How do I install the Forge skill?

Run npx skillmds@latest add shivae372/forge in your terminal (requires Node.js), paste this page's agent-chat prompt into Claude, Cursor, or any MCP-connected agent, or download the SKILL.md file and copy it into your agent's skills directory.

Is Forge safe to use?

This skill has not completed SkillMD's automated safety review yet. SkillMD never runs a skill's scripts for you; review the SKILL.md before installing.

Which AI agents work with Forge?

This skill is tagged as working with Claude Code, Claude.ai, OpenAI Codex. SKILL.md is an open format, so most agents that read a skills directory can load it too.

Is Forge free to use?

Yes. Installing skills from SkillMD is free, and the skill stays under its author's original license.

shivae372 (@shivae372) published this skill. Their other Agent Skills are listed on their SkillMD profile.

Forge by shivae372 — Agent Skill

Name: Forge
Author: shivae372

Forge

Authors a NEW, project-specific Claude Code skill from scratch when the ecosystem has nothing that fits. Use after `augment` finds no good match, or when the user wants a custom workflow captured as a reusable skill. Produces a detailed, validated skill — never a generic stub. Triggers: 'make a skill for…', 'turn this into a skill', 'I keep doing X, automate it'.

shivae372 85961f2 2.9 KB Updated Sep 17, 2026

File contents

Purpose

Capture a capability as a durable, high-quality skill tailored to THIS codebase. The deterministic scaffolder guarantees structure and validity; you supply the real, specific substance. A generic skill is a failed skill.

When to use this

augment searched the ecosystem and nothing good fit.
The user has a repeatable, project-specific workflow worth making one-command.

Steps

1. Understand the task deeply FIRST

Before writing anything, learn how this project actually does the thing:

Read the relevant files (configs, scripts in package.json/Makefile/pyproject, existing patterns). Use the explorer agent for big repos.
Identify the EXACT commands, file paths, env vars, and success checks involved. Specificity here is the whole point.

2. Scaffold a valid skeleton

python3 .claude/engine/skill_forge.py scaffold \
  --name <kebab-name> \
  --description "<what it does> + Use when <concrete triggers>" \
  --allowed-tools "Bash, Read"

This writes .claude/skills/<name>/SKILL.md with the right frontmatter and the section skeleton (Purpose / When to use / Steps / Output / Guardrails / References), plus scripts/ and references/ dirs.

3. Fill it with REAL detail (replace every placeholder)

Edit the SKILL.md so the Steps are the actual procedure for this project:

Real commands with the real flags (pnpm test --filter web, fly deploy --remote-only…).
Real file paths and expected outputs.
A concrete success check and explicit guardrails (what it must never do).
Long reference material → drop into references/ and link it (progressive disclosure).
If the skill needs a helper, write it into scripts/ and call it from the steps.

4. Validate — the quality gate

python3 .claude/engine/skill_forge.py validate .claude/skills/<name>/SKILL.md

It FAILS if placeholders remain or the body is too thin. Do not finish until it passes.

5. Prove it, then learn

Run the new skill once on a real (safe) case to confirm it works. Then record it:

echo '{"category":"workflow","text":"Forged skill <name>: <one-line what it does>","tags":["forge"]}' \
  | python3 .claude/engine/learn.py add

Output

A validated .claude/skills/<name>/SKILL.md (plus any scripts/), specific to this project, ready to auto-invoke — and a note to the user on how to trigger it.

Guardrails

Never ship a skill that still contains skeleton placeholders (the validator blocks this).
One skill = one job. If it's growing two jobs, forge two skills.
Keep the body lean; push detail into references/.

---name: forgedescription: Authors a NEW, project-specific Claude Code skill from scratch when the ecosystem has nothing that fits. Use after `augment` finds no good match, or when the user wants a custom workflow captured as a reusable skill. Produces a detailed, validated skill — never a generic stub. Triggers: 'make a skill for…', 'turn this into a skill', 'I keep doing X, automate it'.---## PurposeCapture a capability as a durable, high-quality skill tailored to THIS codebase.The deterministic scaffolder guarantees structure and validity; you supply the real,specific substance. A generic skill is a failed skill.## When to use this- `augment` searched the ecosystem and nothing good fit.- The user has a repeatable, project-specific workflow worth making one-command.## Steps### 1. Understand the task deeply FIRSTBefore writing anything, learn how this project actually does the thing:- Read the relevant files (configs, scripts in `package.json`/`Makefile`/`pyproject`, existing patterns). Use the `explorer` agent for big repos.- Identify the EXACT commands, file paths, env vars, and success checks involved. Specificity here is the whole point.### 2. Scaffold a valid skeleton```bashpython3 .claude/engine/skill_forge.py scaffold \ --name <kebab-name> \ --description "<what it does> + Use when <concrete triggers>" \ --allowed-tools "Bash, Read"```This writes `.claude/skills/<name>/SKILL.md` with the right frontmatter and thesection skeleton (Purpose / When to use / Steps / Output / Guardrails / References),plus `scripts/` and `references/` dirs.### 3. Fill it with REAL detail (replace every placeholder)Edit the SKILL.md so the Steps are the actual procedure for this project:- Real commands with the real flags (`pnpm test --filter web`, `fly deploy --remote-only`…).- Real file paths and expected outputs.- A concrete success check and explicit guardrails (what it must never do).- Long reference material → drop into `references/` and link it (progressive disclosure).- If the skill needs a helper, write it into `scripts/` and call it from the steps.### 4. Validate — the quality gate```bashpython3 .claude/engine/skill_forge.py validate .claude/skills/<name>/SKILL.md```It FAILS if placeholders remain or the body is too thin. Do not finish until it passes.### 5. Prove it, then learnRun the new skill once on a real (safe) case to confirm it works. Then record it:```bashecho '{"category":"workflow","text":"Forged skill <name>: <one-line what it does>","tags":["forge"]}' \ | python3 .claude/engine/learn.py add```## OutputA validated `.claude/skills/<name>/SKILL.md` (plus any `scripts/`), specific to thisproject, ready to auto-invoke — and a note to the user on how to trigger it.## Guardrails- Never ship a skill that still contains skeleton placeholders (the validator blocks this).- One skill = one job. If it's growing two jobs, forge two skills.- Keep the body lean; push detail into `references/`.

Frequently asked questions