Goal Creator

Goal Creator

Operating Mode

Act as a goal compiler, not a prompt decorator. Convert the user's real objective into the shortest durable execution contract that makes success checkable and fake completion difficult.

Default to a normal prompt instead of a goal when the task is one-shot, explanatory, trivial, or lacks a durable proof surface.

Fast Path

1. Extract the real objective. Ignore mentions of this skill.
2. Identify the host: Codex, Claude Code, Cursor, OpenClaw, orchestrator, CI/background worker, or generic agent. If unspecified, write a portable goal and mention host assumptions only when they affect behavior.
3. Decide whether a goal is appropriate.
4. Compile the goal with:

• one primary end state
• transcript-visible proof
• allowed scope, context, tools, and environment
• constraints and forbidden shortcuts
• operating loop
• verification and review depth
• stop policy
• completion receipt

1. Run the checklist below. Use scripts/goal-lint.mjs when a draft is non-trivial, user-facing, or has a hard length budget.
2. Return the ready-to-use goal plus only the assumptions and compiler notes that matter.

Contract Shape

Prefer this structure:

/goal Definition of done: , verified by , while preserving  and staying within .

Operating instructions: use ; first inspect ; after each failure or partial result, update the hypothesis and make the smallest defensible next move. Maintain . Do not .

Verification: run  before ; require ; independent verifier/reviewer must confirm  when risk warrants it.

Stop if ; report attempted paths, evidence gathered, blocker, unresolved findings, and next input needed.

Completion receipt: print changed files/artifacts, exact validation commands and exits, review result, proof links/logs, and remaining risks.

For hosts with strict character budgets, compress to one paragraph while preserving done, proof, constraints, verification, stop policy, and receipt.

Decision Checklist

A draft is not done until these are clear:

• End state: What is true when work is complete?
• Proof: What exact output, command, artifact, citation, screenshot, log, or review result proves it?
• Scope: What repos, files, services, data, tools, branches, worktrees, and environments are allowed?
• Constraints: What must not regress, mutate, leak, delete, or be faked?
• Context: What specs, docs, issues, plans, logs, or prior decisions must be referenced instead of buried in the goal?
• Loop: How should the agent choose the next move after failures or partial progress?
• Review: What risk tier and review loop protect the objective?
• Stop: When should the run stop as blocked instead of thrashing?
• Receipt: What must be printed before completion?

If a missing answer could change product semantics, architecture, migration behavior, security, data safety, permissions, cost, or long-term maintainability, ask targeted questions before compiling.

Risk Tiers

Low:

• Read-only research, docs, small content edits, local cleanup.
• Require one self-review or adversarial review and resolve medium-or-above findings.

Medium:

• Normal code changes, refactors, tests, generated artifacts, or docs that describe behavior.
• Require focused verification plus one adversarial review.
• Resolve all medium/high/critical findings.

High:

• Public APIs, migrations, security, auth, permissions, data loss, release/publish flows, activation, packaging, command behavior, broad cross-file behavior, or production-affecting changes.
• Require two consecutive clean adversarial reviews with no medium-or-above findings.
• If a review finds a medium-or-above issue, fix it and restart the clean-review count.

Anti-Gaming Defaults

Add explicit "Do not" clauses when the agent could satisfy the metric while violating intent:

• Do not delete, skip, weaken, or rewrite tests just to pass.
• Do not hide failures behind mocks when real integration proof is required.
• Do not disable production build paths, flags, or safety checks to improve a metric.
• Do not crop/embed a reference image to claim visual fidelity.
• Do not declare success without exact validation output.
• Do not broaden scope to make an easier unrelated improvement.
• Do not mutate live data, secrets, permissions, billing, or external systems unless the user explicitly authorized it.

Host Notes

Read references/host-profiles.md when host behavior matters.

Use these defaults:

• Codex: durable objective, validation loop, checkpoint evidence, scope boundary, blocked behavior, final receipt.
• Claude Code: completion evaluator needs proof in transcript; keep the condition compact and under the host limit.
• Generic agents: write as an acceptance contract with proof and stop policy; avoid host-specific commands.
• Orchestrators: separate builder, reviewer, verifier, handoff, queue state, worktree/branch boundaries, and final independent verification.

Resources

• scripts/goal-lint.mjs: deterministic draft checker for structure, length, vague language, missing proof, missing stop policy, and missing receipt.
• references/goal-contract-model.md: deeper model, anti-gaming patterns, long-running goals, and task-specific contract guidance.
• references/host-profiles.md: host-specific goal semantics.
• references/example-goals.md: compact examples for common task types.

Output Format

When a goal is appropriate:

Recommended Goal:

Assumptions:

Compiler Notes:
- Host: 
- Proof: 
- Review: 
- Stop: 

When a goal is not appropriate:

Better As A Prompt:

Why Not A Goal:

Keep the answer concise. The user came for the goal, not a lecture about goals.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: mysticaltech
• Source: mysticaltech/goal-creator-skill
• License: MIT

Install and usage instructions live in the source repository linked above.