Digital Innovation Agents

Appearance

A full V-Model run ​

This tutorial takes a rough idea through all 7 phases of the V-Model workflow, from first conversation to final release closure. You see every phase, every handoff, and every artifact. Think of it as the reference trip. Once you have done this, you understand the whole system.

Timeframe to follow along: about 30-45 minutes of reading, plus another 1-2 hours if you actually run it for a PoC project.

The setup ​

Start with a plain project directory and invoke the orchestrator:

/v-model-workflow

I want to build X. Help me go through the full cycle.

The orchestrator asks where you are in the V-Model. For a new project, pick A) Starting from scratch. It launches /business-analyse.

Phase 1: Business Analysis ​

See the first Business Analysis tutorial for the detailed walkthrough. For this overview, the key outputs are:

  • BA-{PROJECT}.md (Business Analysis document)
  • EXPLORE-{PROJECT}.md (Exploration Board for PoC/MVP)
  • HMW question, personas, critical hypotheses

Methods the BA agent will propose

During this phase the agent stops the interview whenever a gap appears and points you at the matching field method. Most common entry points:

Full catalog under Discovery methods, Ideation methods, and Validation methods.

At the end, the skill asks: "Start /requirements-engineering now?" Say "yes". The orchestrator passes the handoff context automatically.

Phase 2: Requirements Engineering ​

/requirements-engineering reads the BA and produces:

  • Epics (EPIC-{NNN}-{slug}.md) with Hypothesis Statements derived from the HMW question
  • Features (FEATURE-{EPIC}-{NNN}-{slug}.md, epic-local numbering with 3-digit epic and feature numbers) with tech-agnostic Success Criteria and Technical NFRs
  • Architect handoff (architect-handoff.md) aggregating ASRs and NFRs

Key rule: Success Criteria must be technology-neutral. No "OAuth", "REST", "PostgreSQL". Those go into Technical NFRs. See Tech-agnostic Requirements.

Methods the RE agent will propose

If the BA input has gaps, the RE agent does not invent content. It sends you back to the user with the matching method:

  • Jobs to be done when a feature has only functional user stories and the emotional and social layers are missing.
  • Expert conversations when a Technical NFR reads "fast" or "secure" and needs an actual number before it can become an ASR.
  • Expert review when an ASR is suspected but unverified and a full test would be too expensive.
  • User journey focused on the "before" phase when the Epic Hypothesis cannot name the current alternative.
  • Test grid and Value proposition quantification when a Benefits Hypothesis needs a measurable success signal.

Handoff ritual ends with: "Start /architecture now?" -> yes.

Phase 3: Architecture ​

/architecture transforms requirements into proposals:

  • ADRs (ADR-{NNN}-{slug}.md) in MADR format, one per Critical ASR
  • arc42 documentation (scope-dependent section count)
  • plan-context.md as the context bridge to implementation

Each ADR must have at least 2 alternatives with pros and cons plus a justified recommendation. "We chose React because it's popular" is not acceptable.

Handoff: "Start /coding now?" -> yes.

Phase 4: Coding ​

This is the most involved phase. /coding has four steps:

Step 1: Load context ​

Reads plan-context.md, all ADRs, Features, CLAUDE.md, and optionally the backlog, bug log, and last handoff entry.

Step 2: Critical review ​

Reviews the design proposals against the real codebase. This is the v1 differentiator that Digital Innovation Agents never lost. ADRs that conflict with existing patterns are flagged. Changes are written back into the source artifacts before implementation begins.

Step 3: Implementation (delegated to Default agent) ​

Hands off to the Default Claude Code agent with a precise briefing:

  • Phase 3a: Task-breakdown guidelines (bite-size tasks, no placeholders)
  • Phase 3b: Optional TDD mode (activate with "enable TDD")
  • Phase 3c: Debugging protocol if a bug appears (4-phase root-cause process, "architecture alarm" after 3+ failed fixes)

Step 4: Completion and final sync ​

  • Phase 4a: Verification gate. No completion claims without fresh evidence. "Tests pass" requires actual test command output with 0 failures.
  • Phase 4b: Regression test cycle for bug fixes (red-green-red verify)
  • Final synchronization: Feature specs, ADRs, backlog, and bug log updated

Bugs found during implementation land in _devprocess/context/20_bugs.md with FIX-NN IDs, causal chains, and priority.

Handoff: "Start /testing now?" -> yes.

Phase 5: Testing ​

/testing has two modes depending on whether /coding ran in TDD mode:

  • With TDD: focus on integration tests, unit-test gaps, coverage check
  • Without TDD: takes over unit test creation (fallback mode)

Failing tests trigger a fix-loop with 4 user options (fix all, approve one-by-one, adjust tests only, abort). The loop runs until all tests are green or the user aborts.

Handoff: "Start /security-audit now?" -> yes.

Phase 6: Security Audit ​

/security-audit runs 6 phases:

  1. Reconnaissance (tech stack identification)
  2. SAST (static code analysis, CWE-based)
  3. OWASP Top 10 (A01-A10)
  4. OWASP LLM Top 10 (LLM01-LLM10, if applicable)
  5. SCA (software composition analysis, dependency vulnerabilities)
  6. Zero Trust & Code Quality

Findings are prioritized (Critical / High / Medium / Low) and a fix-loop identical to the testing fix-loop runs. Deferred findings land in the backlog with full traceability.

Handoff: "Start Phase 7 Release Closure now?" -> yes.

Phase 7: Release Closure ​

The final phase. /v-model-workflow runs 5 steps:

  1. Final artifact synchronization (cross-phase): BA, Features, ADRs, arc42, plan-context all reflect the actual implemented state
  2. Release notes generation: implemented features, fixed bugs from 20_bugs.md, security findings, breaking changes
  3. CHANGELOG update: [Unreleased] -> [{version}] - {date}, semver bump
  4. Backlog cleanup: open items referenced, future ideas captured
  5. Closing report to the user with a full summary

Artifact trail after a full run ​

_devprocess/
  analysis/
    BA-{PROJECT}.md
    EXPLORE-{PROJECT}.md
    security/AUDIT-{PROJECT}-{DATE}.md
  requirements/
    epics/EPIC-001-*.md
    features/FEATURE-001-001-*.md, FEATURE-001-002-*.md, FEATURE-002-001-*.md, ...
    handoff/architect-handoff.md, plan-context.md
  architecture/
    ADR-001-*.md, ADR-002-*.md, ...
    arc42.md
  context/
    10_backlog.md
    20_bugs.md
    30_handoffs.md   (full audit trail of phase transitions)

Plus the actual source code under src/ and tests under tests/.

Opting out mid-workflow ​

At any handoff, if you do not want the orchestrator to proceed, just say "stop" or "I want to check first" or ask an unrelated question. The workflow pauses immediately. You can resume later by re-invoking /v-model-workflow.

What's next ​