is parked below as a split candidate.

This page is a derived projection — its evidence is the capabilities that implements: [[stage-align]].

Implemented by (backlinks)

GSD:

• gsd-discuss-phase — adaptive questioning, codebase scouting, decision capture → artifact-phase-context.

Matt Pocock — Skills for Real Engineers:

• mp-grill-me — comprehensive interview until decisions resolve.
• mp-grilling — the reusable interview loop.
• mp-grill-with-docs — grilling that also writes a domain model + ADRs.
• mp-to-questionnaire — when the missing knowledge is in someone else’s head: grill the user about the send (who it goes to, what they need back), then write a Markdown questionnaire aimed at the gap, for one recipient to fill in async (delegated-elicitation flavor; v1.2.0). No other framework pages this yet.
• mp-triage — move issues/PRs through a triage state machine → agent-ready briefs (intake flavor; v1.1); the intake counterpart to Compound Engineering’s ce-sweep.
• mp-wayfinder — map an oversized effort (greenfield / bigger than one session) as a tracker map of decision tickets (v1.2.0 rename; formerly “investigation tickets”) that resolve decisions until the destination is clear, then hand off to mp-to-spec / mp-to-tickets (decision-mapping flavor; v1.1). Plan, don’t do — it shapes what to build, upstream of the spec, so it aligns rather than plans.

Addy Osmani — Agent Skills:

• addy-interview-me — one-question-at-a-time interview to ~95% intent confidence.
• addy-idea-refine — divergent/convergent ideation (the explore sub-step).

OpenSpec:

• openspec-explore — investigate the codebase and compare approaches before committing.

Spec Kit:

• speckit-clarify — iterative dialogue that resolves the spec’s [NEEDS CLARIFICATION] markers.
• speckit-constitution — establish immutable project governing principles (governance flavor; see split candidate).

BMAD:

• bmad-analyst — the Analysis persona (Mary); facilitator-stance elicitation over the align workflows.
• bmad-tech-writer — captures project knowledge for AI context.
• bmad-brainstorming — diverge-then-converge ideation (ideation flavor).
• bmad-forge-idea — adversarial pressure-testing of an idea until it hardens or dies cheaply.
• bmad-prfaq — Working-Backwards customer-first stress-test → artifact-product-brief.
• bmad-research — market / domain / technical research → artifact-research-md.
• bmad-product-brief — the pre-PRD scoping document → artifact-product-brief.
• bmad-document-project — brownfield codebase investigation (codebase-investigation flavor).

Compound Engineering:

• ce-brainstorm — dialogue that defines requirements-only → artifact-brainstorm-md (grilling flavor).
• ce-ideate — conceptual-frame + adversarial-filter idea discovery (ideation flavor).
• ce-pov — decisive Adopt/Trial/Hold/Reject verdict on external inputs (evaluate-and-decide flavor).
• ce-strategy — maintain the STRATEGY.md upstream anchor → artifact-strategy-md (governance/strategy flavor; feeds the stage-govern candidate).
• ce-sweep — consolidate external feedback (Slack/GitHub) into inputs for the next cycle (feedback-intake flavor).

gstack (grilling flavor):

• gstack-office-hours — YC-Office-Hours reframing with six forcing questions; challenges premises and generates implementation alternatives → artifact-product-brief (pattern-grilling, pattern-source-grounding). The design doc it writes feeds gstack’s plan reviews.

Superpowers (grilling flavor):

• sp-brainstorming — Socratic one-question-at-a-time refinement; proposes 2-3 approaches; presents the design in sections for approval; HARD-GATE before any implementation → artifact-design-md (pattern-grilling). Its design doc folds straight into sp-writing-plans (Superpowers has no separate stage-specify capability, like GSD).

Agent OS (strategy/product-direction flavor):

• agent-os-plan-product — a one-question interview establishing the standing mission/roadmap/tech-stack docs → artifact-product-docs; the direct counterpart to Compound Engineering’s ce-strategy (both maintain a durable, read-as-context product anchor upstream of the spec).

pstack (codebase-investigation flavor, plus an unattended intake case):

• pstack-how — how does X work; scale-adaptive parallel explorers then one synthesizing explainer → artifact-explainer.
• pstack-why — why it is shaped this way; run-time MCP discovery fans one investigator across each of seven organizational evidence categories (source control, tracker, docs, chat, observability, error tracking, analytics) with an explicit epistemics contract (pattern-source-grounding).
• pstack-benny-triage — unattended Slack issue-report triage with cause-aware routing, tracker dedupe, and fail-closed ticketing (intake flavor; counterpart to mp-triage and ce-sweep, but triggered by an incoming report rather than by a developer).

builder-skills (intake flavor, from production signal):

• builder-factory-collect — scheduled collection and triage of feedback, telemetry, runtime errors, and issues, with explicit coverage accounting (empty ≠ unavailable ≠ truncated), one targeted follow-up question when policy allows, and re-triage when the answer arrives. It joins the intake cluster with mp-triage, ce-sweep, and pstack-benny-triage, and is the first of them designed to run on a schedule with nobody watching.

Cross-framework equivalents

Ten frameworks now realize this stage. The grilling / elicitation cluster (interrogate the human) spans GSD’s gsd-discuss-phase ↔ Matt’s mp-grill-me ↔ Addy’s addy-interview-me ↔ Spec Kit’s speckit-clarify ↔ Compound Engineering’s ce-brainstorm ↔ gstack’s gstack-office-hours ↔ Superpowers’ sp-brainstorming, plus BMAD’s facilitator personas, at pattern-grilling. The codebase-investigation flavor clusters OpenSpec’s openspec-explore ↔ Addy’s addy-idea-refine ↔ BMAD’s bmad-document-project. The ideation flavor clusters addy-idea-refine ↔ bmad-brainstorming ↔ ce-ideate. Compound Engineering adds an evaluate-external-inputs sub-flavor (ce-pov, graded verdicts) with no counterpart elsewhere. Overwhelming evidence this is a genuine canonical stage, not a framework quirk.

Split candidates

stage-govern (establish project-wide governing principles)

• Distinction: govern = set immutable, cross-cutting principles that later phases are mechanically gated against; distinct from aligning on what to build for a given change.
• Evidence so far (one framework — below the ≥2 bar):
  • Spec Kit externalizes principles into artifact-constitution via speckit-constitution and enforces them through speckit-analyze and speckit-plan’s “Phase -1” gates (see pattern-project-constitution).
  • GSD, Matt Pocock, and Addy embed principles inside individual skills (Addy’s per-skill “Red Flags”, MP’s design heuristics) rather than a standing governing document.
  • OpenSpec has no project-wide charter.
  • BMAD (2026-07-04) does not advance it: its project-context.md (from bmad-generate-project-context) and customize.toml carry dev-time rules/conventions — context engineering, not an immutable numbered charter that gates every phase. Principles otherwise live inside the personas.
  • Compound Engineering (2026-07-04) adds a standing upstream anchor (STRATEGY.md via ce-strategy) that ideate/brainstorm/plan defer to — structurally like a charter, but it frames product direction (what/why to build) and is read as context, not an immutable set of principles mechanically gated against every artifact. Related flavor, but not the same govern activity. Still one framework (Spec Kit) for gated immutable principles.
  • Agent OS (2026-08-05) is the sharpest test of the distinction: its entire value is externalized conventions (artifact-standards via agent-os-discover-standards), yet they are injected on demand as advisory context (agent-os-inject-standards), never gated. So it strongly evidences an externalize-standards activity but explicitly not the gated-principles one — it lands with CE on the “read as context” side. The gated bar stays at one framework (Spec Kit); if anything Agent OS suggests the real second cluster is “advisory externalized standards” (→ see topic-harness-engineering, the guide layer), distinct from a gated charter.
• Decisive trigger: a second framework that ships a dedicated set-the-project-principles capability that is mechanically gated → clears the bar → split into stage-govern.

See Also

• stage-specify · stage-plan — the next stages.
• gsd, openspec, speckit — frameworks implementing this stage.