tq-forge

🔨 tq-forge

An autonomous skill + agent factory for Claude Code.

Forge new slash-command skills and multi-file agents from a one-line intent — scored, sandboxed, and promoted into production without leaving your terminal.

What is this?

Claude Code lets you define skills (/your-command) and agents (personas with their own system prompt). Writing good ones by hand is fiddly: you have to remember the file shape, the frontmatter, the section structure, and you have no feedback on whether what you wrote is any good.

tq-forge turns that into a one-liner. You describe what you want; it decides whether you need a skill or an agent, writes it to a sandbox, scores it on structural quality (0–10), and only lets you promote it once it passes. It's a small, opinionated factory for your own Claude Code toolkit.

You:  /tq-forge a skill that summarizes my open GitHub PRs every morning

      → classified: skill
      → slug: github-pr-digest
      → written to ~/.tq-forge/sandbox/forged-skills/github-pr-digest/
      🟢 skill github-pr-digest: 8.5/10 (excellent)
         clarity            9  █████████░
         actionability      9  █████████░
         completeness       9  █████████░
         specificity        7  ███████░░░
      → Next: /tq-forge-promote github-pr-digest

Why you'd want it

• No more blank-page skill writing. Describe the workflow; get a structured, conventions-following SKILL.md.
• Quality gate built in. A heuristic scorer checks clarity, actionability, completeness, and specificity. Agents add consistency + handoff clarity. Nothing reaches production below 7/10.
• Sandbox → promote workflow. Everything starts in ~/.tq-forge/sandbox/. You review and score before a single file touches your live skills dir.
• Eight agent archetypes. researcher, coder, business-analyst, ops-manager, scraper, reviewer, sales-agent, custom — each with a rigid output format and a {{CONTEXT}} slot for your own domain.
• Self-curating. /tq-forge-improve rewrites only the weak sections; /tq-forge-scan finds workflows you keep doing by hand that should be skills.

How it compares

Most ways to make a Claude Code skill are either blank-page guesswork or a one-shot LLM generation you can't measure. tq-forge puts a deterministic quality gate between "generated" and "live."

The thing nothing else here ships is the token-free deterministic scorer: every artifact gets the same 0–10 structural read, every time, for free — and nothing reaches ~/.claude/skills/ until a re-score passes the gate.

It scores structure, not correctness — it catches missing sections and vagueness, not logic bugs. Always read what you promote.

Install

npx tq-forge install

One command. It copies the 13 skills into ~/.claude/skills/, drops the support scripts in ~/.tq-forge/install/, seeds your state home, and checks for bash/python3. Restart Claude Code and you're done.

npx tq-forge doctor      # check the install + deps
npx tq-forge uninstall   # remove skills (keeps your forged data)

/plugin marketplace add tanishq286/tq-forge
/plugin install tq-forge@tq-forge