v0.4 / MIT / Almaty, May 2026

Stop
trusting
the prompt.

43 opinionated rules for AI coding agents. 26 of them ship with a shell script that fails when the agent ignored the rule.

Get the kit → Try the workbench ↓
43
SKILLS
26
VERIFIERS
14
DOMAINS
MIT
LICENSE

Paste
bad code.
Watch it
break.

PICK A SKILL
Live playground

Above is a static preview.
Below the link is the real thing.

The playground loads the actual TypeScript compiler in your browser and runs the same six AST verifiers that ship with the kit. Paste your own code, pick a filename, hit Run. Violations come back with the exact line and column the AST reports. Nothing leaves your browser.

Open the playground →
WHY THIS EXISTS

A directory is not a discipline.

"Awesome" lists hand you 200 unranked links. Other skill packs hand you a long prompt and ask you to trust it.

Forge picks the hill, writes the rules, and ships a check that runs. After the agent edits a file, run ./verify/forge run --diff. You see, in plain text, which rules were ignored and on which line. That output feeds straight back into the agent for a second pass.

Not every domain is grep-checkable. Where it is, Forge greps. Where it is not, Forge says so.

THE KIT

Fourteen domains.
One install.

Each skill is a folder with a SKILL.md and a verifier where applicable. Aesthetic registers are mutually exclusive: pick exactly one of minimalist, brutalist, or soft. Everything else stacks.

Design

Frontend taste, three mutually-exclusive style registers, and the meta-skill for upgrading existing apps without breaking them. forge-frontend · forge-frontend-nextjs · forge-minimalist · forge-brutalist · forge-soft · forge-redesign
06 / 02 VERIFIED

Backend

REST and RPC discipline, TypeScript hard rules, error handling, structured logs, auth, validation. forge-api-design · forge-typescript · forge-error-handling · forge-logging · forge-auth · forge-validation
06 / 06 VERIFIED

LLM apps

For when the AI itself is the product. Prompts, RAG, evals, streaming, tool calling. forge-prompt-engineering · forge-rag · forge-evals · forge-llm-streaming · forge-tool-use
05 / 03 VERIFIED

Infrastructure

Dockerfiles under 200MB. Kubernetes manifests that survive a 3am page. GitHub Actions that don't leak secrets. Caddyfiles that aren't just default. forge-dockerfile · forge-kubernetes · forge-github-actions · forge-caddy
04 / 04 VERIFIED

Dev workflow

Reviewable PR descriptions. Commit messages that survive git blame in five years. Naming. Code review itself. forge-pr-description · forge-commit-messages · forge-code-review · forge-naming
04 / 02 VERIFIED

MCP & multi-agent

Designing servers and tools an LLM uses well. When to spawn subagents, how to brief them, how to check what they returned. forge-mcp-server · forge-mcp-tool-design · forge-mcp-resource · forge-multi-agent · forge-agent-prompt · forge-subagent-eval
06 / 02 VERIFIED

Research

Doing it honestly. Methodology, hygiene around web search, citations that actually exist. forge-research · forge-web-search · forge-citation
03 / 01 VERIFIED

Image generation

Reference comps for websites, mobile, brand kits. Avoiding the floating-iPhone-on-purple-gradient default. forge-imagegen-web · forge-imagegen-mobile · forge-brandkit
03 / 00 VERIFIED

Data

Migrations that don't lock the users table at peak. Schemas that don't leak row counts. forge-migrations · forge-schema-design
02 / 02 VERIFIED

The small ones

Each is one skill, but each is unambiguously its own category. forge-secrets · forge-readme · forge-full-output · forge-tests
04 / 04 VERIFIED
INSTALL

There is no install, really.

Clone the repo and copy whichever SKILL.md you want into the place your agent reads from. For Claude Code, that's usually .claude/skills/. 

# clone, copy, done
$ git clone https://github.com/f4rkh4d/forge-skill
$ cp forge-skill/skills/design/forge-frontend/SKILL.md .claude/skills/

Or use the skills CLI:

$ npx skills add https://github.com/f4rkh4d/forge-skill

After your agent edits a file, check the work:

$ ./verify/forge run forge-api-design path/to/handler.ts
$ ./verify/forge run forge-api-design --diff
$ ./verify/forge run-all path/to/changed/file

Note. The verifiers are intentionally simple - shell, grep, ripgrep. The plan is to move the top half-dozen to tree-sitter AST checks; for now they catch the common cases and are honest about being heuristic.