The reconciled tactical execution behind the AI-Native SDLC — the exact skills, hooks, agents, and file names behind each stage, checked against Anthropic's published playbook.
The AI-Native SDLC is the what and why. This page is the how.
Anthropic's playbook runs six stages: Plan → Design → Build → Test → Deploy → Maintain. Sparq splits the first stage into two gates, sold together as one commercial unit — the Evidence Sprint. This is the actual differentiator.
| Sparq stage | Anthropic stage | What Sparq adds |
|---|---|---|
| 0 · Intent Framing | Plan | Structured workshop (2–4 hrs) + strict inline assumption-tagging before intent.md exists |
| 1 · Evidence Gate | Plan | Required external-signal checkpoint (Promote/Iterate/Pivot/Stop) + tracker sync and merge on close |
| 2 · Design | Design | 1:1 |
| 3 · Build | Build | 1:1 |
| 4 · Test | Test | 1:1 |
| 5 · Deploy | Deploy | 1:1 |
| 6 · Maintain | Maintain | 1:1, plus product-analytics bands widened beyond Anthropic's infra-only example |
Commercially: Stages 0–1 together are the Evidence Sprint — the thing that used to be pitched as discovery, now scoped to roughly 2 weeks because the output is evidence and a costed Build decision, not a stack of workshop artifacts. Everything from Design onward tracks Anthropic's canonical mechanics closely.
First half of the Evidence Sprint.
A 2–4 hour structured workshop. 10 min silent assumption dump sorted into Value / Usability / Feasibility / Viability / Operational → failure premortem → architecture exposure with engineering. No untagged claims survive the file.
Meeting-transcription MCP (Whisper/Fathom/Recall.ai) for live capture, plus Miro/Mural/Slack MCP for stickies and threads.
Interactive elicitation — asks probing questions, rephrases claims as testable assertions.
Parses the transcript in real time and auto-tags into the 5 risk categories.
Red-teams the room on technical constraints and unstated dependencies.
Pre-commit. Rejects the commit if any claim lacks an [ASSUMPTION: Category] tag.
Artifact: raw intent.md, 100% tagged. Facilitator validates tagging before it moves.
Second half of the Evidence Sprint — the execution part of discovery.
Pull the highest-risk Value/Usability assumption. Frame a minimal external-facing slice. Attach it to a real client touchpoint — synthetic-only validation is disallowed without a logged exception. Record signal. Product + Delivery decide: Promote / Iterate / Pivot / Stop. On a Promote, close by merging cleared intent.md into /intent/, mapping scope blocks into the enterprise tracker, and defining hard boundary markers.
Evidence Work skills and the resulting evidence-of-signal live in a per-project repo. That's also where the ESOA lives — transitioning from a solutions-side version into the delivery-side version rather than being recreated.
Distinct from the narrow Evidence Slice. Where an Evidence Slice tests one assumption, a Vision Prototype is a holistic build — full breadth, selective depth — meant to show the client the future-state value, not validate a single risk. Optional per engagement, but when built, it sits alongside the Promote decision and the costed Build proposal.
Conversation Intelligence MCP (Gong/Zoom/Teams/Chorus) for signal capture; Enterprise Tracker MCP (Jira/Azure DevOps/ServiceNow) and a non-interactive Claude CI runner for the close.
Maps quotes to tagged assumptions, evaluates signal strength, and caps internal-only feedback at Directional confidence so it can't be mistaken for a real-occasion validation.
Pre-gate multi-persona simulation so live client time isn't burned on an obviously broken pitch.
Blocks the push to /intent/ unless an explicit decision tag and signature block are present.
Automates mapping between Markdown scope blocks and enterprise epic/story schemas.
GitHub Action that fires only when the intent.md frontmatter reads gate_status: cleared, instantiating Stage 2 (Design).
Artifact: intent.md (cleared, synced, merged to main) + per-project ESOA (carried forward) + optional Vision Prototype.
Compressed design session(s), Claude + core engineering leads. Enterprise skills injected into context. Architecture must satisfy the Evidence Gate's operational bounds without over-building beyond intent.md.
Figma MCP or repo-native design — design capability doesn't have to live in Figma. Where the capability is confidently prototype-driven, the design system is built directly in Claude Code / Cursor and lives in the repo, feeding spec.md without a Figma round-trip. Choose per project; both paths converge on the same artifact.
Organization-wide security compliance skill.
Brand compliance skill.
UX standards skill.
Agent behavioral guardrails skill.
Verifies spec.md has every required section (API specs, failure modes, data models) before the stage can transition.
Artifact: spec.md. Design/Tech Lead review.
Git worktree isolation (cap 2–3 per engineer). Plan Mode first — Claude writes plan.md before touching source. Engineer accepts the plan, then code generation starts. Scoped Auto Mode execution within defined paths. Deliberately mimics Anthropic's own Build stage — plan mode, CLAUDE.md, subagents, hooks. Sparq-specific tooling layers on top.
Claude Code CLI, CLAUDE.md as the repo's context engine.
Spins up a fresh context window, runs the app, verifies behavior against plan.md.
Strips redundant abstraction from generated diffs.
For builds with embedded AI agents — evaluates prompt responses for safety/hallucination/tool-calling accuracy.
Intercepts file edits outside the scope defined in plan.md.
Artifact: plan.md and verified code diffs, in isolated branches. Engineer plan acceptance.
Agent self-verification (build, unit tests, visual regression) before submission. Non-interactive CI eval suite runs 20–50 task scenarios. Test files are read-only during bug-fix tasks — the agent must fix the code, never weaken the assertion.
Headless Playwright/Puppeteer MCP, .github/workflows/agent-evals.yml.
Blocks Edit/Write on tests/** during bug-fix tasks.
Blocks PR merge if task accuracy falls below the threshold.
Artifact: verification logs and CI evaluation pass results. Automated CI pass threshold.
Multi-pass PR review (Bugs → Security/PII → spec.md/plan.md compliance). Environment autonomy is tiered: Dev is fully autonomous, Staging requires a clean CI eval run plus automated review, Production requires named human Release Manager sign-off.
Claude Code PR Review Agent.
Multi-pass review writing findings to REVIEW.md.
allowManagedHooksOnly, permissions.deny, sandboxed shell, disableSideloadFlags, allowManagedMcpServersOnly — matching Anthropic's reference settings.json.
Its own named control rather than folded into managed settings generically.
Artifact: REVIEW.md, PR findings, release log. Human Release Manager sign-off on Production only.
Statistical process control on bands.yaml (Western Electric rules). On a breach, a background agent diagnoses root cause and writes a fresh raw intent.md back into Stage 0's triage queue — closing the loop without a human starting it. On-call engineers can also tag Claude directly in incident threads. Every resolved bug becomes a permanent regression case in the Stage 4 eval suite.
Infra MCP (Datadog/Prometheus/New Relic) and Product Analytics MCP (PostHog/Mixpanel/Pendo/Zendesk) feeding the same sensor.
Background monitor triggering triage agents on band breaches.
Turns post-mortem logs into permanent regression tests.
ChatOps bot for on-call — Anthropic's actual Claude Tag product, not a generic ChatOps MCP Bot.
Artifact: bands.yaml updates, incident records, new regression cases added to Stage 4. Service Owner / on-call triage.
Every place the two internal source documents disagreed and the call that was made.
| Stage | Conflict | Call made |
|---|---|---|
| 0 | assumption-extractor (A) vs intent-elicitor + assumption-linter (B) | Not a conflict — different functions. Keep all three. |
| 0 | Workshop length: 60–90 min (both docs) | Extended to 2–4 hrs to reflect real client kickoffs. |
| 1 | Skill named evidence-signal-check (A) vs evidence-synthesizer (B) | Canonicalize as evidence-synthesizer. |
| 1 | Separate confidence-decision.md (A) vs updated intent.md (B) | Canonicalize as intent.md updated in place, in the per-project repo alongside the ESOA. |
| 1 | Old Stage 2 (tracker sync) as its own numbered stage | Folded into Stage 1 as a closing action — automatic result of a Promote. |
| 2 | 4 compliance skills (A) vs 2 (B) | Keep A's full 4; add B's spec-compliance-linter on top. |
| 2 | Figma-only tooling (both docs) | Added repo-native design as an equal path, not a fallback. |
| 3/4 | Test-lock hook placed at Build (A) vs Test (B) | Move to Test — matches Anthropic's own placement. |
| 3 | prompt-evaluator.md present (A) vs absent (B) | Keep — relevant for embedded-AI-agent builds. |
| 5 | Bundled MDM settings (A) vs split config-engine + network hook (B) | Canonicalize on B's split, matches Anthropic's reference settings.json. |
| 6 | Generic ChatOps MCP Bot (B) | Rename to Claude Tag — it's a real, named Anthropic product. |
Anthropic treats CLAUDE.md with its own feedback loop — the correction goes into CLAUDE.md the second time an agent repeats a mistake, and PR review flags staleness. Neither internal doc gives CLAUDE.md this maintenance loop; it's currently just a context source.
Anthropic's Deploy stage includes scheduled, model-driven security scanning (Claude Security) running independently of PR review — a recurring scan, not point-in-time, with findings fed back as fresh intent.md. Sparq's only security coverage at Deploy is currently the PR review pass.
Anthropic's playbook defines explicit metrics per stage (e.g., time from intent.md commit to spec.md commit; first-pass CI success rate; time from band breach to intent.md in triage). Neither internal doc names how Sparq will measure whether the pipeline itself is working.