Description

What It Does

AI agents hallucinate API calls. They guess function names, invent parameters, and produce code that looks right but doesn't compile. Skill Forge fixes this by extracting real signatures from your code repositories, documentation websites, and developer discourse and compiling them into provenance-backed instruction files that agents can follow without making things up.

The output follows the agentskills.io open standard and works across Claude, Cursor, Copilot, and other AI agents.

Key Features

• AST-Backed Extraction: Parses real function signatures, types, and patterns using ast-grep — not string matching, not guessing. Falls back to source reading when AST isn't available.
• Full Skill Lifecycle: Create, test, export, audit for drift, and smart-update skills as source code evolves — preserving your manual additions through every regeneration.
• Progressive Capability: Works with zero external tools (Quick tier). Add ast-grep for structural truth (Forge). Add cocoindex-code for semantic discovery on large codebases (Forge+). Add QMD for temporal context from issues and changelogs (Deep).
• Architecture Verification: Cross-reference generated skills against your architecture docs to get a feasibility verdict — before writing a single line of code.

12 Workflows, One Agent

Skill Forge ships 12 structured workflows driven by Ferris, a single AI agent that adapts its behaviour to the task:

Before vs After

Without SKF — your agent guesses:

import cognee

# Agent hallucinates: sync call, wrong parameter name, missing await
results = cognee.search("What does Cognee do?", mode="graph")

With SKF — your agent reads the verified skill:

import cognee

# Agent follows the skill instruction:
# `search(query_text: str, query_type: SearchType = GRAPH_COMPLETION) -> List[SearchResult]`
# [AST:cognee/api/v1/search/search.py:L26]
results = await cognee.search(
    query_text="What does Cognee do?",
    query_type=cognee.SearchType.GRAPH_COMPLETION
)

The skill told the agent the real function name, the real parameters, and that the call requires await — all traced to the exact source line. This is from a real generated skill.

Every instruction in the skill links back to the exact file and line it came from. If it can't be cited, it doesn't ship.

Ideal Users

• Developers whose AI agents keep hallucinating API calls — get verified instructions from the actual source code
• Library maintainers who want to ship official, machine-readable skills alongside their package
• Teams with complex dependency stacks that need to understand how everything integrates before writing code
• Architects onboarding large codebases who need to identify what's skillable and verify feasibility
• Teams maintaining skills over time as libraries ship breaking changes — audit, update, re-test

How It Works

1. Install — npx bmad-module-skill-forge install sets up the forge and generates IDE commands
2. Brief — Scope what you want to extract (or skip this with Quick Skill)
3. Create — Ferris analyses your source with AST parsing, enriches it with usage patterns, and compiles a SKILL.md
4. Test & Export — Validate quality, generate token-optimised context snippets, and inject into your agent config
5. Maintain — Audit for drift when source changes, smart-update while preserving your manual edits

Further Resources

• Getting Started — Installation and first skill
• How It Works — Architecture and design
• Workflows — All 12 workflows explained
• agentskills.io — The open standard SKF targets