prompts-to-workflow

An agent skill that turns a pile of prompts you keep pasting into a workflow, then decides how to split it into skills. Most prompt batches aren't one skill, and they aren't one skill per prompt either. They're a chain: research, then draft, then check, then publish. This skill finds the chain first. Then it checks every step against the live skillsmith.tech catalog and reuses the skills that already exist (tune-and-polish, wordsmith, findable and the rest). It writes a thin chain skill that runs the steps in order, and briefs a new skill only where nothing in the catalog covers a step.

It has two confirmation points: the workflow map, and the split. No files are written on a split you haven't seen.

Install

npx skills add arm-pz/prompts-to-workflow --skill prompts-to-workflow

Or upload prompts-to-workflow.zip in your skills settings, or copy the folder into ~/.claude/skills/ (Claude Code), ~/.qoder/skills/ (Qoder) or your agent's skills directory, then start a new session.

Needs Python 3 for scripts/catalog.py (standard library only; it reads the live index, falling back to installed skills when offline). Install tune-skill to validate the chain skill it writes, and prompts-to-skill or draft-skill to build the new skills it briefs.

Try it

• "Here's my prompt log from launching my last three skills. Make it a workflow."
• "[pastes 10 prompts] Which of these are already skillsmith skills, and what's missing?"
• "I do these seven steps by hand every Monday. Can it be one thing?"

What you get

• workflow.md: the workflow map (the steps, what each one receives and hands on, its gates, which prompts showed it), the split table with evidence for each verdict, and install lines.
• <chain-name>/SKILL.md: a chain skill that runs the steps and stops at each skill's own gate.
• briefs/<new-skill>.md: one per genuinely new skill, ready for prompts-to-skill.

Boundaries

• Handles: a batch of prompts that together do one multi-step job, turned into a workflow plus a split.
• Does not handle:
  • prompts that each become their own skill (prompts-to-skill)
  • a single idea (draft-skill or skill-launch)
  • an outcome described without prompts (caffeinity-routing)
  • an existing skill (tune-and-polish or tune-skill)
  • packaging what it built (prompts-to-skill-release)
• If the chain is already a brew (for example "crawl, fix keywords, rewrite copy" is Brew 4 · Cold brew), it says so, hands off to caffeinity-routing, and writes nothing new.

Test results

Version 0.1.0:

• Trigger cases: 0 pass, 0 fail, 12 reviewed-only, 0 not run. All 12 were judged against the descriptions; none was run in a live session.
• catalog.py: executed in all three modes.
• Worked example: an 8-prompt brand-identity pack was run by hand. It produced 8 steps, 2 reused skills, 5 new skills and a chain skill that passes the validator.

Version

See CHANGELOG.md. MIT licence.