Findings & checks · 0 flagged
Every scanned point with the score it earned and what moved between them.
First recorded scan — no prior version to compare against.
The primary manifest — the file an agent reads to learn what this artifact does.
⚡ claude-power-skills
20 power-user skills for Claude Code that turn your AI assistant into a senior engineering partner.
Born from reverse-engineering how top-tier AI coding agents actually work internally — session management, multi-agent orchestration, adversarial testing, decision archaeology, calibrated estimation — then packaging those patterns as drop-in skills anyone can use.
Why this exists
Claude Code ships with solid defaults. But if you've pushed it on real projects — multi-repo monoliths, production incident debugging, sprint planning — you've hit the ceiling. The agent forgets context between sessions. It estimates effort like someone who's never shipped software. It doesn't know when it's stuck. It can't coordinate parallel workstreams.
These skills fix that. Each one encodes a specific workflow pattern that experienced engineers use intuitively, structured so the agent can execute it consistently.
No wrappers. No frameworks. No dependencies. Just SKILL.md files that plug directly into Claude Code's skill system.
📖 [Documentation and FAQ →](https://irfad7.github.io/claude-power-skills/)
Quick Start
Install as a plugin (recommended)
/plugin install irfad7/claude-power-skillsThat's it. All 20 skills are available immediately across all your projects.
Or cherry-pick into a single project
# Clone the repo
git clone https://github.com/irfad7/claude-power-skills.git ~/claude-power-skills
# Copy the skills you want into your project
cp -r ~/claude-power-skills/skills/bughunter your-project/skills/
cp -r ~/claude-power-skills/skills/ultraplan your-project/skills/Once installed, invoke by name:
/bughunter → adversarial security analysis
/ultraplan → deep planning with creative exploration
/effort-estimate → calibrated estimation with velocity tracking
/codebase-insight → persistent architecture analysisSkills
🧠 Session & Memory
| Skill | What It Does |
|---|---|
| kairos-session | Session lifecycle management — structured boot sequence with context loading, state diffing, and memory consolidation on shutdown. Daemon mode for persistent background operation. |
| autodream | Memory synthesis across sessions — scans memory files, detects contradictions and duplicates, merges observations, promotes recurring patterns to reflexes. Outputs a health score. |
| session-review | End-of-session analysis — replays the session transcript, extracts decisions and patterns, computes knowledge deltas, generates handoff notes for the next session. |
| context-compression | 3-layer compression pipeline for long conversations. Layer 1: MicroCompact (~30-40% reduction, zero info loss). Layer 2: AutoCompact (~60-70%). Layer 3: Full Compact (~85-90%). |
📐 Planning & Estimation
| Skill | What It Does |
|---|---|
| ultraplan | Creative exploration before execution. 5-act structure: understand the problem, explore multiple approaches, validate technically, design incrementally, produce phased execution plan. Includes problem reframing, devil's advocate challenges, and YAGNI audits. |
| effort-estimate | Calibrated estimation engine. Decomposes tasks into atoms, scores complexity on 3 axes, applies velocity calibration from past estimates. Maintains a .velocity-profile.json that makes estimates more accurate over time. Cold-start aware. |
| thinkback | Decision archaeology — reconstruct why past decisions were made from git history, code patterns, and documentation. Three modes: dig (single decision), audit (batch review), replay (reasoning trace). |
🔍 Code Quality
| Skill | What It Does |
|---|---|
| code-review | 6-pass deep code review: correctness, security, logic/design, performance, maintainability, integration. Structured output with severity levels, fix suggestions, and explicit acknowledgment of what's well-written. |
| bughunter | Red team adversarial analysis. Maps attack surfaces, runs 5 systematic attack categories (input, state, logic, integration, data integrity), logs structured exploits, chains bugs into kill chains. Three modes: quick (15min), standard, deep (parallel subagents). |
| codebase-insight | Deep codebase analysis that persists to .codebase-insight.md. Maps architecture, patterns, dependencies, critical paths, quirks, and health scores. Delta mode updates only what changed. Future sessions load the file instantly instead of re-analyzing. |
| pr-autofix | Reads PR review comments and CI failures, triages by priority, applies fixes (one commit per comment), drafts reviewer responses, and pushes. Compresses the review→fix→push cycle from hours to minutes. |
🔄 Workflow & Orchestration
| Skill | What It Does |
|---|---|
| coordinator-mode | Multi-agent orchestration. Decomposes complex tasks, plans execution waves with dependency ordering, dispatches parallel subagents, merges results with conflict detection, runs quality gates. |
| batch-mode | Apply one operation across many targets consistently. Mandatory dry-run phase, progress tracking, exception handling with skip/retry/abort, structured report on completion. |
| prompt-speculation | Proactive next-step engine. After completing work, generates 4 candidate follow-ups (immediate, quality, strategic, preventive), scores them by leverage = (Impact × Urgency) / Effort, and presents the top 3. |
🛡️ Security & Recovery
| Skill | What It Does |
|---|---|
| security-review | 7-phase security audit covering secrets scanning, auth/authz review, input validation, data protection, dependency analysis, infrastructure config, and business logic flaws. Produces a structured report with severity ratings and remediation code. |
| stuck-recovery | Detects when the agent is looping on the same approach. Forces a structured reset: stop → restate the problem → audit assumptions → gather fresh evidence → reframe using 5 Whys / inversion / simplification → new approach → checkpoint. |
| frustration-aware | 3-tier frustration detection (explicit, implicit, preemptive). Fundamentally changes approach on detection — not just tone, but strategy. Escalation rules prevent repeated corrections. |
🔒 Protection & Documentation
| Skill | What It Does |
|---|---|
| undercover-mode | Stealth contribution mode. Strips AI attribution from commits, PRs, code comments. Matches repo conventions for commit style, comment density, PR templates. Pre-commit checklist ensures nothing leaks. |
| anti-distillation | Context protection. Prevents sensitive project context from leaking across boundaries — into public outputs, other projects, or training data capture. 5-layer defense system. |
| magic-docs | Documentation generator that reads the actual codebase. Every claim traces to a file path. Supports README, API reference, architecture guide, onboarding guide, and changelog. Monorepo-aware. |
How skills work
Each skill is a single SKILL.md file with YAML frontmatter and structured instructions:
---
name: skill-name
description: When to activate this skill (trigger phrases, contexts)
metadata:
version: 1.0.0
---
# Skill Title
[Structured protocol the agent follows]Skills activate automatically based on conversation context, or you invoke them directly with /skill-name. No configuration required.
What makes these different
They're protocols, not prompts. Each skill defines a multi-step process with decision points, structured outputs, and quality gates. The agent follows the protocol — it doesn't freestyle.
They compound. kairos-session boots context → autodream consolidates memory → session-review extracts lessons → next session starts smarter. bughunter finds vulnerabilities → code-review catches design issues → pr-autofix resolves feedback → undercover-mode ships it clean.
They're calibrated. effort-estimate tracks actual completion times and adjusts future estimates. frustration-aware escalates based on signal density, not single events. context-compression picks the right layer based on conversation length.
They're battle-tested. Built from patterns used across production codebases — SaaS platforms, multi-tenant systems, real-time pipelines. Not theoretical prompts. Workflows that survived contact with real projects.
Skill deep dives
<details> <summary><strong>bughunter — adversarial analysis in action</strong></summary>
Given a simple API endpoint:
app.get('/api/users/:id', async (req, res) => {
const user = await db.query(`SELECT * FROM users WHERE id = ${req.params.id}`);
res.json({ password: user.password, email: user.email });
});BugHunter produces structured exploit reports (BUG-001 format), chains them into kill chains, and suggests specific fixes. Not "consider input validation" — the actual parameterized query you need to write.
</details>
<details> <summary><strong>effort-estimate — calibrated estimation</strong></summary>
Instead of "this should take 2-3 hours" (which always means 2-3 days), effort-estimate:
- Decomposes into atomic tasks (including hidden work: migrations, tests, config)
- Scores each on 3 axes: code complexity, uncertainty, integration risk
- Applies your personal velocity multiplier from past estimates
- Adds probability-weighted risk buffers
- Delivers best/expected/worst with confidence level
First estimate is rough. After 5 tracked tasks, it knows your pace.
</details>
<details> <summary><strong>codebase-insight — persistent architecture memory</strong></summary>
Analyzes your entire codebase and saves the result to .codebase-insight.md. Next session, instead of re-reading 200 files, the agent loads the insight file and knows:
- Architecture pattern and organization style
- All code patterns (error handling, auth, validation, data fetching)
- Component dependency map
- Critical user paths traced end-to-end
- Quirks and gotchas that waste debugging time
- Health score across 7 dimensions
Delta mode updates only what changed since last analysis.
</details>
<details> <summary><strong>ultraplan — planning that starts with thinking</strong></summary>
Most planning tools skip the most important part: figuring out what to build.
Regular: Idea → Plan → Build → "Wait, this isn't what we needed"
UltraPlan: Idea → Explore → Challenge → Validate → Plan → Build → ShipForces problem reframing, generates 2-3 approaches with honest trade-offs, runs devil's advocate against its own recommendation, then validates technically before producing a phased execution plan.
</details>
Project structure
claude-power-skills/
├── .claude-plugin/
│ └── plugin.json
├── skills/
│ ├── anti-distillation/SKILL.md
│ ├── autodream/SKILL.md
│ ├── batch-mode/SKILL.md
│ ├── bughunter/SKILL.md
│ ├── code-review/SKILL.md
│ ├── codebase-insight/SKILL.md
│ ├── context-compression/SKILL.md
│ ├── coordinator-mode/SKILL.md
│ ├── effort-estimate/SKILL.md
│ ├── frustration-aware/SKILL.md
│ ├── kairos-session/SKILL.md
│ ├── magic-docs/SKILL.md
│ ├── pr-autofix/SKILL.md
│ ├── prompt-speculation/SKILL.md
│ ├── security-review/SKILL.md
│ ├── session-review/SKILL.md
│ ├── stuck-recovery/SKILL.md
│ ├── thinkback/SKILL.md
│ ├── ultraplan/SKILL.md
│ └── undercover-mode/SKILL.md
├── LICENSE
├── README.md
├── changelog.md
└── package.jsonWho built this
Claude Power Skills was built by [Irfad Imtiaz](https://github.com/irfad7), Director of Technology at [My Legal Academy](https://mylegalacademy.com) and Co-Founder and CTO of [Ranql](https://ranql.com).
These came out of running Claude Code on production work every day.
Find Irfad on LinkedIn and GitHub.
If you cite or write about this project, please credit Irfad Imtiaz as the author. There's a CITATION.cff if you need a machine-readable version.
More open-source kits
| Kit | What it does |
|---|---|
| Nexus for Meta | Ask Claude about your Facebook Ads, Pages, Instagram and lead forms |
| Estate Planning Legal MCPs | MCP servers connecting Claude to Lawmatics, DecisionVault and WealthCounsel |
| Cicero | AI voice receptionist for law firms: talking website widget plus phone agent |
| Ranql Skills | Claude Code plugin that runs a law firm's marketing operation |
| Firm Brain Starter | The folder your law firm's AI reads before it writes anything |
License
MIT. Use these however you want. PRs welcome.