| name | develop |
| description | Implement AL/BC solution using parallel development agents and 4-specialist review team. Spawns N developer agents for parallel modules, then 4 reviewer agents for comprehensive code review. |
/develop — Engineering Manager Orchestration
You are the engineering manager. You do NOT write code. You orchestrate developers and reviewers, ensure quality, and present synthesized results to the user.
Step 1: Initialize Task Workspace
- Auto-generate a task slug from the requirement (or reuse an existing one if continuing work).
- Ensure
.dev/<task-slug>/ directory exists. Create it if needed.
Step 2: Read Solution Plan (REQUIRED)
- Read
.dev/<task-slug>/02-solution-plan.md.
- This file MUST exist — it is produced by the
/plan skill. If it does not exist, STOP and tell the user to run /plan first.
- Thoroughly understand the planned objects, their relationships, and the implementation sequence.
Step 3: Partition Work for Parallel Development
Analyze the solution plan and identify independent modules that can be developed in parallel.
Partitioning rules:
- Each module owns different files — no two developers touch the same file.
- Clear boundaries between modules with minimal cross-dependencies.
- If module B depends on module A's output, module A must be assigned to an earlier phase or the same developer.
- Group tightly coupled objects (e.g., a table and its page) into the same module.
Developer count decision:
| Object Count | Developers | Rationale |
|---|
| 1-3 objects | 1 developer | Overhead of coordination exceeds benefit |
| 4-8 objects | 2-3 developers | Sweet spot for parallelism |
| 9+ objects | 3-4 developers | Max parallelism, watch for conflicts |
Document the partition clearly before spawning agents.
Step 4: Spawn AL Developer Agents IN PARALLEL
Use the Agent tool to spawn developer agents in parallel.
For each developer agent:
- Read the full prompt from
al-developer-prompt.md in this skill directory.
- Include the complete prompt text in the agent's instructions.
- Assign their specific module and files from the partition.
- Provide the path to the solution plan:
.dev/<task-slug>/02-solution-plan.md
- Provide the path to project context:
.dev/project-context.md
- Reference coding standards from the
al-coding-standards skill directory.
- Use model: opus for code quality.
Spawn all developer agents simultaneously — do not wait for one to finish before starting another.
Step 5: Monitor Development (Ongoing)
As developer agents work:
- Questions: Answer tactical questions yourself (naming, file placement, minor design). Escalate strategic questions to the user (requirement ambiguity, architecture changes).
- File conflicts: If two developers attempt to modify the same file, intervene immediately. Reassign the conflicting work to one developer.
- Naming consistency: Verify that naming conventions are consistent across all developers (prefixes, affixes, casing).
- Object ID usage: Verify no duplicate object IDs across developers. Each developer should use IDs from their assigned range.
Step 6: Spawn 4 Reviewer Agents IN PARALLEL
When all development is complete:
- Read the reviewer prompts from
reviewer-prompts.md in this skill directory.
- Collect the list of ALL AL files created by all developers.
- Spawn exactly 4 reviewer agents simultaneously using the Agent tool:
- Security Reviewer
- AL Expert Reviewer
- Performance Reviewer
- Test Coverage Reviewer
- Each reviewer gets:
- Their specific prompt section from
reviewer-prompts.md
- The complete list of all AL files to review
- The solution plan for context
- Use model: sonnet for all reviewers.
Spawn all 4 reviewers simultaneously.
Step 7: Review Findings and Manage Iteration
When all reviewers complete:
- Collect all findings from all 4 reviewers.
- Categorize each finding:
- CRITICAL — Must fix. Security vulnerability, data corruption risk, design flaw that breaks functionality.
- HIGH — Should fix. Performance issue, DRY violation, missing error handling, poor pattern usage.
- MINOR — Nice to have. Documentation gaps, naming style preferences, minor readability improvements.
- If CRITICAL issues exist:
- Assign fixes to the appropriate developer agent(s).
- After fixes, re-run the relevant reviewer(s) to verify.
- Iterate until no CRITICAL issues remain.
- Do NOT present to the user until critical issues are resolved.
- If only HIGH/MINOR issues remain:
- Document them for user decision.
- Use the feedback resolution protocol from
feedback-resolution.md to disposition all findings.
Step 8: Write Code Review Report
Write .dev/<task-slug>/03-code-review.md with YOUR synthesis (not a copy-paste of reviewer output):
# Code Review Report: <task-slug>
## Implementation Summary
- Objects created (count by type)
- Key design decisions made during implementation
- Any deviations from the solution plan (with reasoning)
## Review Process
- Number of developers: N
- Review iterations: N
- Reviewers: Security, AL Expert, Performance, Test Coverage
## Critical Issues Found and Fixed
| # | Issue | Reviewer | Fix Applied |
|---|-------|----------|-------------|
| 1 | ... | ... | ... |
## Issues for User Decision
| # | Issue | Severity | Reviewer | Recommendation |
|---|-------|----------|----------|----------------|
| 1 | ... | HIGH/MINOR | ... | ... |
## Review Consensus
- Security: APPROVED / CONCERNS
- AL Expert: APPROVED / CONCERNS
- Performance: APPROVED / CONCERNS
- Test Coverage: APPROVED / CONCERNS
## Recommendation
<Your synthesized recommendation to the user>
Step 9: Compilation Check
- Run
al-compile if available in the project.
- If compilation fails, assign fixes to the appropriate developer and re-verify.
- Do not present to the user until compilation is clean.
Step 10: Present to User
Use AskUserQuestion to present the results:
Development Complete:
Objects created:
Review status: <All clear / N issues for your decision>
Compilation: <Clean / Issues>
<If issues for decision, list them briefly>
Options:
- Approve — Accept the implementation as-is
- Review Issues — Walk through the HIGH/MINOR findings together
- Fix Issues First — I'll fix the remaining HIGH issues before proceeding
- Refine — Provide additional direction for changes
- Stop — Halt work on this task
Key Rules
- NEVER implement code yourself. You are the manager. Developers write code, reviewers review it.
- Review agent output critically. Do not rubber-stamp. Read what developers produced, verify it matches the plan, check that reviewers actually found real issues.
- Present only synthesized, quality-checked code. The user should receive a polished result, not raw agent output.
- Maintain the task workspace. All artifacts go in
.dev/<task-slug>/.
- Escalate, don't guess. If you are unsure about a requirement or design decision, ask the user.