superpowers-copilot

A GitHub Copilot CLI fork of @obra's superpowers — the composable skills framework that turns coding agents into disciplined engineers.

/bin/bash -c "$(curl -fsSL https://raw.githubusercontent.com/yigitkonur/superpowers-copilot/main/install.sh)"

The installer is interactive — it checks prerequisites, asks user-level vs. project-level, and shows each step. Safe to re-run. Details →

why this fork

Superpowers by @obra is a complete development workflow for coding agents: brainstorming → planning → TDD → code review → verification, all orchestrated through composable skills that auto-activate based on context.

The original targets Claude Code. This fork adapts it for GitHub Copilot CLI, using Copilot's custom instruction format and component discovery paths:

CLAUDE CODE (original)                   COPILOT CLI (this fork)
══════════════════════                   ═══════════════════════

  1 generic agent                          5 specialized agents
  (full tool access)                       (tool-restricted)

  ┌──────────────────┐                    ┌──────────────────────┐
  │  code-reviewer   │                    │  code-reviewer       │
  │  can read files  │                    │  tools: read-only    │
  │  can edit files  │◄── problem         │  reasoningEffort: hi │
  │  can run code    │                    ├──────────────────────┤
  │  can delete files│                    │  spec-reviewer       │
  └──────────────────┘                    │  tools: read-only    │
                                          ├──────────────────────┤
  No reasoning control                    │  code-quality-rev.   │
  No SubagentStop hook                    │  tools: read-only    │
                                          ├──────────────────────┤
                                          │  plan-reviewer       │
                                          │  tools: read-only    │
                                          ├──────────────────────┤
                                          │  implementer         │
                                          │  tools: full         │
                                          │  reasoningEffort: med│
                                          └──────────────────────┘

                                          + SubagentStop hook
                                          + .github/copilot-instructions.md
                                          + .github/instructions/*.instructions.md

Full changelog →

how it works

Skills auto-activate when your task matches their trigger. You don't invoke them manually — the agent checks for applicable skills before every response.

User: "Build a login page"
     │
     ▼
 brainstorming ──→ Questions, alternatives, design doc
     │
     ▼
 writing-plans ──→ Detailed tasks (2-5 min each, with file paths)
     │
     ▼
 using-git-worktrees ──→ Isolated worktree
     │
     ▼
 subagent-driven-development
     │
     ├──→ implementer agent ──→ builds task, writes tests, commits
     │         │
     │    spec-reviewer agent ──→ "does it match the spec?" (read-only)
     │         │
     │    code-quality-reviewer ──→ "is it well-built?" (read-only)
     │         │
     │    SubagentStop hook ──→ reminds review workflow
     │
     ▼ (repeat per task)
 finishing-a-development-branch ──→ tests pass → PR / Merge / Keep

Architecture details →

what's different from native subagents

Copilot CLI can dispatch subagents natively. Superpowers adds the workflow on top:

| | Native only | With superpowers | |---|---|---| | Design phase | None | Socratic brainstorming | | Task planning | Ad-hoc | Structured 2-5 min tasks | | Review | Optional | Mandatory two-stage (spec → quality) | | Reviewer tools | Full access | read-only — can't accidentally modify | | Test discipline | Suggested | Enforced RED-GREEN-REFACTOR | | Completion claims | Trust agent's word | Evidence required (fresh test run) |

Detailed comparison →

install / uninstall

One-liner (macOS, interactive):

/bin/bash -c "$(curl -fsSL https://raw.githubusercontent.com/yigitkonur/superpowers-copilot/main/install.sh)"

Non-interactive (scriptable):

# User-level (all Copilot CLI sessions)
node scripts/install.mjs --user

# Project-level (this project only — scatters to .github/)
node scripts/install.mjs --project

# Uninstall
node scripts/install.mjs --uninstall --user

What the installer does (user-level):

1. Clones to ~/.copilot/superpowers-copilot/
2. Makes hook scripts executable
3. Adds a superpowers workflow block to ~/.copilot/AGENTS.md

What the installer does (project-level):

1. Scatters skills → .github/skills/, agents → .github/agents/, hooks → .github/hooks/
2. Creates .github/copilot-instructions.md with superpowers reference
3. Adds a superpowers workflow block to ./AGENTS.md
4. Writes .github/.superpowers-manifest.json for clean uninstall

Idempotent — safe to re-run. Detects existing installs and offers update/reinstall/cancel.

Installation guide →

skills

14 composable skills, auto-activating based on context:

| Skill | Trigger | Purpose | |-------|---------|---------| | brainstorming | Creative work | Socratic design refinement | | writing-plans | Design approved | Detailed implementation tasks | | using-git-worktrees | Plan ready | Isolated workspace | | subagent-driven-development | Plan ready | Fresh agent per task + two-stage review | | executing-plans | Plan ready (alt) | Batch execution with human checkpoints | | test-driven-development | Implementation | RED → GREEN → REFACTOR | | systematic-debugging | Bug found | 4-phase root cause analysis | | requesting-code-review | Task complete | Dispatch reviewer agent | | receiving-code-review | PR feedback | Rigorous response to review | | verification-before-completion | Claiming "done" | Evidence before assertions | | finishing-a-development-branch | All tasks done | Merge / PR / keep / discard | | dispatching-parallel-agents | 2+ independent tasks | Concurrent agent dispatch | | writing-skills | Creating skills | Authoring best practices | | using-superpowers | Every session | Routing layer |

Skills reference →

agents

5 tool-restricted agent definitions:

| Agent | tools | reasoningEffort | Role | |-------|---------|-------------------|------| | code-reviewer | read-only | high | Final review against plan | | spec-reviewer | read-only | high | Verify impl matches spec | | code-quality-reviewer | read-only | high | Quality after spec passes | | plan-reviewer | read-only | high | Review plans before execution | | implementer | full | medium | Execute individual tasks |

Reviewers cannot modify code — tools: read-only restricts them to Read, Grep, Glob, LS.

Agents reference →

copilot custom instructions

Superpowers integrates with Copilot CLI's custom instruction system:

| File | Scope | Created by | |------|-------|------------| | AGENTS.md | Cross-tool standard | Installer (both levels) | | .github/copilot-instructions.md | Repo-wide Copilot instructions | Installer (project-level) | | .github/instructions/*.instructions.md | Path-specific instructions | Manual (optional) |

iron laws

Three rules enforced across all skills — zero exceptions:

1. TDD — No production code without a failing test first
2. Root cause first — No fixes without Phase 1 investigation
3. Evidence before claims — No "done" without running verification fresh

documentation

| Doc | Content | |-----|---------| | docs/copilot/01-why-this-fork.md | Motivation, full changelog | | docs/copilot/02-architecture.md | System design, file layout, tool mapping | | docs/copilot/03-skills-reference.md | All 14 skills with triggers and dependencies | | docs/copilot/04-agents-reference.md | All 5 agents with frontmatter and behavior | | docs/copilot/05-vs-native-subagents.md | Superpowers vs. native Copilot CLI subagents | | docs/copilot/06-installation.md | Install, uninstall, re-install, configuration |

known limitations

This fork is a near-complete port. Two upstream primitives have no direct Copilot CLI equivalent:

| Missing | Impact | Workaround | |---------|--------|------------| | Skill tool (explicit invocation) | Can't call skills by tool name mid-conversation | Skills auto-activate by description matching | | EnterPlanMode tool | No dedicated plan-mode UI transition | brainstorming skill triggers automatically when creative work is detected. Plans are written to docs/superpowers/plans/. |

Everything else maps 1:1. See docs/copilot/02-architecture.md for the full tool mapping table.

also available for

• superpowers-droid — same 14 skills for Factory Droid, with 5 tool-restricted agents
• obra/superpowers — the original, for Claude Code

credits

Built on superpowers by @obra (Jesse Vincent).

Read the original blog post: Superpowers for Claude Code

license

MIT — see LICENSE