Build With Quality

Build with Quality — Unified Meta-Skill

A development + quality-engineering meta-skill: implement features with EDD/TDD, debug hard bugs feedback-loop-first, stress-test designs before coding, and run a coverage/security/accessibility/chaos quality-gate pipeline. Combines Claude Flow V3 with Agentic QE — one cohesive system in place of several specialized skills.

Pointers (progressive disclosure)

• Activation prompt (start here): BUILD-WITH-QUALITY-PROMPT.md — copy-paste to spin up.
• Worked examples: USAGE-EXAMPLES.md — 5 complete project examples.
• EDD loop: EDD-PROTOCOL.md — Expectation-Driven Development, evidence categories, anti-fox separation.
• Debugging: DEBUGGING-PROTOCOL.md — feedback-loop-first protocol, design interrogation, Diagram-Driven Diagnosis (complex multi-function bug / suspected parallel implementations), and Reasoning Without a Runtime (Static-Oracle Mode) (implementing from a spec or restoring a stub with no shell/compiler/test runner).
• Agents catalog: references/agents.md — 34 agents by domain.
• Methodologies: references/methodologies.md — DDD, ADR (+ ruflo ADR tooling), TDD.
• Quality gates & workflow: references/quality-gates-and-workflow.md — gate thresholds and the 5-phase flow.
• Security gate (deepsec): references/deepsec-security-gate.md — run scripts/deepsec-gate.sh --diff origin/main; policy from [security.deepsec], receipts in .deepsec-gate/reports/. ADR-2033.
• Architecture: references/architecture.md — learning system, memory, model routing, consensus, MCP/CLI execution, config.
• BHIL traceability: references/bhil-traceability.md — PRD→SPEC→ADR→TASK artifact chain, AI-native ADR types, eval/guardrail specs.
• Migration: references/migration.md — moving off agentic-qe / reasoningbank-* / pair-programming.
• Empirical tuning: to optimize any part of this skill against a measurable reward rather than by intuition, use the skill-tuning skill (SkillOpt loop + held-out A/B). The Static-Oracle Mode section was produced by that loop.

When to use

• Implementing a feature that warrants tests and quality gates (TDD/EDD).
• Chasing a stubborn, multi-function bug — use the feedback-loop-first protocol.
• Interrogating a design before writing code.
• Running a coverage / security / accessibility / chaos quality-gate pass.
• Reviewing changed code for vulnerabilities before merge (scripts/deepsec-gate.sh --diff <base>).

When not to use

• A quick code change that does not need quality gates or swarm coordination — edit files directly with Claude Code.
• GitHub-specific PR review without the full quality pipeline — use github-code-review.
• Standalone documentation validation and alignment — use docs-alignment.
• SPARC methodology orchestration without quality-engineering agents — use sparc-methodology.
• Simple unit test generation without the full agent system — write tests directly or use the TDD workflow in sparc-methodology.

Quick start

Option 1 — Copy-paste prompt (recommended). Take the prompt from BUILD-WITH-QUALITY-PROMPT.md and fill in:

Build with Quality skill.
Project: [NAME] | Stack: [TECH] | Task: [DESCRIPTION]
Methodology: DDD + ADR + TDD (EDD design-time first)
Quality: 85% coverage, security scan, WCAG AA

Option 2 — CLI. There is no claude-flow skill ... subcommand; compose the same workflow from primitives:

claude-flow swarm init --topology hierarchical --v3-mode
claude-flow agent spawn --type architect
claude-flow agent spawn --type coder
claude-flow agent spawn --type test-strategist
claude-flow task create --type implementation --description "[TASK]"

Option 3 — MCP tools (when available).

mcp__claude-flow__swarm_init { topology: "hierarchical", maxAgents: 15, strategy: "specialized" }
mcp__claude-flow__agent_spawn { agentType: "architect" }
mcp__claude-flow__agent_spawn { agentType: "coder" }
mcp__claude-flow__agent_spawn { agentType: "test-strategist" }
mcp__claude-flow__task_orchestrate { task: "[PROJECT]", strategy: "parallel" }

Option 4 — Task tool (Claude Code). TinyDancer routes the model optimally:

Task({ prompt: "Implement user authentication with JWT, following TDD", subagent_type: "coder", model: "sonnet" })
Task({ prompt: "Generate tests for auth module with 95% coverage", subagent_type: "tester", model: "haiku" })

Claude Code only: Option 4 uses the Task tool. On Codex / GPT-6 Astra: use Option 2 (CLI) or Option 3 (MCP, if the ruflo MCP proxy is reachable from Codex).

MCP is preferred when mcp__claude-flow__* tools are available; otherwise fall back to the claude-flow CLI (baked on PATH — never npx claude-flow@alpha, which resolves a stale cached version). See references/architecture.md for the detection logic and the per-agent coordination hooks.

Methodology at a glance

• EDD — the design-time conversation layer between human intent and AI implementation. Captures qualitative expectations, ordering invariants, systemic properties, and explicit counter-examples ("must NOT happen") that TDD assertions and BDD templates can't. 7-step loop: formulate → implement → produce evidence → audit → challenge → iterate → stabilize. Executed evidence required (command, raw output, timestamp, git SHA); narrative evidence is auto-rejected. evidence-producer and evidence-auditor must be different agents on different model families. Full playbook: EDD-PROTOCOL.md. EDD runs before TDD/BDD, then hands proven scenarios off to them for permanent regression coverage — it does not replace them.
• DDD / ADR / TDD — bounded contexts and tactical patterns; graph-backed ADRs with depends_on/amends/supersedes plus a verification/staleness contract (verified_commit + verified_paths so CI catches a record drifting from its code — an ADR you can't falsify against the code is documentation, not a control); Red-Green-Refactor with TDD agents that also stabilize proven EDD expectations into tests. See references/methodologies.md.

Security gate at a glance

The Security gate is executed, not asserted: scripts/deepsec-gate.sh runs deepsec in PR mode on the changed files, applies the manifest's [security.deepsec].fail_on threshold and writes receipt.json. Exit 78 means the gate is disabled or the CLI is not baked — record it as SKIPPED, never as passed. See references/deepsec-security-gate.md.

Related skills (complementary)

These work alongside build-with-quality:

• sparc-methodology — higher-level orchestration framework
• swarm-orchestration / swarm-advanced — lower-level swarm primitives
• verification-quality — truth scoring and rollback
• github-* — GitHub-specific integrations
• hive-mind-advanced — specialized Byzantine consensus features

License

MIT