AI-Native SDLC

Standard operating procedures for the Sparq AI-Native Software Development Life Cycle. Adapted from the Anthropic AI-Native SDLC pattern, this methodology integrates the Evidence Gate as a non-negotiable gate inside the pipeline rather than treating it as an external meta-process.

The base pipeline optimizes for speed and build correctness. However, it requires a mechanism to validate whether we are building the right thing before scaling execution. Therefore, the Evidence Gate explicitly gates the transition from Intent Framing to Design — closing with the tracker sync and merge that was previously modeled as its own Plan stage, now folded into the Gate as its automatic closing action rather than a separate decision point. The commit chain (Git-tracked artifacts at every stage) remains both the operational workflow and the audit trail.

Pipeline is 7 stages, not 8. Stages 0 and 1 together are sold commercially as one unit — the Evidence Sprint — matching how discovery has traditionally been packaged, just with real signal behind it.

The pipeline

StageCore artifactGate
0 · Intent Framingraw intent.md—
1 · Evidence Gateintent.md (cleared, synced)Promote / Iterate / Pivot / Stop + PO merge
2 · Designspec.mdSkills validation
3 · Buildplan.md + code diffsCLAUDE.md + hooks
4 · Testverification logs + eval resultsEval pass thresholds
5 · DeployREVIEW.md + PR findingsEnvironment autonomy tiers
6 · Maintainbands.yaml + incident recordsMetric-band triggers

How stages and risk shapes work together

Stages are artifact order. They describe the sequence of what you produce and which governance gate you pass through. Risk shapes are the real-time operating picture — they describe which capabilities run hot right now, regardless of which stage the work is in.

A risk shape fires wherever it fires. The Feasibility shape can spike during Intent Framing if an architecture constraint surfaces early. The Value shape can stay active deep into Build if a prior assumption gets challenged. Shapes co-occur, recur, and persist — they are not bound to a single stage.

The stage tells you what artifact you owe. The live risk mix tells you what to worry about while you produce it. You are always in a stage, and you are always responding to risk shapes.

Stages have a default gravity — 0 through 6 — but change events and failed gates send you back. The change-response doctrine runs at any stage: re-read the risk, re-set the dials, decision gate, re-staff, re-price and re-time, name the next slice.

Stage 0 · Intent Framing

Generate the raw material and problem framing required for the pipeline without prematurely settling on narrative agreement. This session acts as structured elicitation to expose unvalidated assumptions.

First half of the Evidence Sprint.

Discipline

The output must be a stack of tagged, unvalidated assumptions, not settled prose. Confidence is not agreement in a workshop.

Client alignment

As a consultancy, we often do not have standing authority over intent.md. Use this session to bridge the vocabulary gap and secure client agreement on the problem statement before proceeding.

Assumption Dump

Ask: "What must be true for this initiative to succeed?" Individually list, group, and rephrase as testable statements.

Failure Premortem

Ask: "It's 12 months from now. This failed. Why?" List causes and translate them into risks.

Domain Walkthrough

Force the identification of at least one risk per category: Value, Usability, Feasibility, Viability, and Operational. This prevents tunnel vision.

Architecture Exposure

Ask: "What technical constraint could invalidate this?" Identify platform limitations, integration dependencies, performance ceilings, and security requirements.

Output: A raw intent.md (problem statement, proposed outcomes, affected systems, constraints) where every claim is explicitly tagged as an assumption. This file does not trigger CI compilation — it must proceed to the Evidence Gate.

Stage 1 · Evidence Gate

Decide whether the raw intent.md is trustworthy enough to compile into spec.md. This is a human-driven evaluation gate to validate risk categories that cannot be caught by CI rubrics. On a Promote, the gate closes automatically: tracker sync, merge, and spec compilation fire without a separate decision.

Second half of the Evidence Sprint. This stage now includes the tracker sync and merge that was previously modeled as a separate Plan stage — it is the mechanical result of a Promote, not an independent decision point.

Validation loop

Run the cycle of Risk → Assumption → Slice → Signal → Confidence → Decision on tagged assumptions. Address one assumption per slice. The validation loop is prototyping — build the smallest thing that tests one assumption, read the signal, decide.

Test in the real world

For Value/Usability risks, attach the test to a real occasion (next client conversation, stakeholder review, prospect encounter) rather than manufacturing synthetic scenarios. A slice needs something behaving in the world to react to.

Internal convergence

Only use internal convergence checks when a real external occasion genuinely does not exist. Record the reason explicitly. Internal-only signals cap at Directional confidence, never Validated.

Explicit decisions

Make a clear decision for the intent: Promote, Iterate, Pivot, or Stop. No numeric confidence scores.

Close the gate

Once Promote is decided, immediately execute the dual-linkage: Markdown remains the operational source of truth in the repository, synced to legacy enterprise trackers (Jira/Azure DevOps) via MCP connectors for client governance. This merge triggers a non-interactive CI job that compiles spec.md, loading Sparq compliance skills (security, UX, brand guidelines). The Product Owner performs the final review.

Output: The intent.md either clears — merged, synced, and moving to Stage 2 — or gets sent back to Stage 0 for re-framing, or the initiative is stopped/pivoted.

Stage 2 · Design

Synthesize requirements and architecture into a formal specification.

Compressed synthesis

Execute a compressed, single-session requirements and architecture synthesis.

Skill integration

Guide the synthesis utilizing organization-wide skills for security, compliance, and UX standards.

Domain focus

Building is the primary domain — product-interface-building and core-systems-engineering drive the synthesis. Framing constraint-checks run as secondary, ensuring the architecture stays within the boundaries set at the Evidence Gate.

Output: spec.md.

Stage 3 · Build

Draft implementation plans and execute code generation within isolated agent environments.

CLAUDE.md context engine

Maintain a robust context engine detailing build commands, linting rules, architectural patterns, and team-specific mistakes to avoid.

Plan mode as default

AI agents must draft plan.md before writing code. Human acceptance commits the plan to Git.

Parallel worktrees

Isolate agent sessions across separate Git worktrees. Enforce a parallel session cap of 2–3 concurrent worktree sessions per engineer to preserve review quality and prevent fatigue.

Scoped auto mode

Allow autonomous edit execution only where existing test coverage and build hooks are fully mature.

Shared subagents

Utilize .claude/agents/ to store standardized helper agents across repositories — verifier agents (run the app, check behavior) and simplifier agents (strip redundant complexity post-implementation).

Output: plan.md and verified code diffs.

Stage 4 · Test

Continually verify code behavior and assess Feasibility and Operational risks through automated CI suites.

Self-verification loop

Agents must run builds, tests, and visual diffs, proving success in context before reporting a task complete.

Continuous CI eval suite

Maintain 20–50 real-world task evaluations in .github/workflows/agent-evals.yml. These run non-interactively upon updates to code, skills, or hooks to establish baseline agent performance benchmarks.

Focused scope

CI evaluations assess Feasibility and Operational risks only. Value and Usability risks are resolved upstream at the Evidence Gate (Stage 1).

Test-file locking

Implement pre-tool hooks that block agents from modifying existing test files during bug-fix tasks. The agent must fix the code to pass the tests — it is explicitly denied the ability to alter assertions to force a passing status.

Output: Verification logs and evaluation results.

Stage 5 · Deploy

Conduct multi-pass reviews and securely deploy artifacts across environments using defined autonomy tiers.

Multi-pass PR review

Automate first-pass scans for logical bugs, security gaps, and compliance against spec.md and plan.md (REVIEW.md).

Hooks as release gates

Enforce deterministic pre-execution scripts for security controls, blocking unauthorized path edits.

Environment tiers

Dev/Sandbox: full agent execution allowed. Staging: automated PR review with passing CI required. Production: explicit human release-manager authorization required.

Managed settings

Deploy immutable settings centrally (e.g., allowManagedHooksOnly, permissions.deny) to block shell network egress, prevent credential leakage, and disable plugin sideloading.

Output: REVIEW.md, PR findings, and deployed code.

Stage 6 · Maintain

Monitor system health continuously and trigger automated triage and incident resolution loops.

Metric drift triggers

Apply deterministic monitoring using statistical control rules against operational metrics (e.g., test failure rate, post-deploy 5xx rates) via bands.yaml.

Autonomous intent feed

If a metric breach occurs (e.g., 3σ deviation), a background agent automatically diagnoses the root cause and writes a new intent.md into the triage queue.

ChatOps incident response

Triage live production alerts via Claude Tag in Slack/Teams, execute diagnostic runbooks via MCP, and log root-cause analyses directly to the thread record.

Incident-to-eval pipeline

For every resolved production incident, automatically compile a regression evaluation case into the CI suite (Stage 4) to guarantee prevention of recurrence.

Output: bands.yaml updates and incident/eval loop records.

Domains and risk shapes across the pipeline

Each stage has a primary domain driving the work and risk shapes that are typically hottest at that point. Framing tapers over time but does not hard-stop at the Evidence Gate. Proof runs at the Gate, Test, and Maintain stages, handling distinct but related evaluation tasks.

StagePrimary domainSecondaryRisk shapes typically hot
0 · Intent FramingFraming—Problem clarity, Commercial / viability
1 · Evidence GateProofFramingValue, Problem clarity
2 · DesignBuildingFramingFeasibility
3 · BuildBuildingProofFeasibility, AI reliability
4 · TestProofBuildingProof / acceptance, Feasibility, AI reliability
5 · DeployBuildingContinuityProof / acceptance, Adoption
6 · MaintainContinuityEnablementContinuity / operational, Adoption

Risk shapes are listed where they are typically hottest, not where they only fire. Any shape can spike at any stage.

Commercial Commercial frames the pipeline — it sets the envelope and price before Stage 0, and the proof → commercial seam updates it as evidence arrives. It is not absent from the pipeline; it operates across stages rather than owning one.

Enablement Enablement activates when the Adoption risk shape fires, typically hottest at Deploy and Maintain but possible at any stage where organizational change is needed to land the work.

Adaptation matrix

StageAI-Native approachArtifactGovernance
0 · Intent FramingStructured elicitation producing tagged, unvalidated assumptions.raw intent.mdFacilitator ensures output stays assumption-tagged.
1 · Evidence GateRisk validation loop for Value/Usability/Viability, closing automatically into tracker sync and spec compilation on Promote.intent.md (cleared, synced)Product + Delivery sign-off; Promote/Iterate/Pivot/Stop, then PO merge.
2 · DesignCompressed single-session requirements and design synthesis.spec.mdSkills validation (Security, UX, Brand).
3 · BuildPlan-mode drafting, worktree-isolated implementation.plan.md + code diffsCLAUDE.md + path-blocking hooks.
4 · TestContinuous self-verification and CI-driven eval suites.verification logs + eval resultsTest-locking hooks and eval pass thresholds.
5 · DeployMulti-pass agentic PR review with human risk evaluation.REVIEW.md + PR findingsBranch protection and production deploy hooks.
6 · MaintainMetric-band monitoring auto-generating new intent items.bands.yaml + incident recordsTiered automated response and on-call approval.

Known gaps

bands.yaml widening

Maintenance metrics primarily monitor infrastructure health (error rates). A sensor is needed to monitor Value/Usability drift post-launch (adoption rates, usage depth) so the pipeline does not lose signal on a shipped feature nobody wants. This remains an unresolved gap — the tactical playbook includes a candidate widened bands.yaml design as a proposal, not yet ratified.

Naming the Intent Framing Session

The Stage 0 session itself is deliberately left undressed. Naming it before the shape is settled through repeated real use is an anti-pattern. We will call it what it is until it earns a name through institutional habit. This is separate from Evidence Sprint, which names the combined Stage 0 + Stage 1 commercial unit, not the Stage 0 session on its own.

Client-specific authority

How much of the Intent Framing Session includes the client versus internal Sparq preparation is inherently variable by design, not a gap to standardize away. Some engagements run heavy client involvement, others very little; headcount in the room has ranged from as few as 2 to as many as 20. This is engagement-specific and should stay that way.