agentleFS
ulpi-ioulpi-io/browse/.agents/skills/run-parallel-agents-feature-build/SKILL.md

run-parallel-agents-feature-build

Skill15 starsChanged 8 months agoApache-2.0View on GitHub

Orchestrate multiple specialized agents working in parallel when building independent features. Use when the task list contains 3+ unrelated tasks that don't share state or files. Matches each task to the right expert agent and launches them concurrently via the Agent tool with worktree isolation. Supports all available subagent types.

The file

Show its rules and text

We split this skill into the rules it gives an agent

29 rules. About 0 are common, 25 are distinctive, and 4 are very specific to this project.

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈25

    Distinctive · estimate

    This author's own choices. Worth a look.

  • 4

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Distinctive:

    Before launching parallel agents, you ABSOLUTELY MUST:

    line 12
  2. Distinctive:

    Verify tasks are truly independent (no shared files, no data flow between them)

    line 14
  3. Distinctive:

    Create complete briefs with scope, files, output, and success criteria

    line 16
  4. Distinctive:

    Before launching ANY parallel agents, you MUST complete this checklist:

    line 28
  5. Distinctive:

    Use this skill automatically when:

    line 45
  6. Distinctive:

    No shared state or cross-dependencies between tasks

    line 53
  7. Distinctive:

    Do NOT use parallel agents when:

    line 73
  8. Distinctive:

    Shared state: - Do tasks modify the same files or data structures?

    line 103
  9. Distinctive:

    If all conditions for parallelization are met, proceed to Step 2.

    line 106
  10. Very specific:

    Use isolation: "worktree" — without this, parallel agents clobber each other's files

    line 184
  11. Very specific:

    Use run_in_background: true — without this, agents run sequentially

    line 185
  12. Distinctive:

    No placeholder values — all parameters must be complete

    line 187
  13. Very specific:

    After each task agent completes and before merging its worktree, check the task's review field from the plan JSON.

    line 220
  14. Distinctive:

    Review failures: If the review tool finds P0/P1 issues, fix them before merging the worktree and before launching dependent tasks.

    line 233
  15. Distinctive:

    After aggregating results, verify the parallel execution was successful:

    line 237
  16. Distinctive:

    Gate: Do NOT mark task complete until all 5 checks pass.

    line 259
  17. Distinctive:

    Launch: all three in a single message with Agent tool

    line 275
  18. Distinctive:

    If agents modify overlapping files:

    line 361
  19. Distinctive:

    Always inform the user BEFORE launching parallel agents:

    line 370
  20. Distinctive:

    If an agent fails or gets blocked:

    line 383
  21. Distinctive:

    Don't retry automatically without user input

    line 388
  22. Distinctive:

    0 points: No briefs or generic briefs

    line 407
  23. Distinctive:

    0 points: No summary provided

    line 417
  24. Distinctive:

    Fix: Run dependency checks. If Output(A) ∈ Input(B), execute sequentially.

    line 447
  25. Very specific:

    Fix: Use technology detection patterns.

    line 452
  26. Distinctive:

    Fix: ALL Agent tool calls in a SINGLE message.

    line 457
  27. Distinctive:

    Fix: Every brief must have Scope, Files, Output, Success criteria.

    line 462
  28. Distinctive:

    Fix: After aggregation, check file overlap.

    line 467
  29. Distinctive:

    When parallel execution is complete, announce:

    line 510

≈ “Common” is an estimate: at least 1 in 5 of the 350,495 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md
<EXTREMELY-IMPORTANT> Before launching parallel agents, you **ABSOLUTELY MUST**:
  1. Verify tasks are truly independent (no shared files, no data flow between them)
  2. Match each task to the correct specialized agent
  3. Create complete briefs with scope, files, output, and success criteria
  4. Launch ALL agents in a SINGLE message with multiple Agent tool calls

Launching agents without verification = conflicts, wasted work, merge hell

This is not optional. Parallel execution requires independence verification. </EXTREMELY-IMPORTANT>

Run Parallel Agents Feature Build

MANDATORY FIRST RESPONSE PROTOCOL

Before launching ANY parallel agents, you MUST complete this checklist:

  1. ☐ Count tasks — are there 3+ independent tasks?
  2. ☐ Check dependencies — do any tasks depend on others?
  3. ☐ Check file overlap — do tasks modify the same files?
  4. ☐ Match agents — assign correct agent type per task
  5. ☐ Prepare briefs — scope, files, output, success criteria for each
  6. ☐ Announce to user — "Launching N agents in parallel for [tasks]"

Launching agents WITHOUT completing this checklist = conflicts and failures.

Overview

Automatically detect opportunities for parallel execution and orchestrate multiple specialized agents working concurrently on independent features, modules, or investigations. Match each task to the appropriate domain expert and coordinate their work to deliver faster results without sacrificing quality.

When to Use This Skill

Use this skill automatically when:

Task List Indicators:

  • The task list contains 3 or more independent tasks or phases
  • Receiving a large plan with multiple unrelated features
  • Tasks are clearly scoped and don't overlap
  • Each task can be understood without context from others
  • No shared state or cross-dependencies between tasks

User Triggers:

  • "Build in parallel"
  • "Split this work across agents"
  • "Use multiple agents for this"
  • "Speed this up with parallel execution"

Common Scenarios:

  • Building independent features or endpoints (e.g., wishlist API, checkout flow, user dashboard)
  • Implementing multiple UI components or microservices simultaneously
  • Generating documentation or schema files for several modules
  • Analyzing multiple code files for specific issues
  • Debugging failures in isolated subsystems
  • Processing multiple data files or resources

When NOT to Use This Skill

Do NOT use parallel agents when:

Dependencies Exist:

  • Tasks have sequential dependencies (Task B needs Task A's output)
  • Shared state, data, or ownership between tasks
  • The sequence of execution affects the outcome
  • Agents would interfere with each other's work

Problem Not Decomposed:

  • The problem isn't yet broken into independent units
  • Need to understand full system state first
  • Related failures where fixing one might fix others
  • Careful coordination or debugging across layers is required

Single Cohesive Task:

  • Working on a single integrated feature
  • Refactoring that touches multiple interconnected parts
  • Tasks that require constant communication between components

Core Workflow

Step 1: Analyze the Task List or Request

Examine the current task list or user request to identify:

  1. Number of independent work streams - Are there 3+ separate features/tasks?
  2. Dependencies - Can each task be completed without waiting for others?
  3. Shared state - Do tasks modify the same files or data structures?
  4. Scope clarity - Is each task clearly defined with known requirements?

If all conditions for parallelization are met, proceed to Step 2.

Step 2: Match Features to Specialized Agents

For each independent task, determine the best agent type based on:

Technology Stack Detection:

  • Laravel backend → laravel-senior-engineer
  • Next.js frontend or full-stack → nextjs-senior-engineer
  • React + Vite + Tailwind frontends → react-vite-tailwind-engineer
  • Express.js APIs → express-senior-engineer
  • Node.js CLI tools → nodejs-cli-senior-engineer
  • Python backend, Django, data pipelines → python-senior-engineer
  • FastAPI specifically, async DB, JWT auth → fastapi-senior-engineer
  • Go backend, services → go-senior-engineer
  • Go CLI tools → go-cli-senior-engineer
  • iOS/macOS, Swift, SwiftUI → ios-macos-senior-engineer
  • Expo React Native mobile → expo-react-native-engineer
  • AWS infrastructure, CDK → devops-aws-senior-engineer
  • Docker, containers → devops-docker-senior-engineer
  • General tasks (exploration, research) → general-purpose

File Pattern Analysis:

*.php + /app/ + /routes/ → Laravel
*.tsx + /app/ or /pages/ + next.config.* → Next.js
*.tsx + vite.config.* + tailwind.config.* → React/Vite/Tailwind
*.ts + express imports → Express
*.ts + commander/inquirer imports → Node CLI
*.py + fastapi imports → FastAPI
*.py + django/flask/general → Python
*.go + go.mod → Go
*.swift + Package.swift or *.xcodeproj → iOS/macOS
*.tsx + app.json (Expo) → Expo React Native
Dockerfile + docker-compose.* → Docker
CDK, CloudFormation, Terraform → AWS DevOps

Step 3: Prepare Agent Briefs

For each agent, create a clear, focused brief containing:

Required Elements:

  1. Scope of work - Exactly what to build/analyze/fix
  2. Expected output - What deliverables to produce
  3. Context - Relevant file paths, existing patterns, constraints
  4. Success criteria - How to verify completion

Brief Template:

Build [feature name]:
- Scope: [specific feature boundaries]
- Files: [relevant paths or patterns]
- Requirements: [bullet points]
- Output: [expected deliverables]
- Patterns: [existing code patterns to follow]

Step 4: Launch Agents in Parallel

Execute all agents simultaneously using a single message with multiple Agent tool calls:

Use the Agent tool with these parameters for each task:
- prompt: Complete brief (scope, files, output, success criteria)
- subagent_type: Matched agent from the Agent Table
- isolation: "worktree" (MANDATORY for parallel file modifications)
- run_in_background: true (for concurrent execution)
- model: "opus" for complex tasks, "sonnet" for straightforward tasks

Critical Requirements:

  • Send ALL Agent tool calls in ONE message (this is what makes them parallel)
  • Use isolation: "worktree" — without this, parallel agents clobber each other's files
  • Use run_in_background: true — without this, agents run sequentially
  • Each agent gets its complete brief in the prompt parameter
  • No placeholder values — all parameters must be complete
  • For verifying built features in the browser (e.g., testing a UI component the agent just created), use browse --session <agent-name> for session isolation

Step 5: Monitor and Aggregate Results

After agents complete:

  1. Collect outputs from each agent
  2. Verify deliverables match expected outputs
  3. Check for conflicts (e.g., overlapping file changes)
  4. Merge results into a coherent summary
  5. Report to user with consolidated findings

Aggregation Template:

Parallel execution complete. Results:

**[Feature 1]** (via [agent-type])
- Status: [completed/blocked/partial]
- Delivered: [summary]
- Files modified: [list]

**[Feature 2]** (via [agent-type])
- Status: [completed/blocked/partial]
- Delivered: [summary]
- Files modified: [list]

**Overall:** [X/Y features completed, any conflicts, next steps]

Step 5b: Review Dispatch (Mandatory When Plan Specifies Review)

After each task agent completes and before merging its worktree, check the task's review field from the plan JSON. If it is not none, invoke the corresponding review skill using the Skill tool:

review valueSkill invocationWhat it runs
codexSkill("codex-review")OpenAI Codex CLI (codex review) — external binary, independent AI
claudeSkill("claude-review")Separate Claude Code agent in isolated worktree
kiroSkill("kiro-review")Kiro CLI (kiro-cli chat) — external binary, independent AI
noneSkip reviewNo review needed

CRITICAL: These are real external CLI tools, not prompts for a general-purpose agent. Do NOT substitute a Claude agent with "review this diff" in the prompt — that is a self-review, not an independent review. The entire point of specifying codex or kiro is to get a second opinion from a different AI system.

Review timing: Reviews can run in parallel with other task agents that don't depend on the reviewed task. For example, if TASK-001 and TASK-002 are both in Layer 0, you can review TASK-001 while TASK-002's agent is still running.

Review failures: If the review tool finds P0/P1 issues, fix them before merging the worktree and before launching dependent tasks. P2/P3 issues can be noted and fixed during merge.

Step 6: Verification (MANDATORY)

After aggregating results, verify the parallel execution was successful:

Check 1: All Agents Completed

  • Every launched agent returned a result
  • No agents timed out or crashed

Check 2: No File Conflicts

  • No two agents modified the same file
  • If conflicts exist, they are resolved

Check 3: Deliverables Match Briefs

  • Each agent delivered what was specified
  • Output matches expected format

Check 4: No Integration Issues

  • Features work together (if they interact)
  • No broken imports or references

Check 5: User Informed

  • Completion announced with summary
  • Any issues or partial results documented

Gate: Do NOT mark task complete until all 5 checks pass.

Example Scenarios

Example A: Feature Build

User Says: "Build the wishlist API, checkout summary, and user dashboard in parallel."

Execution:

  1. Detect three separate feature scopes
  2. Match agents:
    • Wishlist API → laravel-senior-engineer (backend)
    • Checkout summary → nextjs-senior-engineer (frontend)
    • User dashboard → nextjs-senior-engineer (frontend/UI)
  3. Launch all three in a single message with Agent tool
  4. Aggregate results into merged summary

Output:

Built 3 features in parallel:
- Wishlist API: Complete (app/Http/Controllers/WishlistController.php, routes/api.php)
- Checkout summary: Complete (app/checkout/summary/page.tsx)
- User dashboard: Complete (app/dashboard/page.tsx)

Example B: Debug Parallel Subsystems

User Says: "Run parallel agents to debug these failing tests."

Execution:

  1. Cluster failures by subsystem:
    • Laravel backend tests
    • Next.js frontend tests
    • Node.js API tests
  2. Spawn agents:
    • laravel-senior-engineer for Laravel failures
    • nextjs-senior-engineer for Next.js failures
    • express-senior-engineer for Node API failures
  3. Let each diagnose and fix independently
  4. Gather results, merge patches, re-run tests
  5. Output consolidated fix report

Example C: Code Analysis

User Says: "Analyze these 5 code files in parallel for performance bottlenecks."

Execution:

  1. Split files across appropriate agents based on file type
  2. Run analysis concurrently (each agent gets 1-2 files)
  3. Merge findings into summarized report with:
    • File-by-file breakdown
    • Common patterns across files
    • Prioritized recommendations

Agent Type Reference

Quick reference for matching tasks to agents:

Agent TypeBest ForKey Indicators
laravel-senior-engineerLaravel backends, APIs, Eloquent models*.php, /app/, Eloquent, Artisan
nextjs-senior-engineerNext.js apps, React Server Components, App Router*.tsx, /app/, /pages/, next.config.*
react-vite-tailwind-engineerReact + Vite + Tailwind TypeScript frontends*.tsx, vite.config.*, tailwind.config.*
express-senior-engineerExpress.js APIs, middleware, REST endpoints*.js/*.ts, express imports
nodejs-cli-senior-engineerNode.js CLI tools, commander.js, inquirercommander, inquirer, ora, CLI patterns
python-senior-engineerPython backends, Django, data pipelines*.py, requirements.txt, pyproject.toml
fastapi-senior-engineerFastAPI, async DB, JWT auth*.py, fastapi imports, uvicorn
go-senior-engineerGo backends, services, APIs*.go, go.mod, go.sum
go-cli-senior-engineerGo CLI tools, cobra, viper*.go, cobra imports, CLI patterns
ios-macos-senior-engineerSwift, SwiftUI, Xcode, SPM, StoreKit*.swift, Package.swift, *.xcodeproj
expo-react-native-engineerExpo mobile apps, cross-platform*.tsx, app.json, Expo modules
devops-aws-senior-engineerAWS infrastructure, CDK, CloudFormationCDK, CloudFormation, Terraform, AWS
devops-docker-senior-engineerDocker, Docker Compose, containerizationDockerfile, docker-compose.*
general-purposeExploration, research, general tasksNon-framework-specific work

See references/agent_matching_logic.md for detailed matching rules and edge cases.

Best Practices

Scoping Tasks Effectively

Good Task Scopes:

  • "Build user authentication endpoint with JWT"
  • "Create product listing page with filters"
  • "Implement cart total calculation service"

Poor Task Scopes:

  • "Build the entire checkout flow" (too broad, likely has dependencies)
  • "Fix the app" (undefined, not decomposed)
  • "Refactor database layer" (touches too many interconnected parts)

Handling Conflicts

If agents modify overlapping files:

  1. Review changes from each agent
  2. Identify conflicts (same lines modified)
  3. Merge intelligently or ask user for guidance
  4. Re-test affected areas

Communication Pattern

Always inform the user BEFORE launching parallel agents:

I've identified 3 independent features that can be built in parallel:
1. [Feature 1] - using [agent-type]
2. [Feature 2] - using [agent-type]
3. [Feature 3] - using [agent-type]

Launching agents now...

Handling Partial Success

If an agent fails or gets blocked:

  • Continue with successful agents
  • Report partial results
  • Provide clear next steps for blocked work
  • Don't retry automatically without user input

Quality Checklist (Must Score 8/10)

Score yourself honestly before marking parallel execution complete:

Independence Verification (0-2 points)

  • 0 points: Launched without checking dependencies
  • 1 point: Checked some dependencies
  • 2 points: Full dependency check (files, data flow, state)

Agent Matching (0-2 points)

  • 0 points: Wrong agents for tasks
  • 1 point: Agents assigned without justification
  • 2 points: Each task matched to correct agent with indicators

Brief Quality (0-2 points)

  • 0 points: No briefs or generic briefs
  • 1 point: Partial briefs (missing fields)
  • 2 points: Complete briefs with scope, files, output, success criteria

Execution Correctness (0-2 points)

  • 0 points: Agents launched sequentially
  • 1 point: Parallel launch but incomplete briefs
  • 2 points: Single message with all Agent tool calls, complete briefs

Result Aggregation (0-2 points)

  • 0 points: No summary provided
  • 1 point: Partial summary
  • 2 points: Complete summary with status, deliverables, conflicts, next steps

Minimum passing score: 8/10


Common Rationalizations (All Wrong)

These are excuses. Don't fall for them:

  • "These tasks look independent" → STILL check for file overlap
  • "I know the right agents" → STILL document the matching rationale
  • "Briefs are obvious" → STILL write complete scope, files, output, success criteria
  • "Sequential is safer" → If tasks are independent, parallel is FASTER with no downside
  • "I'll merge conflicts later" → Check for conflicts BEFORE launching
  • "One message is too long" → ALL Agent tool calls MUST be in one message
  • "The user just wants it done" → Correct execution = faster completion
  • "Two tasks don't need parallel agents" → Correct, need 3+ tasks
  • "I'll just have a Claude agent review the diff, same thing as codex" → It is NOT the same thing. Codex and Kiro are independent AI systems with different training and different blind spots. Self-review catches fewer bugs than cross-model review. Use the tool specified in the plan.
  • "The review tool isn't installed" → WARN the user, don't silently substitute. which codex or which kiro-cli to check availability.

Failure Modes

Failure Mode 1: Launching Dependent Tasks in Parallel

Symptom: Agents fail or produce conflicting output because Task B needed Task A's result Fix: Run dependency checks. If Output(A) ∈ Input(B), execute sequentially.

Failure Mode 2: Wrong Agent Assignment

Symptom: Laravel work done by nextjs-senior-engineer, poor results Fix: Use technology detection patterns. Match *.php → Laravel agent.

Failure Mode 3: Sequential Launch Instead of Parallel

Symptom: Sent multiple messages with Agent tool calls, agents ran one at a time Fix: ALL Agent tool calls in a SINGLE message. This is critical for parallel execution.

Failure Mode 4: Incomplete Briefs

Symptom: Agents ask clarifying questions or deliver wrong output Fix: Every brief must have Scope, Files, Output, Success criteria. No placeholders.

Failure Mode 5: No Conflict Check After Completion

Symptom: Agents modified same files, changes overwrite each other Fix: After aggregation, check file overlap. If conflicts, merge intelligently.


Quick Workflow Summary

STEP 1: ANALYZE
├── Count tasks (need 3+)
├── Check dependencies
├── Check file overlap
└── Gate: Tasks are independent

STEP 2: MATCH AGENTS
├── Identify technology per task
├── Assign correct agent type
└── Document matching rationale

STEP 3: PREPARE BRIEFS
├── Scope, Files, Output, Success criteria
└── No placeholders allowed

STEP 4: LAUNCH PARALLEL
├── Single message with ALL Agent tool calls
├── Do NOT wait between launches
└── Gate: All agents launched

STEP 5: AGGREGATE RESULTS
├── Collect outputs
├── Check for conflicts
└── Merge into summary

STEP 6: VERIFICATION
├── All agents completed
├── No file conflicts
├── Deliverables match briefs
└── Gate: All 5 checks pass

Completion Announcement

When parallel execution is complete, announce:

Parallel execution complete.

**Quality Score: X/10**
- Independence Verification: X/2
- Agent Matching: X/2
- Brief Quality: X/2
- Execution Correctness: X/2
- Result Aggregation: X/2

**Results:**
- Features completed: X/Y
- Agents used: [list]
- Conflicts: [none/resolved/pending]

**Summary:**
[Brief description of what was delivered]

**Next steps:**
[Any remaining work or follow-up needed]

Integration with Other Skills

The run-parallel-agents-feature-build skill integrates with:

  • plan-to-task-list-with-dag — Use to generate a structured plan, then this skill to execute it
  • plan-founder-review — Review the plan before execution (quality gate)
  • start — Use start to identify if parallel agents are appropriate
  • run-parallel-agents-feature-debug — For debugging, use the debug variant instead
  • codemap — Use codemap search and codemap deps when preparing briefs to understand what code exists and what each file depends on
  • browse — For agents to verify their own work in the browser (e.g., testing a page they just built), use browse --session <agent-name> to isolate each agent's browser state

Workflow: start → plan-to-task-list-with-dag → plan-founder-review → run-parallel-agents-feature-build


Resources

references/

  • agent_matching_logic.md - Detailed rules for matching features to agent types, edge cases, and technology detection patterns

This skill does not require scripts or assets - it orchestrates existing Claude Code agent capabilities.

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

No reviews yet

People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.

  • It worked
  • Partly
  • It didn't
What did you use it for? e.g. Reviewing pull requests in a Rails app

Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.