Skip to content
Back to skills

Goal Prompt

ASecurity

Turn a task, plan, or feature request into a ready-to-paste Claude Code /goal command — a single completion condition with a measurable end state, a demonstrable proof, and the constraints that must not drift. Use when the user says "give me a goal", "goal prompt", "make this a /goal", "turn this into a goal", or wants an autonomous long-running objective for Claude Code.

  • 129 stars
  • 0 votes
  • 0 copies
  • 1 view
  • Added August 31, 2026
ai-agentspythongo

Works with

  • claude code
  • cli

Security analysis

A100/100

Scanned August 31, 2026

npx -y skills add techwolf-ai/ai-first-toolkit --skill goal-prompt --agent claude-code

Installs into .claude/skills of the current project.

Are you the author of Goal Prompt?

Add the live security badge to your README. It updates with every re-scan.

Security grade badge for Goal Prompt
[![Security: A — Skills Directory](https://www.skillsdirectory.com/api/skills/techwolf-ai-goal-prompt/badge)](https://www.skillsdirectory.com/skills/techwolf-ai-goal-prompt)

More formats (shields.io, HTML) on the badges page. Keep it an A: scan every change in CI with Pro.

Download with Pro
SKILL.md
---
name: goal-prompt
description: Turn a task, plan, or feature request into a ready-to-paste Claude Code /goal command — a single completion condition with a measurable end state, a demonstrable proof, and the constraints that must not drift. Use when the user says "give me a goal", "goal prompt", "make this a /goal", "turn this into a goal", or wants an autonomous long-running objective for Claude Code.
---

# goal-prompt

Produce a ready-to-paste `/goal ...` command from whatever the user is trying to accomplish.

## Why the shape matters
`/goal` runs Claude autonomously until a separate fast-model evaluator decides, after a turn, that the condition is met. The evaluator does NOT run commands or read files itself — it only reads Claude's output. So the completion condition must be demonstrable by Claude's own output, never by hidden side effects.

## A good goal has three parts
1. **Measurable end state** — one concrete finish line: a test/exit code, a file that must exist, a count, an empty queue, named sections present.
2. **Stated proof** — exactly how Claude demonstrates it: the command to run and its expected result, or the grep/check whose output shows done. Phrase it as "Prove it by showing X."
3. **Constraints that must not drift** — what stays unchanged on the way there: files not to touch, framing to keep, no network/prod, don't modify tests.

## How to write it
- One sentence of objective, then `Done when: <end state + proof>`, then `Constraints that must not change: <list>`.
- If the full spec is long, point to a plan/doc file (e.g. a path under `~/.claude/plans/` or `docs/`) and keep the goal itself scannable.
- Make the proof something the transcript can show: prefer `command exits 0` + a summary line, or `grep` for markers, over vague "it works".
- Translate conditions the evaluator can't see ("the UI looks good") into an observable check.
- Keep constraints tight enough to stop scope creep, not so rigid they block the obvious path.

## Output
Give the user a single fenced block starting with `/goal`, then 2-3 lines explaining the end state, the proof, and why it's demonstrable. Nothing else.

## Example
```
/goal Add a --json flag to the export CLI per docs/export-json.md. Done when: `pytest tests/test_export.py -q` exits 0 and `python -m app.export --json` prints valid JSON whose top-level keys include "rows" and "meta". Prove it by showing the pytest summary and the piped `... --json | jq keys` output. Constraints that must not change: only edit app/export.py and add tests/test_export.py; do not alter the existing CSV output path; no network.
```
Its finish line is a passing test plus a schema check, both visible in Claude's transcript; the constraints pin the blast radius so the autonomous run can't wander.

Attribution

Is this your skill, or is something wrong with this listing? Request removal or report an issue. Author removals are honored within 72 hours.

Comments

Loading comments…

Newsletter

Get the weekly list of safe new skills, and the ones we pulled.