LightFlow Slashes AI Coding Overhead to 2.4K Tokens with a 3-Skill Workflow
Behind the Superpowers Uninstall Wave, I Built a Lightweight Alternative (Open Source)
Save 89% Tokens, Smart On-Demand Activation, 3 Skills to Get the Job Done
Background
Superpowers is the most popular skill plugin in the Claude Code ecosystem, with over 25K stars on GitHub.
It solves a real problem: Agents start out a bit "reckless," writing code as soon as they get a requirement without planning or testing. Superpowers wraps it in a set of SOPs—first Brainstorm → write a Spec → make a Plan → TDD → Code → Review.
But recently, the community has started uninstalling it en masse.
The reason is surprisingly consistent: bloated, token-hungry, and it forces the full process even for small tasks.
It consumes 22,000 tokens just to start up, using 11% of the context window before any work is done. To rename a variable, it first spends a minute brainstorming, then writes a spec, then makes a plan, and only then starts making the change.
A 10-second task turns into a 5-minute ordeal.
One user even said, "After installing Superpowers, Claude actually made more mistakes"—because the context was stuffed with Skills, diluting the attention given to the actual requirements.
My Thinking
Superpowers has 14 Skills and over 900 lines of code.
Does an agent really need this many rules?
A project called fable-skills validated one thing: lightweight guidance is completely sufficient. You don't need 14; 6 is enough. For agents that have already evolved to the Fable 5 level, keeping just the brainstorm and review modules is almost enough.
So I thought, could I build something that is:
- Token-efficient (startup < 3K, compared to Superpowers' 22K)
- On-demand activation (each Skill has an independent switch, instead of 14 bundled together)
- Intelligently judges task complexity (small tasks go straight to coding, large tasks follow the process)
And so LightFlow was born.
What is LightFlow
A lightweight AI coding workflow framework, compatible with Claude Code / Codex / Cursor / Trae.
3 core Skills:
1. Brainstorm — Requirement Clarification
Triggers only for complex tasks or vague requirements. Asks a maximum of 3 follow-up questions, each with options for the user to choose from, and stops once things are clear. Does not generate a Spec document.
2. Plan — Implementation Plan
Enabled only when changes involve cross-module/database/API modifications. A 3-8 step implementation plan, marking risk points and rollback strategies. Does not create a Git Worktree.
3. Review — Code Review
Executes automatically after every change. Checks 5 dimensions: type safety, error handling, performance, security, and testing. Does not modify code, only flags issues.
Smart Thresholds
This is the core design. Configured in config/threshold.md:
max_files_changed: 3 # >3 files triggers Plan
max_lines_changed: 100 # >100 lines triggers Plan
risk_level: medium # low/medium/high
Below threshold → Direct coding → Review → Done Above threshold → Brainstorm → Plan → Code → Review
Users can also manually override:
- "Use the full process" → Forces the complete workflow
- "Just change it, don't ask" → Skips everything
Comparison Test
I pulled the latest code from the Superpowers repository for a comparison:
Startup Token Consumption
| Project | Characters | Approx. Tokens |
|---|---|---|
| Superpowers Core Files | 156,180 chars | ~39,045 tokens |
| LightFlow Core Files | 9,778 chars | ~2,444 tokens |
| Savings | 93.7% | 93.7% |
Installation Method
- Superpowers: Install plugin → configure permissions → load Skills (different for each platform)
- LightFlow: git clone → cp CLAUDE.md ./ → done
File Count
- Superpowers: 91 files (including docs, tests, design drafts, specs…)
- LightFlow: 10 files (5 core files)
More comparison details in the README.
Quick Start
git clone https://github.com/leallee00/lightflow.git .lightflow
cp .lightflow/CLAUDE.md ./
# Edit CLAUDE.md to fill in your project's tech stack and conventions
Or install with a single command:
curl -fsSL https://raw.githubusercontent.com/leallee00/lightflow/main/install.sh | bash
It's that simple.
Usage Scenarios
Scenario 1: Rename a variable (Simple)
User: "Change userName to displayName"
→ Direct search → Replace → Review (✅ Passed) → Done
Time: 15 seconds Token cost: Near zero
Scenario 2: Add a new API (Medium)
User: "Add an API to get a user list"
→ Confirm boundaries → Write code → Review (⚠️ Suggests pagination)
Time: 2 minutes
Scenario 3: Refactor payment module (Complex)
User: "Migrate payment from Stripe to Alipay"
→ Brainstorm (Confirm scope)
→ Plan (6-step plan)
→ Code → Review (Check security)
Choosing Between LightFlow and Superpowers
| Scenario | Recommendation |
|---|---|
| AI coding novice, needs full process guidance | Superpowers |
| Veteran/small team, pursuing efficiency | LightFlow |
| Simple project, few changes | LightFlow |
| Large refactor, needs Sub-agent collaboration | Superpowers |
| Troubled by token consumption | LightFlow |
| Chinese-speaking developer | LightFlow (native Chinese) |
What's Missing
- Lacks Sub-agent multi-agent collaboration (on the roadmap)
- Lacks stress test cases
- Lacks Visual Companion visualization capabilities
- Currently only tested by me alone
These need community feedback to gradually improve.
Conclusion
Ultimately, tools serve people, not the other way around. Superpowers provided real value when agents weren't mature enough, but agents have grown up now and don't need so much hand-holding.
LightFlow's idea is simple: a gentle nudge at the right moment is enough.
GitHub Repo: leallee00/lightflow
Stars, Issues, and PRs are welcome. Feel free to share your thoughts in the comments too, and let's polish this tool together 🪶
Published on 2026-07-21