Skip to content

Adit-Jain-srm/skill-forge

Repository files navigation

Version License Stars Skills Tests

Your agent keeps using the same stale knowledge. These skills fix that — straight from my daily workflow.


The Problem

Your AI coding agent has a fixed set of capabilities. Meanwhile, 20,000+ skills exist on marketplaces, new patterns emerge weekly, and the agent you set up last month is already behind.

You don't have time to browse, evaluate, and install manually. You need an agent that feeds itself.

The Fix

skill-forge is a self-improving intelligence amplifier that:

  1. Hunts — Discovers skills across GitHub, marketplaces, blogs, communities
  2. Devours — Reads deeply, extracts philosophy and novel techniques
  3. Judges — Validates against the absolute best, rejects mediocrity
  4. Acts — Installs what fills gaps, creates what doesn't exist
  5. Ships — Publishes as repos others can install
  6. Evolves — Rewrites its own heuristics based on outcomes

Quick Start

npx skills@latest add Adit-Jain-srm/skill-forge
Claude Code (Plugin Marketplace)
/plugin marketplace add Adit-Jain-srm/skill-forge
/plugin install skill-forge@skill-forge
Cursor
npx skills@latest add Adit-Jain-srm/skill-forge

Or clone to personal skills:

git clone https://github.com/Adit-Jain-srm/skill-forge.git ~/.cursor/skills/skill-forge
Codex CLI
git clone https://github.com/Adit-Jain-srm/skill-forge.git
cp -r skill-forge/skills/* ~/.codex/skills/
Gemini CLI / Copilot / Other
git clone https://github.com/Adit-Jain-srm/skill-forge.git
# Copy skills/ to your agent's skill directory

18 Skills — Each Solves a Named Problem

# Problem Skill What Changes
1 "AI code has hidden bugs" ai-debt-detector Catches orphaned resources, missing error handling, hallucinated packages
2 "Visual edits miss the target" design-mode Precision UI editing with Cursor Design Mode — targeted, minimal, correct
3 "Agent says 'done' but it's NOT" prove-it Forces evidence before any completion claim
3 "Agent shows buggy first drafts" self-review Reviews own code BEFORE presenting
4 "Building the wrong thing" grill Interview yourself until every decision is resolved
5 "Something broken, don't know why" diagnose Forced loop: reproduce → minimise → hypothesise → fix
6 "Agent doesn't speak my language" context-builder Builds shared vocabulary. 50-75% less verbosity
7 "Lost in unfamiliar code" zoom-out Forces perspective before modifying anything
8 "Need to continue later" handoff Compacts session into doc another agent picks up
9 "Long sessions degrade" session-guard Monitors health, splits proactively before corruption
10 "Complex task needs parallel work" dynamic-workflow Fan-out 10-100 subagents with verification gates
11 "Agent uses one tool, should chain 5" mcp-conductor Multi-MCP orchestration pipelines
12 "Git history is unreadable" git-workflow Enforced discipline: atomic commits, PR readiness
13 "Site is slow, don't know why" web-perf Measure-first loop: baseline → identify → fix → re-measure
14 "Code breaks in production" error-resilience Retry, circuit breakers, graceful degradation
15 "Nobody knows the architecture" arch-from-code Generates diagrams from actual codebase analysis
16 "Database breaks at scale" db-schema Queries-first discipline before every DDL
17 "First-time setup confusing" setup One-time preference config all skills consume

The Meta-Agent

The root SKILL.md is itself an intelligence amplifier:

skill-forge              # Full pipeline (discover → judge → act → learn → improve)
skill-forge route <task> # "What skill solves this?" → best match + invocation prompt
skill-forge create       # Validate-against-best → Create → Ship → Promote
skill-forge devour       # Maximum: learn everything, apply everything, faster each cycle
skill-forge guide <proj> # Apply ALL learnings to your project
skill-forge hackathon    # Creative ideation with latest approaches
skill-forge promote      # Advertise published skills across platforms
skill-forge improve      # Force self-check + meta-optimize
skill-forge status       # Metrics for both objectives

Design Philosophy

"The rate of feedback is your speed limit." — The Pragmatic Programmer

"Invest in the design of the system every day." — Kent Beck

"The best modules are deep: a lot of functionality through a simple interface." — John Ousterhout

Applied to skills:

  • Each solves a NAMED PROBLEM (not "a tool that does X")
  • Behavioral change > reference material
  • Self-improvement is mandatory, not optional
  • NEVER ship without validating against the absolute best
  • The internet is the perimeter. NO LIMITATIONS TO LEARNING.

Quality Proof

✓ 101 automated tests (structure, skills, scripts, memory, routing, skillopt, orchestration, quality)
✓ 17/17 skills pass validation (0 warnings)
✓ TF-IDF semantic routing with compound skill matching
✓ SkillOpt self-improvement loop (data-driven prompt optimization)
✓ Project orchestration (7-phase guided plan generation)
✓ CLI with 8 commands (discover, validate, test, status, self-check, publish, monetize, route)
✓ Multi-platform: Cursor, Claude Code, Codex CLI, Gemini CLI, Copilot
✓ Evals directory with behavioral proof scenarios
✓ CONTEXT.md shared vocabulary
✓ CI-ready (run: node tests/run-tests.js)

Project Structure

skill-forge/
├── SKILL.md                 # Intelligence amplifier (meta-agent)
├── CONTEXT.md               # Shared vocabulary
├── AGENTS.md                # Agent instructions + script reference
├── CLAUDE.md                # Claude-specific context
├── skills/                  # 16 behavioral skills (0 validation warnings)
├── scripts/                 # 13 automation scripts
│   ├── route-task.js        # TF-IDF semantic routing + compound matching
│   ├── index-memory.js      # Category/keyword indexed retrieval
│   ├── skillopt.js          # Self-improvement loop (SkillOpt)
│   ├── orchestrate.js       # End-to-end project orchestration
│   ├── discover.js          # Multi-strategy skill discovery
│   ├── self-improve.js      # Weakness detection + suggestions
│   ├── rl-update.js         # Reinforcement learning signals
│   ├── validate-skill.js    # Quality validation
│   ├── analyze.js           # Deep skill analysis
│   ├── install.js           # Skill + MCP installer
│   ├── scaffold-repo.js     # New skill repo scaffolding
│   ├── publish.js           # Marketplace publishing
│   └── capafy-publish.js    # Capafy monetization
├── bin/                     # CLI (npx skill-forge <cmd>)
├── memory/                  # Persistent state (indexed, RL, skillopt)
├── evals/                   # Behavioral proof scenarios
├── tests/                   # 98-test validation suite
├── docs/                    # Setup guides + skill anatomy
├── .cursor-plugin/          # Cursor marketplace manifest
└── .claude-plugin/          # Claude marketplace manifest

Contributing

See CONTRIBUTING.md. Key rule: every skill must solve a named problem, not describe a feature. See docs/skill-anatomy.md for the quality bar.


Changelog

See CHANGELOG.md for version history.


Star History

Star History Chart


License

MIT — See LICENSE


About

18 AI agent skills that catch what your agent misses: hidden code debt, false completions, missing error handling, context amnesia, imprecise UI edits. Self-improving with TF-IDF routing + SkillOpt loop. Cursor, Claude Code, Codex, Copilot. Install: npx skills@latest add Adit-Jain-srm/skill-forge

Topics

Resources

License

Contributing

Stars

Watchers

Forks

Packages

 
 
 

Contributors