hdb-design

name: hdb:design description: Design a new software feature with a PRD and detailed implementation task list

hdb:design

Usage

/hdb:design <feature description>

Description

Designs a new software feature in five phases (triage → understand → PRD → tasks → write), grounded in the project's actual codebase. Produces two deliverables written to docs/design/<feature-slug>-{prd,tasks}.md:

1. PRD — overview, goals, acceptance criteria (as test descriptions), technical decisions, design, test strategy, rollback, and implementation stages.
2. Task List — test-first steps per stage: tests, code, verification commands, risks.

Instructions

Phase 0: Triage

1. Classify scope and state reasoning:

   • Small (1 file, <50 lines, no new patterns): task list only → skip to Phase 3.
   • Medium (2–5 files, existing patterns): PRD + task list → all phases.
   • Large (multiple subsystems, new patterns): PRD + task list. Consider splitting into sub-designs.
   • Ambiguous (no concrete problem/solution): clarify in Phase 1, then re-classify.

   The feature description is the scope — no tangential improvements or unnecessary abstractions.

Phase 1: Understand

2. Clarify (if needed) using AskUserQuestion — only when the description lacks a concrete what or how. At most 3–4 questions: problem/user, constraints, out-of-scope, prior art. If clear, skip to exploration.

3. Explore the codebase with parallel tool calls. Every technical decision in the PRD must cite a specific file or pattern discovered here:

   • Glob("**/*.{go,rs,py,ts}") → project structure and file layout
   • Read build/config files (go.mod, Cargo.toml, package.json) → tech stack and dependencies
   • Grep for error handling, test patterns, naming conventions → coding standards
   • Grep for feature-related keywords → related existing code
   • For unfamiliar codebases: use Agent subagents to explore multiple areas concurrently

4. Identify integration points. For each, record: the file path, the function/type involved, and the convention new code must follow (naming, error handling, test style). Flag novel patterns (no codebase precedent) for Phase 2 decisions. Done when every file and convention is named.

Phase 2: PRD

5. Write the PRD:

   • Overview: What the feature does and the problem it solves.
   • Goals / Non-Goals: Bulleted. Non-goals prevent scope creep.
   • Acceptance Criteria (as test descriptions): TestName: When X, assert Y. If you can't write a criterion as a test, the design is too vague.
   • Technical Decisions: Cite the codebase file/pattern justifying each choice. For novel patterns: present ≥2 options with trade-offs, recommend one, flag for developer. For unknowable runtime behavior: recommend a time-boxed spike.
   • Design and Operation: User perspective, system perspective (data flow, concurrency, state), error handling, edge cases.
   • Test Strategy: Levels (unit/integration/e2e), infrastructure (fixtures, mocks, temp dirs), key scenarios.
   • Rollback and Safety: Can the feature be disabled without data loss? One sentence for trivial features.
   • Implementation Stages: Ordered. Each stage: working system, ≤5 files, deliverable verifiable in under a minute.

6. Present PRD to the user. Wait for explicit approval before proceeding.

Phase 3: Task List

7. Build test-first tasks per stage. Be specific — name files, functions, types:

   • Tests: functions, assertions, fixtures — always first.
   • Code: minimum to pass tests (files, functions, config, deps, migrations, API surface as applicable).
   • Verify: exact test command.
   • Risks: known blockers.

8. Order tasks top-to-bottom, no forward dependencies. Note parallelizable tasks.

9. Validate coverage: name the task covering each acceptance criterion. Add tasks for gaps.

10. Present task list to the user. Incorporate feedback.

Phase 4: Write Files

11. Write files with Write tool to docs/design/ (or ask user). Offer next steps: implement stage 1, commit docs, or refine.

Implementation Protocol

Per task: (1) write tests (compile but fail), (2) confirm correct failures (add stubs if needed), (3) implement minimum to pass, (4) run tests (green = done; red = fix implementation), (5) lint/format, (6) next task.

No untested code — test or delete. Tests are the specification. Never weaken assertions.

Guidelines

• Ground decisions in existing codebase patterns — cite specific files.
• Isolate risk — put unknowns and external dependencies in early stages so later work isn't blocked.
• Be specific — name files, functions, types. Vague tasks waste time.
• Flag trade-offs — present options with pros/cons for the developer.
• Adapt to scope — small: 1 stage, no PRD. Large: 5+ stages, consider sub-designs.

Example

User: /hdb:design webhook notifications for review completion

Triage: Medium — config, storage, worker, CLI (4 areas), existing patterns.

Clarify: "Multiple URLs per repo? Retry logic? Configurable payload format?"

Explore: Glob("**/*.go") maps structure (12 packages). Read on go.mod (no HTTP client deps → stdlib is sufficient), internal/config/config.go (uses [[repos]] TOML array pattern). Grep("ReviewComplete\|jobDone") finds the hook in internal/worker/run.go:142 and state tracking in internal/storage/jobs.go:89.

PRD (written to webhook-prd.md):

• Overview: HTTP POST on review completion for external tools.
• Goals: Reliable delivery, multiple URLs, verdict in payload. Non-goals: payload transforms, auth beyond shared secret, webhook UI.
• Acceptance Criteria: TestWebhookDelivery_SendsPostOnComplete (POST within 5s), TestWebhookDelivery_NoConfigNoSend (no URL → no request), TestWebhookConfig_ParsesTOML, TestWebhookPayload_IncludesVerdict.
• Technical decisions: stdlib net/http (no HTTP client deps in go.mod — adding one is unjustified for simple POST requests), TOML config with [[webhooks]] (matches existing [[repos]] pattern in config.go), SQLite webhook_deliveries table (follows existing internal/storage/jobs.go table pattern).
• Design: Job done → enqueue per URL → goroutine POSTs and records result.
• Stages: (1) Config + storage, (2) Delivery engine, (3) Worker integration, (4) CLI commands.

User feedback: "Skip retry for v1." → removed retry section from PRD, simplified stages.

Task list (Stage 1):

Task 1.1: Config parsing

• Tests: TestConfig_WebhookSection (TOML → structs), TestConfig_NoWebhooks (→ empty slice)
• Code: WebhookConfig type + [[webhooks]] in internal/config/config.go
• Verify: go test ./internal/config/...

Task 1.2: Delivery storage

• Tests: TestInsertWebhookDelivery (insert + read back), TestListWebhookDeliveries (count + order)
• Code: Migration in migrations.go, methods in webhooks.go
• Verify: go test ./internal/storage/...
• Risks: File-backed SQLite for WAL — use t.TempDir().

Validate: TestWebhookConfig_ParsesTOML → Task 1.1, TestWebhookDelivery_SendsPostOnComplete → Task 2.1. All covered. Write to docs/design/. Next: implement stage 1.