Develop Team

Develop Team

Orchestrate a full feature development cycle using parallel sub-agents. Takes a ticket reference or freeform task and drives it through deep research, planning (with design/test/migration specs), implementation with per-phase test generation, review, and automatic PR creation — fully autonomous with no checkpoints unless a genuine blocker is detected.

When to Use This Skill

• Starting a new feature - Hand off a ticket and let the pipeline research, plan, and implement
• Exploring before implementing - Use plan-only=true to get a researched plan without writing code
• Structured development - Parallel research + review instead of ad-hoc implementation

Parameters

Invocation

/develop-team PROJ-123
/develop-team PROJ-123 plan-only=true
/develop-team "Add currency preferences to user settings"
/develop-team PROJ-45 auto-commit=false skip-review=true
/develop-team PROJ-67 type=ui-heavy
/develop-team PROJ-89 type=backend-only skip-tests=true skip-pr=true

Prerequisites

This skill expects the following project conventions:

1. CLAUDE.md at project root — defines coding standards, patterns, folder structure, and conventions
2. Ticket system (optional) — Jira MCP or similar for fetching ticket details
3. Spec/backlog location — configurable; defaults to searching common locations (/backlog/, /specs/, /docs/specs/)
4. Companion skills (optional) — /refactor and /review-fix for Phases 5-6. If not available, those phases are skipped gracefully.

Architecture

ORCHESTRATOR (main session - manages state, cross-references outputs)
|
+-- Phase 0: INIT — parse input, create state, determine ticket vs freeform
|
+-- Phase 1: CONTEXT + CLASSIFY + DB PRE-SCAN
|   Ticket (MCP or manual) + spec file + git state
|   Past-ticket context (git log, gh pr list)
|   DB schema pre-scan (if DB tools available, scan related tables)
|   Classify: ui-heavy | full-stack | backend-only | async-heavy
|
+-- Phase 2: RESEARCH (4-6 parallel agents, run_in_background: true)
|   Agent 1: Deep Codebase Explorer    (system-architect)          [always]
|   Agent 2: CLAUDE.md Compliance Mapper (code-refactorer)         [always]
|   Agent 3: Requirements Mapper       (product-strategy-advisor)  [always]
|   Agent 4: Design System Integrator  (premium-ux-designer)       [if ui-heavy|full-stack]
|   Agent 5: Async Flow Analyzer       (system-architect)          [if async-heavy]
|   Agent 6: DB Migration Analyzer     (system-architect)          [if DB changes needed]
|   (auto-proceed — pause only on blockers)
|
+-- Phase 3: PLANNING — synthesize into plan + design spec + test plan + migration plan
|   (auto-proceed — plan-only=true pauses here)
|
+-- Phase 4: IMPLEMENTATION — per-phase: agent spawn → build/lint → tests → commit
|   UI phases → premium-ux-designer | Backend → general-purpose
|
+-- Phase 4.5: DB MIGRATIONS — apply via migration tooling, save locally
+-- Phase 5: REFACTOR → /refactor scope=session max-iterations=1 auto-fix=true
+-- Phase 6: REVIEW → /review-fix max-iterations=2 auto-commit={inherited}
+-- Phase 7: PR CREATION → gh pr create with full summary
+-- Phase 8: SYNTHESIS — report with commits, tests, migrations, PR link
|
+-- STATE: .develop-team/state.json

Workflow

Phase 0: Initialize

1. Parse $ARGUMENTS: ticket reference (pattern: {PROJECT_KEY}-\d+, e.g. PROJ-123) or freeform task description
2. Parse parameters: type, auto-commit, skip-review, skip-tests, skip-pr, skip-migrations, plan-only
3. Create .develop-team/ directory if needed
4. Check for existing state.json — if found, offer to resume from last completed phase
5. Initialize state file

State file at .develop-team/state.json: See references/state-schema.md for full schema and backward compatibility notes.

Phase 1: Gather Context + Classify Feature Type

If ticket reference:

Fetch ticket details using the project's configured ticket system (Jira MCP, GitHub Issues, Linear, etc.):

Extract: key, summary, description, acceptance criteria,
         status, issue type, priority, labels

Find spec file:

Search common spec/backlog locations for files matching ticket number or feature name. Typical locations:

• /backlog/ (organized by version or priority)
• /specs/
• /docs/specs/
• Custom location defined in CLAUDE.md

Read spec for requirements, user stories, success criteria.

Past-Ticket Context

Load context from previous related work to avoid re-discovering solved problems:

git log --oneline --all --grep="{ticketKey}" -20
gh pr list --state all --search "{ticketKey}" --json number,title,state,mergedAt --limit 5

Store summary of past PRs and patterns learned in state.json.pastTicketContext.

DB Schema Pre-Scan

If ticket mentions database tables/schema/data model, query the project's database tooling for schema information: table names, columns, policies, indexes. Feeds into Agent 6.

Git state:

git branch --show-current && git status --short && git log --oneline -5

Branch Creation (ticket references only)

When input is a ticket reference, create a feature branch off the current branch:

# Extract ticket number and kebab-case summary
# PROJ-111 + "Add currency preferences" → 111-add-currency-preferences
git checkout -b {ticketNumber}-{kebab-case-summary}

Store branch name in state.json.branchName. Skip for freeform tasks (user is already on their working branch).

Classify Feature Type

If type != auto, use directly. Otherwise, scan combined ticket text for keyword signals:

Logic: UI > 2x backend → ui-heavy. Backend > 2x UI and UI=0 → backend-only. Async >= 2 → async-heavy. Else → full-stack.

Phase 2: Research (4-6 Parallel Sub-Agents)

Spawn all applicable agents with run_in_background: true. Each receives ticket context, spec file, and past-ticket context.

For full agent prompts, read references/agent-prompts.md.

Post-Research: Auto-Proceed with Blocker Detection

Collect results via TaskOutput. Auto-proceed UNLESS a blocker is detected.

Blockers (pause and ask user only if):

• Research agents recommend incompatible approaches
• Feature requires tables with no migration plan to create them
• Acceptance criteria are ambiguous or contradictory
• A research agent completely failed (total failure, not just partial)

If no blockers, log summary and proceed directly:

Research complete ({N} agents, 0 blockers)
  Codebase Explorer: {N} files, {N} flow chains | Compliance Mapper: {N} rule cards
  Requirements Mapper: {N} phases, {N} tasks | Design/Async/Migration: {status}
Proceeding to planning...

Phase 3: Planning (Inline)

The orchestrator synthesizes all research into a cohesive plan. Done inline (not sub-agent) to cross-reference multiple outputs.

Cross-reference: Codebase Explorer's file map + Compliance Mapper's rule cards + Requirements Mapper's phases + Design System Integrator's specs (if ran) + Async Analyzer (if ran) + DB Migration Analyzer (if ran).

Plan structure: Save to .develop-team/plan.md. See references/plan-template.md for the full template. Includes: context, per-phase tasks with convention notes, design specification (if UI), test plan, migration plan, AC coverage, and files summary.

Save design spec to state.json.designSpec. If plan-only=true, present plan and stop. Otherwise auto-proceed.

Phase 4: Implementation

Execute plan phase by phase:

1. Select agent:

2. Spawn agent (run_in_background: false):

For UI phases → premium-ux-designer:

Implement Phase {N}. Follow the design specification EXACTLY — do not improvise colors/shadows/spacing.

## Phase Tasks: {tasks from plan}
## Design Specification: {full design spec — hierarchy, tokens, states, responsive}
## Convention Rules: {component rule card from Compliance Mapper}
## Anti-Slop Checklist: {from Design System Integrator}

Instructions: Implement tasks in order. Run build and lint commands after. Do NOT implement other phases.

For non-UI phases → general-purpose:

Implement Phase {N}.

## Phase Tasks: {tasks from plan}
## Convention Rules: {relevant rule cards — api-route, service, schema, store}
## Async Analysis: {if applicable, from Async Flow Analyzer}

Instructions: Implement tasks in order. Run build and lint commands after. Do NOT implement other phases.

3. Verify: Run build and lint commands (per CLAUDE.md configuration). If fails: spawn debug agent. If still fails after 2 attempts, stop.

4. Generate tests (unless skip-tests=true): Spawn general-purpose with test cases from the plan for this phase. Run test command. Update state.json.tests.perPhase.

5. Commit (if auto-commit=true): create atomic commit with phase name and files list.

6. Update state and proceed to next phase.

Phase 4.5: DB Migrations (Conditional)

Skip if: skip-migrations=true OR no migrations planned.

For each migration in the plan:

1. Apply: Use the project's migration tooling (Supabase MCP, Prisma, Drizzle, raw SQL, etc.)
2. Save locally (CRITICAL): Save migration file to the project's migrations directory following existing naming conventions
3. Update related files: Edit any select/query builder files, ORM schemas, or type definitions as specified in the plan
4. Commit if auto-commit enabled

On migration failure: Save SQL to .develop-team/failed-migration.sql, ask user: skip or stop.

Phase 5: Refactor (Companion Skill Delegation)

Skip if skip-review=true or /refactor skill not available. Invoke: /refactor scope=session max-iterations=1 auto-fix=true

After: read .refactor/state.json, update state, clean up .refactor/.

Phase 6: Review (Companion Skill Delegation)

Skip if skip-review=true or /review-fix skill not available. Invoke: /review-fix max-iterations=2 auto-commit={inherited}

After: read .review-fix/state.json, update state, clean up .review-fix/.

Phase 7: PR Creation

Skip if: skip-pr=true.

gh pr create --title "{type}({scope}): {summary} ({ticket key})" --body "$(cat <<'EOF'
## Summary
{1-3 bullets from ticket}

## Changes
{Phases implemented with key files}

## Test Plan
| Phase | Tests | Passed | Failed |
{from state.json.tests.perPhase}
- [ ] Build passes
- [ ] Lint passes

## Database Migrations
{List of migrations applied, or "None"}

## Acceptance Criteria
| Criterion | Status |
{from plan}

## Ticket
{ticket key}: {URL if available}

---
Generated with [Claude Code](https://claude.com/claude-code) via /develop-team
EOF
)"

Update state.json.pullRequest. On failure: save body to .develop-team/pr-body.md, note in synthesis.

Phase 8: Synthesis

# Development Complete: {Ticket Key} - {Summary}

## Summary
| Metric | Value |
|--------|-------|
| Feature type | {featureType} |
| Phases implemented | {N} |
| Total commits | {N} |
| Tests | {N} generated ({N} passed, {N} failed) |
| Migrations | {N} applied |
| Refactor | {N} fixed, {N} strategic |
| Review | {N} iterations, {N} fixed |
| Pull request | {PR URL or "Skipped"} |

## Commits — {all with hashes}
## Acceptance Criteria — {criterion → status}
## Test Summary — {per-phase results}
## Database Migrations — {migration → status → local file}
## Files Changed — {action → path}
## Next Steps — {merge PR, remaining items, manual testing}

Update state.json with status: "completed".

State File Reference

Location: .develop-team/state.json — enables resumability.

If existing state found, ask user: "Resume from Phase {N+1}" or "Start fresh".

Error Handling

Integration Points

• Ticket system: Jira MCP, GitHub Issues, Linear, or manual input
• Database tooling: Supabase MCP, Prisma, Drizzle, or raw SQL — auto-detected from project setup
• GitHub CLI: gh pr create for PRs, gh pr list for past-ticket context
• Spec/backlog files: Searched in configurable locations (see Prerequisites)
• CLAUDE.md: Per-file-type rule cards extracted and injected into all agents
• Design system: Theme config, global styles, and design docs read by Agent 4 (paths auto-detected from project)
• Companion skills: /refactor and /review-fix delegated to in Phases 5-6 (gracefully skipped if unavailable)

Tips

1. plan-only=true for complex features — review before committing to implementation
2. Provide a ticket reference when possible — acceptance criteria drive better research and validation
3. Override with type= to specify the feature type directly — ensures correct agent routing
4. Check .develop-team/state.json if interrupted — resumable from last completed phase
5. skip-review=true for quick iterations — run /refactor and /review-fix manually later
6. skip-tests=true for prototypes — skip per-phase test generation
7. skip-pr=true when not ready — create PR manually from synthesis report
8. skip-migrations=true for code-only changes
9. Zero checkpoints — fully autonomous, pauses only on genuine blockers