dreamlab-ai BHIL AI-First Development Methodology
Specification-driven development methodology where artifacts flow through a traceable chain: PRD → SPEC → ADR → TASK → CODE → REVIEW → DEPLOY, with retrospectives feeding back into planning.
Core premise: The bottleneck in AI-assisted development is not code generation — it is specification quality.
When to Use This Skill
- Starting a new feature: "create a PRD for X", "scaffold a feature", "new feature"
- Architecture decisions: "should we use model X or Y", "document this decision"
- Sprint planning: "plan the next sprint", "break this into tasks"
- AI-specific decisions: model selection, prompt strategy, agent orchestration patterns
- Quality gates: evaluation suites, guardrails specs, artifact validation
- Traceability: "trace this bug back to requirements", "impact analysis"
When Not to Use
- For single-file quick fixes — use direct editing
- For code quality/testing execution — use
build-with-quality(complementary, not competing) - For pure code review without methodology — use
code-review-qualityorsherlock-review - For SEO content — use
toprank
Artifact Chain
PRD-NNN (Product Requirement)
└── SPEC-NNN (Technical Specification)
├── ADR-NNN (Architecture Decision — standard, model, prompt, or agent)
└── TASK-NNN (Implementable Work Unit)
└── Sprint S-NN (Planned Execution)
└── Code + Review + Deploy
└── Retrospective → next PRD cycle
Every artifact carries YAML frontmatter with parent/child IDs for machine-actionable traceability.
Templates
| Template | Purpose | ID Format |
|---|---|---|
PRD-TEMPLATE.md |
Product requirements document | PRD-NNN |
SPEC-TEMPLATE.md |
Technical specification | SPEC-NNN |
ADR-MODEL-SELECTION.md |
LLM model choice with benchmarks, cost, re-eval triggers | ADR-NNN |
ADR-PROMPT-STRATEGY.md |
Prompting approach, versioning, quality thresholds | ADR-NNN |
ADR-AGENT-ORCHESTRATION.md |
Agent architecture patterns (swarm, mesh, pipeline) | ADR-NNN |
TASK-TEMPLATE.md |
Implementable work unit with acceptance criteria | TASK-NNN |
SPRINT-PLAN-TEMPLATE.md |
Sprint planning with velocity and capacity | S-NN |
EVAL-SUITE-TEMPLATE.yaml |
LLM evaluation configuration | — |
GUARDRAILS-SPEC-TEMPLATE.md |
Safety and guardrails specification | — |
PROMPT-REGISTRY.md |
Prompt versioning and tracking | PV-NNN |
Guides
| Guide | Content |
|---|---|
00-getting-started.md |
5-minute setup |
01-methodology-overview.md |
Philosophy and principles |
03-sprint-workflow.md |
Step-by-step sprint execution |
04-context-management.md |
Preventing context fragmentation |
05-ai-native-patterns.md |
LLM-specific development patterns |
07-ruflo-ruvector-setup.md |
Ruflo/RuVector integration |
Tools
Run from the skill directory (paths are skill-relative; the baked copy lives at /opt/agentbox/skills/bhil-methodology/tools/).
# Initialise a new project with BHIL structure
bash tools/init.sh
# Create a new ADR
bash tools/new-adr.sh "Adopt bge-small-en-v1.5 for RuVector embeddings"
# Validate all artifacts for completeness and traceability
bash tools/validate-artifacts.sh
Integration with Other Skills
| Skill | Relationship |
|---|---|
build-with-quality |
BWQ executes the code/test/review phases; BHIL provides the spec/ADR/traceability wrapper |
sparc-methodology |
SPARC's 5-phase model maps to BHIL's artifact chain (Spec=PRD+SPEC, Arch=ADR, Code=TASK) |
prd2build |
Generates docs from PRD; BHIL provides the PRD template and traceability |
wardley-maps |
Strategic analysis feeds into PRD and ADR context |
lazy-fetch |
Context management complements BHIL's context fragmentation prevention |
Quality Framework
- Specification quality is the primary quality lever, not test coverage alone
- Evaluation suites (EVAL-SUITE-TEMPLATE.yaml) define LLM quality thresholds
- Guardrails specs separate safety concerns from functional requirements
- Artifact validation scripts check frontmatter completeness and link integrity
- Retrospective feedback closes the loop from deployment back to requirements
Attribution
BHIL AI-First Development Toolkit by Barry Hurd (barryhurd.com). MIT License.