Minh Đỗ
Iron Forge Documentation

Superpowers

Bộ 14 skills workflow hoàn chỉnh cho coding agents. Biến AI thành đồng đội viết code có kỷ luật, có quy trình, có chất lượng.

14
Skills
84.1k
GitHub Stars
6+
Platforms
Tác giả: Jesse Vincent (obra) | GitHub Repository | v5.0.2 (Mar 2026)

Chương 1: Giới thiệu Superpowers

Superpowers là gì?

Superpowers là một complete software development workflow (quy trình phát triển phần mềm hoàn chỉnh) dành cho coding agents. Nó không chỉ là tập hợp lệnh - mà là một hệ thống tư duy mang tính kỷ luật, được thiết kế để AI agent làm việc như một lập trình viên chuyên nghiệp.

Khi bạn kích hoạt Superpowers, agent sẽ không nhảy vào viết code ngay. Thay vào đó, nó sẽ:

  1. Hỏi bạn đang thực sự muốn làm gì (Brainstorming)
  2. Đề xuất 2-3 phương án và recommend (Thiết kế)
  3. Lập kế hoạch chi tiết từng bước nhỏ 2-5 phút (Planning)
  4. Giao nhiệm vụ cho subagent thực thi, review 2 giai đoạn (Execution)
  5. Tự động verify mọi claim trước khi báo "xong" (Verification)

Tại sao 84k stars chỉ với 14 skills?

Tác giả uy tín

Jesse Vincent (obra) là developer kỳ cựu, nổi tiếng trong cộng đồng Perl/open-source. Ông là tác giả của Request Tracker (RT) - hệ thống issue tracking được dùng rộng rãi bởi các tổ chức lớn trên toàn thế giới. Với hàng thập kỷ kinh nghiệm xây dựng công cụ cho developer, Jesse hiểu sâu sắc về quy trình phát triển phần mềm và đã đúc kết tất cả vào Superpowers.

Giải quyết Pain Point lớn nhất của AI Coding

AI agents hiện tại có một vấn đề nghiêm trọng: viết code bừa bãi, không test, không plan, nói "xong" mà chưa verify. Cụ thể:

Superpowers fix chính xác tất cả vấn đề này bằng kỷ luật tuyệt đối - mỗi skill có "Iron Law" (luật sắt) không được vi phạm dưới bất kỳ lý do nào.

Triết lý cốt lõi

TDD First

Viết test trước, xem nó fail, rồi mới viết code. Không ngoại lệ.

YAGNI

You Aren't Gonna Need It - loại bỏ mọi thứ không cần thiết khỏi design.

Evidence > Claims

Bằng chứng trước lời nói. Chạy verify rồi mới claim kết quả.

Cài đặt trên các Platform

PlatformCách cài
Claude Code (Official)/plugin install superpowers@claude-plugins-official
Claude Code (Custom)/plugin marketplace add obra/superpowers-marketplace rồi /plugin install superpowers@superpowers-marketplace
Cursor/add-plugin superpowers hoặc search "superpowers" trong plugin marketplace
Gemini CLIgemini extensions install https://github.com/obra/superpowers
CodexNói với Codex: "Fetch and follow instructions from https://raw.githubusercontent.com/obra/superpowers/refs/heads/main/.codex/INSTALL.md"
OpenCodeNative skill discovery via ~/.agents/skills/superpowers/ symlink

Sau khi cài, bạn không cần làm gì đặc biệt. Skills tự kích hoạt khi agent nhận ra context phù hợp. Verify bằng cách yêu cầu agent: "help me plan this feature" hoặc "let's debug this issue".

Chương 2: The Core Workflow - 7 bước chính

Đây là "happy path" (luồng chính) khi bạn yêu cầu Superpowers xây dựng một tính năng mới. Mỗi bước tương ứng một skill cụ thể, và agent tự động kích hoạt skill phù hợp:

1

Collaboration brainstorming

Hỏi từng câu một, khám phá ý tưởng qua đối thoại Socratic. Đề xuất 2-3 phương án với trade-offs. Trình bày thiết kế từng phần. HARD-GATE: không được viết bất kỳ dòng code nào cho đến khi user approve design.

2

Isolation using-git-worktrees

Tạo workspace cách ly trên branch mới. Auto-detect project setup (npm install / cargo build / pip install). Chạy test baseline để đảm bảo clean start. Verify directory ignored bằng git check-ignore.

3

Collaboration writing-plans

Chia công việc thành bite-sized tasks 2-5 phút. Mỗi step: viết test → chạy fail → viết code → chạy pass → commit. File paths chính xác, code đầy đủ, lệnh verify cụ thể.

4

Execution subagent-driven-development

Dispatch fresh subagent cho mỗi task. Two-stage review: (1) Spec Compliance - code đúng spec? (2) Code Quality - code sạch? Loop cho đến khi pass cả 2. Model selection tự động theo complexity.

5

Quality test-driven-development

Iron Law: NO PRODUCTION CODE WITHOUT A FAILING TEST FIRST. Mỗi subagent tuân thủ RED-GREEN-REFACTOR. Viết code trước test? Delete it. Start over. Không ngoại lệ.

6

Review requesting-code-review

Dispatch reviewer subagent với context chính xác (WHAT, PLAN, BASE_SHA → HEAD_SHA). Critical issues block progress. Important issues fix trước khi tiếp. Minor: ghi nhận.

7

Collaboration finishing-a-development-branch

Verify tests pass → Present 4 options: Merge locally / Create PR / Keep as-is / Discard → Execute choice → Cleanup worktree. Option Discard yêu cầu gõ chính xác "discard" để confirm.

Chương 3: brainstorming - Thiết kế trước khi code

HARD-GATE: KHÔNG VIẾT CODE CHO ĐẾN KHI USER APPROVE DESIGN

Brainstorming là skill bắt buộc trước bất kỳ creative work nào. "Creative work" bao gồm: tính năng mới, thay đổi architecture, refactor lớn, hoặc bất kỳ thay đổi nào cần thiết kế.

Quy trình chi tiết (7 bước)

BướcHành độngChi tiết
1. ExploreTìm hiểu contextĐọc codebase, hiểu constraints, xác định scope
2. One QuestionHỏi từng câu mộtKHÔNG hỏi nhiều câu cùng lúc. Mỗi câu phải dẫn đến insight mới.
3. ApproachesĐề xuất 2-3 hướngMỗi hướng có trade-offs rõ ràng. Recommend 1 hướng và giải thích WHY.
4. PresentTrình bày design theo sectionsTừng phần một, không dump toàn bộ. Dùng diagrams khi cần.
5. SpecViết specificationTừ design → spec chi tiết. Đây là tài liệu chính để writing-plans dùng.
6. Review LoopSpec review (max 5 vòng)User review spec. Nếu feedback → sửa → review lại. Max 5 iterations.
7. TransitionUser approve → writing-plansSau khi approve, chuyển sang planning. KHÔNG quay lại brainstorming.

Visual Companion (Trợ thủ trực quan)

Superpowers v5.0 thêm browser-based Visual Companion - một server nhẹ (zero-dependency, tự động tắt sau 30 phút idle) để hiển thị:

Quy tắc: Chỉ dùng Visual Companion cho câu hỏi mang tính trực quan. Câu hỏi conceptual (ví dụ "nên dùng REST hay GraphQL?") → text là đủ.

Anti-Patterns - KHÔNG BAO GIỜ làm

Agent nghĩ...Vấn đề
"Quá đơn giản, không cần design"MỌI project đều cần design, dù là todo list hay config change. Simple ≠ trivial.
"User đã nói rõ, khỏi hỏi"User biết WHAT, bạn phải tìm ra HOW. Hỏi để discover hidden constraints.
"Hỏi 5 câu cùng lúc cho nhanh"User bị overwhelm. Mỗi câu trả lời thay đổi câu hỏi tiếp theo. One at a time.
"Viết code song song với design"HARD-GATE violation. Code viết trong brainstorming sẽ bị delete.
"Đã approve approach, bắt đầu code"Approve approach ≠ approve spec. Phải có written specification trước khi chuyển sang writing-plans.

Checklist trước khi chuyển sang Planning

Chương 4: writing-plans - Kế hoạch chi tiết

MỖI TASK PHẢI ĐỦ NHỎ ĐỂ MỘT ENGINEER KHÔNG CÓ CONTEXT CŨNG LÀM ĐƯỢC TRONG 2-5 PHÚT

Plan trong Superpowers không phải "danh sách việc cần làm". Nó là blueprint chi tiết - giả định engineer đọc plan không có context, không có taste (thẩm mỹ code), không biết gì về project.

Cấu trúc Plan Document

Mỗi plan document bắt đầu bằng header bắt buộc:

# Implementation Plan: [Feature Name]
## Context
[Tóm tắt từ spec: mục tiêu, constraints, decisions đã đưa ra]

## Tasks
### Task 1: [Tên task mô tả]
**File:** `src/components/UserProfile.tsx`
**Test file:** `src/components/__tests__/UserProfile.test.tsx`

1. Viết test:
   ```typescript
   test('renders user name', () => { ... })
   ```
2. Chạy: `npm test -- UserProfile`
3. Verify: test FAILS với "Cannot find module"
4. Viết implementation: [code đầy đủ]
5. Chạy lại: `npm test -- UserProfile`
6. Verify: test PASSES
7. Commit: `git commit -m "feat: add UserProfile component"`

Quy tắc DRY, YAGNI, TDD trong Planning

Nguyên tắcÁp dụng trong plan
DRYNếu 2 tasks tạo ra code tương tự → extract helper trong task riêng trước
YAGNIKhông có task nào "để sau có thể cần". Mỗi task phải giải quyết requirement hiện tại.
TDDMỗi task bắt đầu bằng test. Test file path + test code cụ thể. Không bao giờ skip.

Scope Check

Nếu spec quá lớn (nhiều subsystem độc lập), tách thành nhiều plan riêng biệt:

Plan Review Loop

Sau khi viết plan, bắt buộc review trước khi execute:

  1. Agent tự review: tasks đúng thứ tự? Dependencies hợp lý? Có task nào quá lớn?
  2. Present plan cho user (hoặc reviewer subagent)
  3. Feedback → sửa plan → review lại
  4. Approve → chuyển sang execution

Chương 5: subagent-driven-development - Giao việc cho subagent

Đây là "động cơ chính" (Core Engine) của Superpowers. Thay vì agent chính tự làm mọi thứ, nó đóng vai trò Controller - đọc plan, dispatch từng task cho fresh subagent, review kết quả.

Tại sao cần fresh subagent?

Vấn đề của 1 agent liên tụcGiải pháp subagent
Context dài → agent quên instructionFresh agent với task-specific context duy nhất
Sai ở task 3 → ảnh hưởng task 4+Mỗi subagent độc lập, lỗi không lan truyền
Agent tự review code mình → thiên vịReviewer subagent không biết ai viết code

Quy trình Controller

  1. Load plan - đọc plan document 1 lần, extract tất cả tasks
  2. Dispatch task - tạo fresh subagent với prompt:
    You are implementing: [Task description]
    Project: [project context]
    Specification: [relevant spec section]
    File to modify: [exact path]
    Test to write first: [exact test code]
    Verification command: [exact command]
    IMPORTANT: Follow TDD - write test first, verify it fails,
    then implement, verify it passes.
  3. Two-stage Review sau mỗi task (xem chi tiết bên dưới)
  4. Handle status - dựa trên kết quả subagent
  5. Dispatch next hoặc pause nếu có issue

Two-stage Review (Đánh giá 2 giai đoạn)

Stage 1: Spec Compliance

Code có đúng specification không? Tests cover đúng requirements không? Thiếu gì so với plan?

Stage 2: Code Quality

Code sạch không? Naming conventions đúng? DRY? YAGNI? Có unnecessary complexity không?

Task chỉ được coi là DONE khi pass cả 2 stage. Nếu fail ở stage nào → dispatch fix subagent → review lại.

Model Selection (Chọn model tự động)

Task typeComplexityModel
Mechanical (rename, move, format)1-2 files thay đổiCheap / fast model
Integration (connect APIs, wire components)3-5 filesStandard model
Architecture / ReviewCross-cutting concernsMost capable model

Status Protocol

StatusÝ nghĩaController action
DONETask hoàn thành, tests passDispatch next task
DONE_WITH_CONCERNSXong nhưng có vấn đề tiềm ẩnReview concerns → quyết định fix hay tiếp
NEEDS_CONTEXTCần thêm thông tinCung cấp context → re-dispatch
BLOCKEDKhông thể tiến hànhSTOP. Escalate to user. Không đoán.

Fallback: executing-plans

Nếu platform không hỗ trợ subagent (ví dụ một số Codex setups), dùng executing-plans thay thế:

dispatching-parallel-agents (Chạy song song)

Khi có 2+ independent problems (vấn đề hoàn toàn độc lập), dispatch nhiều agent song song:

Khi nào dùng

6 test failures ở 3 files khác nhau → 3 agents chạy song song (1 per file) → 3x nhanh hơn. Bugs ở API layer + CSS bugs ở UI → 2 agents song song.

KHÔNG dùng khi

Failures có thể liên quan (fix 1 → fix cái khác tự động). Shared state (cả 2 cần modify cùng file). Chưa biết root cause (đoán domain → giao nhầm).

Prompt Structure cho Parallel Agents

## Agent Prompt Template
1. Problem domain: [chỉ 1 domain cụ thể]
2. Files to investigate: [chỉ files trong domain]
3. Expected behavior: [output mong muốn]
4. Current behavior: [output hiện tại]
5. Rule: DO NOT modify files outside your domain
6. Rule: If you find root cause is OUTSIDE your domain, report BLOCKED

Chương 6: test-driven-development & systematic-debugging

test-driven-development (TDD)

NO PRODUCTION CODE WITHOUT A FAILING TEST FIRST

Đây là skill nghiêm ngặt nhất trong Superpowers. "No exceptions" nghĩa là literally no exceptions:

RED-GREEN-REFACTOR Cycle (Chi tiết)

RED

Viết 1 test failure tối thiểu. Chạy nó. Phải fail. Fail vì feature chưa có, không phải vì syntax error hay typo. Nếu test pass ngay → test sai (tautology).

GREEN

Viết code tối thiểu nhất để test pass. Hardcode nếu cần. KHÔNG thêm error handling, logging, optimization, hay bất kỳ gì khác. YAGNI.

REFACTOR

Sau khi green: loại bỏ duplication, improve naming, extract helpers. Tests phải vẫn green. Nếu tests fail sau refactor → refactor sai.

Viết code trước test? Iron Law Response:

DELETE IT. START OVER.

Không giữ lại làm "reference". Không "adapt" trong lúc viết test. Không nhìn vào code cũ. Delete means delete.

Lý do: Nếu bạn giữ code, bạn sẽ viết tests that prove the code works (testing-after) thay vì tests that define what the code should do (testing-before). Hai cách tiếp cận này khác nhau hoàn toàn về giá trị.

Master Rationalization Table cho TDD

Excuse (Lý do viện)Reality (Thực tế)
"Quá đơn giản để test"Code đơn giản vẫn break. Test viết mất 30 giây. Regression prevention miễn phí.
"Tôi sẽ test sau"Tests viết sau = "code này làm gì?" Tests viết trước = "code này PHẢI làm gì?" → khác nhau hoàn toàn.
"Xóa X giờ code lãng phí"Sunk cost fallacy. Code không đúng quy trình = nợ kỹ thuật. Xóa bây giờ rẻ hơn fix sau.
"TDD chậm quá"TDD nhanh hơn debug. Tìm bug trước commit, prevent regression. Pragmatic = test-first.
"Giữ lại tham khảo rồi viết test trước"Bạn sẽ adapt nó. Đó là testing-after với extra steps. Delete means delete.
"Emergency, no time"Emergency = higher risk = MORE reason for TDD, not less.
"Test framework chưa setup"Setup test framework IS the first task. BEFORE any production code.
"Code này chỉ là prototype"Prototypes become production code. Always. Test it now.

verification-before-completion

NO COMPLETION CLAIMS WITHOUT FRESH VERIFICATION EVIDENCE

Nói "xong" mà không verify = nói dối, không phải hiệu quả. Skill này thiết lập 5-step gate function:

StepActionChi tiết
IDENTIFYXác định lệnh verifyTest command, build command, lint command - tuỳ context
RUNChạy lệnhChạy FRESH, không dùng cached results
READĐọc FULL outputKhông scan lướt. Đọc từng dòng. Đặc biệt chú ý warnings.
VERIFYOutput khớp claim"Tests pass" → phải thấy "X tests passed, 0 failed" trong output
CLAIMBáo kết quảCHỈ sau khi verify xong. Kèm evidence (dẫn chứng) cụ thể.

Red Flags - STOP ngay lập tức nếu thấy:

systematic-debugging (Debug có hệ thống)

NO FIXES WITHOUT ROOT CAUSE INVESTIGATION FIRST

Random fixes lãng phí thời gian và tạo bug mới. Quick patches che giấu vấn đề gốc. Superpowers chia debugging thành 4 phases bắt buộc:

Phase 1: Root Cause Investigation

Phase 2: Pattern Analysis

Phase 3: Hypothesis and Testing

Phase 4: Implementation

Quy tắc 3 lần

3+ Failed Fixes = Architectural Problem

Nếu đã thử 3+ fixes mà chưa work → STOP. Không phải bạn fix sai - mà bạn đang nhìn sai vấn đề. 3+ failures = architecture problem, không phải implementation problem. Quay lại Phase 1 với mindset "root cause có thể ở tầng khác hoàn toàn".

Multiple fix cùng lúc?

KHÔNG BAO GIỜ. Fix nhiều thứ cùng lúc → nếu pass, bạn không biết fix nào thực sự cần thiết. Nếu fail, bạn không biết fix nào gây vấn đề mới. 1 fix = 1 test = 1 commit.

Chương 7: Code Review - Đánh giá code nghiêm túc

Code review trong Superpowers chia thành 2 skills: requesting (yêu cầu review) và receiving (nhận feedback). Cả hai đều có quy trình nghiêm ngặt.

requesting-code-review

Sau khi hoàn thành implementation, bắt buộc request review trước khi merge:

  1. Chuẩn bị context cho reviewer:
    • WHAT: Tóm tắt những gì đã implement
    • PLAN: Link hoặc nội dung spec/plan
    • BASE_SHA → HEAD_SHA: Range commit thay đổi
  2. Dispatch reviewer subagent (agent riêng, KHÔNG dùng agent chính):
    Review these changes:
    What: [implementation summary]
    Requirements: [from spec]
    Changes: `git diff BASE_SHA HEAD_SHA`
    Focus on: spec compliance, code quality, test coverage
  3. Action based on severity:
    SeverityAction
    CriticalFix ngay lập tức. Block mọi progress.
    ImportantFix trước khi tiếp tục task tiếp theo.
    MinorGhi nhận (log). Fix khi convenient.

receiving-code-review

Skill này đặc biệt vì nó chống lại một tendencies (xu hướng) tự nhiên của AI agents: performative agreement (đồng ý để tỏ ra lịch sự).

CODE REVIEW YÊU CẦU ĐÁNH GIÁ KỸ THUẬT, KHÔNG BIỂU DIỄN CẢM XÚC
KHÔNG BAO GIỜ nóiTHAY BẰNG
"You're absolutely right!""Fixed. [mô tả ngắn thay đổi]"
"Great point!""Good catch - [vấn đề cụ thể]. Fixed in [vị trí]."
"Thanks for catching that!""Resolved: [mô tả fix]"
"I totally agree""Verified: [reference evidence]. Implementing."

Response Pattern (Quy trình nhận feedback)

StepHành độngChi tiết
1. READĐọc hết feedbackĐọc xong tất cả rồi mới phản ứng. Không reply từng comment.
2. UNDERSTANDDiễn đạt lại requirementNếu không rõ: "Are you suggesting X because of Y?" - hỏi cụ thể.
3. VERIFYCheck codebaseReviewer có thể sai. Verify bằng code thực, không tin review mù quáng.
4. EVALUATEÁp dụng cho codebase NÀYBest practice chung ≠ đúng cho project này. Context matters.
5. RESPONDTechnical replyAcknowledgment hoặc reasoned pushback. Không performative.
6. IMPLEMENT1 item 1 lúcFix từng feedback item, test mỗi cái, commit riêng.

YAGNI Check khi reviewer đề xuất

Khi reviewer nói "implement this properly" hoặc "add error handling for X": Grep codebase xem feature đó có ai dùng không. Nếu endpoint/function/feature không được gọi ở đâu → "This isn't called anywhere. Remove it? (YAGNI)" Đây là pushback hợp lệ, không phải lazy.

Khi nào Push Back?

Chương 8: Git Worktrees, Meta Skills & Kết thúc

using-git-worktrees (Workspace cách ly)

Git worktree cho phép nhiều branches được checkout đồng thời trong cùng repo. Superpowers dùng worktrees để tạo isolated workspace cho mỗi feature.

Quy trình tạo Worktree

  1. Chọn directory: Kiểm tra .worktrees/ (convention) → đọc CLAUDE.md → hỏi user
  2. Tạo branch + worktree:
    git worktree add .worktrees/feature-name -b feature/feature-name
  3. Safety verification:
    git check-ignore .worktrees/feature-name
    # Phải return path (= đã ignored). Nếu không → thêm vào .gitignore
  4. Project setup: Auto-detect và chạy:
    • package.jsonnpm install
    • Cargo.tomlcargo build
    • requirements.txtpip install -r requirements.txt
  5. Baseline tests: Chạy test suite → tất cả phải pass. Nếu fail → worktree có vấn đề, không phải code bạn.

using-superpowers (Kích hoạt Skills)

Skill "meta" quan trọng nhất - quy định cách agent sử dụng toàn bộ hệ thống:

IF EVEN 1% CHANCE A SKILL MIGHT APPLY, YOU MUST INVOKE THE SKILL

Priority Hierarchy (Thứ tự ưu tiên)

MứcSourceOverride
1 (Cao nhất)User instructions (CLAUDE.md, direct requests)Override tất cả
2Superpowers skillsOverride system default
3 (Thấp nhất)Default system promptBị override bởi cả 2 trên

Skill Priority (Process trước Implementation)

Skill check phải xảy ra TRƯỚC bất kỳ response nào, kể cả clarifying questions.

writing-skills (Viết skills mới)

Superpowers dùng TDD để viết... documentation! Chính skill này cũng tuân thủ RED-GREEN-REFACTOR:

TDD PhaseÁp dụng cho Skill
REDRun pressure scenario WITHOUT skill → ghi baseline failures. Agent skip steps nào? Rationalize bằng gì?
GREENViết skill address specific failures từ RED phase. Không viết skill "dự phòng".
REFACTORTìm new rationalizations agent dùng để bypass skill → thêm chặn. Đây là bước lặp liên tục.

CSO - Claude Search Optimization

Skill discovery (tìm skill phù hợp) dựa trên description trong frontmatter. Quy tắc CSO:

finishing-a-development-branch

Skill kết thúc workflow. Sau khi verify tests:

OptionMergePushKeep WorktreeCleanup
1. Merge locally--
2. Create PR--
3. Keep as-is---
4. Discard---✓ (force)

Option 4 (Discard): Yêu cầu gõ chính xác "discard" để confirm. Agent không bao giờ delete work mà không hỏi user.

Chương 9: Iron Laws & Anti-Patterns tổng hợp

Tổng hợp tất cả "luật sắt" và anti-patterns từ 14 skills tại một nơi. Đây là checklist để bạn hoặc agent kiểm tra mình có đang vi phạm không.

4 Iron Laws

SkillIron Law
test-driven-developmentNO PRODUCTION CODE WITHOUT A FAILING TEST FIRST
verification-before-completionNO COMPLETION CLAIMS WITHOUT FRESH VERIFICATION EVIDENCE
systematic-debuggingNO FIXES WITHOUT ROOT CAUSE INVESTIGATION FIRST
writing-skillsNO SKILL WITHOUT A FAILING TEST FIRST

Master Red Flags

Bạn/Agent đang nghĩ...Phải làm gì
"Quá đơn giản, không cần design"STOP → Brainstorming. Mọi project cần design.
"Quick fix giờ, investigate sau"STOP → Phase 1 Root Cause. Quick fix = technical debt.
"Viết code trước, test sau"DELETE code. Start over with RED phase.
"Should work now" / "Probably fine"RUN the verification command. Evidence over claims.
"Thử thêm 1 fix nữa" (đã fail 2+)STOP → Reassess. 3+ failures = architecture problem.
"Tôi biết skill này rồi"Skills evolve between versions. Invoke current version.
"You're absolutely right!"DELETE response. State the technical fix instead.
"Emergency, no time for TDD"Emergency = higher risk = MORE reason for TDD.
"Fix nhiều thứ cùng lúc cho nhanh"Can't isolate what worked. 1 fix = 1 test = 1 commit.
"Test framework chưa setup"Setup IS task 1. BEFORE any code.

Chương 10: Release History & Tương lai

Superpowers phát triển rất nhanh, từ v2.0 (Oct 2025) đến v5.0.2 (Mar 2026) - 5 tháng, 22 releases. Dưới đây là milestone quan trọng:

VersionNgàyHighlights
v2.0Oct 2025Skills repo tách riêng. 9 skills mới. Auto-update trên session start.
v3.0.1Oct 2025Chuyển sang Anthropic first-party skills system.
v3.2.0Oct 2025Design docs trong brainstorming. superpowers: namespace.
v3.3.0Oct 2025Codex support (experimental). Cross-platform mở đầu.
v4.0.0Dec 2025Two-stage review. DOT flowcharts. Debugging tools bundled. Test infrastructure.
v4.2.0Feb 2026Worktree isolation bắt buộc. Main branch protection (yêu cầu consent).
v4.3.0Feb 2026Brainstorming HARD-GATE. Cursor support chính thức.
v5.0.0Mar 2026Visual brainstorming companion. Document review system. Architecture guidance.
v5.0.2Mar 2026Zero-dependency brainstorm server. Auto-exit 30 phút idle. Context isolation all delegation skills.

Cross-platform Evolution

Mục tiêu: mọi coding agent đều có thể dùng Superpowers, bất kể platform.

Windows Support

Nhiều bản fix dành riêng cho Windows: hook quoting issues (single vs double quotes), bash path discovery, symlink handling, CRLF line endings, O(n^2) performance fix trong escape_for_json. Windows là môi trường khó nhất nhưng được hỗ trợ đầy đủ.

Community & Links

Superpowers là open-source (MIT License). 24 contributors, 6.6k forks.

GitHub: obra/superpowers | Blog: Superpowers for Claude Code