Goal Creator Skill

goal-creator turns plain-language work into concise, verifiable agent goal contracts.

It is designed for Codex, Claude Code, Cursor, OpenClaw, orchestrators, CI/background workers, and generic coding or research agents. The skill helps decide whether a task deserves a /goal at all, then compiles the strongest practical contract: done state, proof, constraints, operating loop, review depth, stop policy, and completion receipt.

Install

Global install:

npx skills add mysticaltech/goal-creator-skill -g
npx skills update -g -y

Install only this skill explicitly:

npx skills add mysticaltech/goal-creator-skill -g --skill goal-creator -y

Use without installing:

npx skills use mysticaltech/goal-creator-skill@goal-creator

Use

Ask your agent:

Use $goal-creator to turn this into a Codex goal: migrate the billing module to the v2 API without breaking current billing behavior.

Expected shape:

Recommended Goal:
/goal Definition of done: ...

Assumptions:
...

Compiler Notes:
- Host: Codex
- Proof: focused billing tests plus typecheck/build
- Review: high
- Stop: API behavior or verification blocked

What Makes It Different

Most goal prompts are too vague, too long, or too easy for an agent to game. This skill treats a goal as an execution contract:

• one durable end state
• transcript-visible proof
• explicit scope and constraints
• anti-gaming rules
• realistic verification environment
• risk-based review depth
• blocked/stop behavior
• completion receipt

If a task does not need durable execution, the skill returns a better normal prompt instead of forcing /goal.

Acknowledgements And Sources

This skill is a synthesis and implementation of public goal-mode practices, with credit to:

• Dominik Kundel, @dkundel, for A guide to /goal, especially the framing of goals as verifiable exit criteria with measurable progress, realistic environments, visual-task cautions, progress tracking, and cleanup/review before finalizing.
• Shubham Saboo, @Saboo_Shubham_, for The ultimate guide to /goal, especially the framing of /goal as an agent primitive, the builder/reviewer/orchestrator split, independent verification, and safe parallelism boundaries.
• Peter Petrash, @petekp, for the open-source write-goal skill, especially the host-aware compiler model, goal anatomy, risk-based review depth, and "better as a prompt" fallback.

The implementation here is intentionally generic: it preserves those core ideas while targeting any agent or orchestrator that benefits from durable done criteria, proof, constraints, stop policy, and completion receipts.

Lint A Goal Draft

The skill includes a zero-dependency Node.js checker:

node goal-creator/scripts/goal-lint.mjs --text "/goal Definition of done: ..." --max-chars 4000

JSON output:

node goal-creator/scripts/goal-lint.mjs goal.txt --json

The linter checks for done state, proof, scope, constraints, operating loop, verification/review, stop policy, completion receipt, vague language, and length.

Repository Layout

goal-creator/
  SKILL.md
  agents/openai.yaml
  references/
    goal-contract-model.md
    host-profiles.md
    example-goals.md
  scripts/
    goal-lint.mjs

License

MIT