A repeatable system for building AI coding-skill packages that actually change how an agent writes code.
Not a library of skills. The factory that produces them.
You give your AI coding agent a "skill" (a big markdown file full of best practices), and it nods politely, then writes the exact code you told it not to.
That's not the agent being dumb. It's the skill being built wrong. After months of failed attempts, one package (Rust, 8 skills, zero-to-deployed in a single session) finally worked. Skill Forge is that process reverse-engineered from the one success, so it's repeatable instead of accidental.
| The mistake | What actually happened |
|---|---|
| One giant SKILL.md | 2,000+ lines. The model got "lost in the middle" and ignored the rules buried in the center. |
| No research phase | Rules written from general knowledge: generic, uncited, blind to ecosystem-specific footguns. |
| No anti-rationalization rules | The skill said what to do but not what the model would be tempted to do wrong. The model found the loopholes. Every time. |
| No difficulty layering | Beginner and expert patterns mixed together. The model couldn't tell "always do this" from "maybe, in advanced cases." |
| No provenance | Rules without citations are opinions. Cited rules are verifiable. Models trust verifiable. |
The whole framework lives in SKILLS-FACTORY-FRAMEWORK.md. The short version:
| Phase | Time | What happens |
|---|---|---|
| 0 · Scope | 30 min | Break the language/platform into 6-10 modes of work, each with a distinct AI failure mode. If two modes fail the same way, merge them. |
| 1 · Discovery | 15 min | 3-5 searches for existing rule sets so you're not reinventing wheels (or repeating their mistakes). |
| 2 · Quicksearch saturation | 2-3 hrs | 40-100 targeted searches. Yes, that many. This is where ecosystem-specific knowledge comes from. |
| 3 · Deep research | 30-60 min | 7-10 deep-research prompts, run in parallel. |
| 4 · Writing | 1-2 hrs | Synthesize into layered skills: Level 1 always-works → Level 2 intermediate → Level 3 advanced, plus Performance, Observability, and the secret sauce, Enforcement: Anti-Rationalization Rules. |
| 5 · Promotion | 5 min | Copy the finished drafts into ~/.claude/skills/. |
| 6 · Pressure test | next session | A fresh agent tries to weasel out of the rules. Whatever loophole it finds becomes a new enforcement rule. |
| 7 · Reference population | optional | Backfill citations and reference docs. |
Anti-rationalization rules. Don't just tell the model the right thing to do. Name the wrong thing it's about to rationalize, and forbid that specifically:
❌ "Use proper error handling." ✅ "You will be tempted to
.unwrap()because it compiles and the demo works. Do not. Every.unwrap()in non-test code is a panic waiting for production. Use?or handle the error."
The model can't argue with a rule that already predicted its excuse. Competence is the only credential, and competence here means closing the loopholes before the model finds them.
Pure system, no library:
SKILLS-FACTORY-FRAMEWORK.md: the full 7-phase pipeline, the skill-file template, and the anti-pattern catalogSKILL-FEEDBACK-LOOP.md: how skills keep improving after they shipROADMAP.md: which language/platform packages are nextSTRATEGIC-GAPS.md: the honest list of what's still missing
Anyone maintaining a fleet of AI coding agents who's tired of writing skills that get ignored. Model-agnostic in spirit; the templates lean toward Claude Code and the mx-{lang}-{mode} naming convention, but the method ports anywhere skills are markdown.
A factory that builds the tools that build the code. It's recursion all the way down, and this is the base case.