Skip to main content
skill CX Codex

mega-goal-prompt — Claude Code skill: what it does, how to install

mega-goal-prompt interviews you step by step to build a structured /goal mega prompt for autonomous tasks in Claude Code or Codex CLI.

MIT 195 installs ★ 635 updated 2026-09-29

Install into Claude Code · Skills CLI (Vercel) All methods and platforms ↓
npx skills add agentara/skills --skill mega-goal-prompt -a claude-code
Open on GitHub

What mega-goal-prompt does

When delegating autonomous work to an agent, a vague prompt leads to drift. This skill conducts an interactive interview to extract what completion looks like before any execution starts. It asks one question at a time and offers a recommended default for each step based on prior answers.

The process covers six branches, starting with a falsifiable outcome, project context, and 3 to 5 provable success criteria. Without resolving these initial three branches, the skill refuses to generate output. Once complete, it outputs a ready-to-paste /goal block containing rules, constraints, and required verification artifacts.

What it helps with

  • define a falsifiable end goal for autonomous agent runs
  • set concrete success criteria verified by test runs or command outputs
  • specify off-limits directories and actions that require human approval
  • build a complete context specification before starting a long refactor or migration

Example prompts for the agent

  • Help me write a goal for migrating the database
  • Set a /goal to build this feature end-to-end
  • I want Claude to keep working until all tests pass

How to install mega-goal-prompt

1Skills CLI (Vercel)

npx skills add agentara/skills --skill mega-goal-prompt -a claude-code

2Manually: put the files in place

~/.claude/skills/mega-goal-prompt/SKILL.md

Copy the whole skill folder; for one project use .claude/skills/mega-goal-prompt/

Step by step, with commands for each platform: How to install skills in Claude Code: five ways

Questions about mega-goal-prompt

How to install mega-goal-prompt in Claude Code?

Run npx skills add agentara/skills --skill mega-goal-prompt -a claude-code.

What does mega-goal-prompt do?

It interviews you one question at a time to generate a complete /goal prompt for long-horizon autonomous tasks.

Can you skip questions during the interview?

The outcome, context, and success criteria branches are strictly required. Branches covering constraints, quality bars, and deliverable proofs can fall back to defaults if you ask to move quickly.

Source text

SKILL.md · MIT
---
name: mega-goal-prompt
description: >
  Generate a /goal mega prompt for Claude Code or Codex CLI by interviewing the user about their task.
  Use when the user wants to define a long-horizon autonomous goal — migration, refactor, feature build,
  optimization loop, test fixing, research project, learning system, or any task where the agent should
  run end-to-end without hand-holding. Trigger on: "help me write a goal", "I want Claude to keep working
  until...", "run this autonomously", "set a /goal", or any request that implies sustained agentic execution
  toward a non-trivial outcome. The skill conducts a structured interview (one question at a time) to extract
  outcome, context, success criteria, constraints, and quality bar — then outputs a filled-in mega prompt
  ready to paste into Claude Code or Codex.
---

# Mega-Goal Prompt Generator

Interview the user about their task, one question at a time, until you have everything needed to fill in the `/goal` mega prompt template. Then output the completed prompt.

## The Mega Prompt Template

Read `references/mega-prompt-template.md` to internalize the full template structure before starting. You will fill this in at the end.

## Interview Protocol

Ask **one question at a time**. Wait for the answer before continuing. For each question, offer your own recommended answer based on what you know so far — the user can accept, tweak, or override it.

Work through these branches in order, but skip questions that are already answered from context:

### Branch 1 — The Outcome
What is the single final outcome? What does "done" look like in one sentence?

This becomes the `/goal` line. It must be falsifiable — something the evaluator model can confirm from the transcript.

### Branch 2 — Context
Fill in the CONTEXT block:
- **Project**: what are you building?
- **Stack**: languages, frameworks, infra?
- **Current state**: what exists today — is this greenfield, or modifying existing code?
- **Working dir**: path or repo name?
- **Constraints**: anything off-limits, budget or time constraints?
- **Audience**: who is this for?

Ask about context fields that are missing or ambiguous. If the user has described the project already, confirm rather than re-ask.

### Branch 3 — Success Criteria
What are the 3–5 measurable conditions that must ALL be true for the goal to be met?

Each criterion must map to something the agent can produce as proof in the transcript (command output, file count, test result, URL). Push back on vague criteria like "it works" — ask what command proves it.

### Branch 4 — Constraints and Off-Limits
What must the agent NOT do or touch?
- Files or directories that must stay unchanged
- Actions that require human approval (commits, deploys, etc.)
- Any tool restrictions (no internet, no DB writes, etc.)

### Branch 5 — Quality Bar
Adjust the quality bar to fit the task:
- Is there a UI? (if no, drop the design line)
- What coding conventions matter?
- Is this internal tooling or production-facing?

### Branch 6 — Final Deliverable Proof
What should the agent produce at the end to confirm completion?
- Default: test output + file list + how to run
- Ask if a screenshot, URL, or specific artifact is needed

## NEVER

- NEVER ask multiple questions in one turn — one question, then wait.
- NEVER accept "I'll figure that out later" — require a decision or explicitly mark it as an open question before moving on.
- NEVER let the user redirect to implementation details before all branches are resolved.
- NEVER output the mega prompt until all blocking branches (1, 2, 3) are resolved. Branches 4–6 can use sensible defaults if the user wants to move fast.

## When to Stop Interviewing

Stop and output the prompt when:
- Branches 1–3 are fully resolved, AND
- The user says "go" / "looks good" / "that's enough", OR
- You've asked about all branches and have reasonable answers for everything

## Output

Once done interviewing, output the completed mega prompt as a code block, ready to paste directly into Claude Code or Codex CLI.

After the code block, add a one-line summary of any open questions or assumptions you made.

---

## References

- `references/mega-prompt-template.md` — the full template to fill in

© agentara · MIT · agentara/skills

Similar skills

All skills for Agent setup: memory, context, orchestration →

From the same set20

The whole set →
video-poster-design video-poster-design creates cinematic poster concepts and final images from a brief, storyboard, or project context. Useful for key art, teaser posters, and video campaigns. CX Codex +1 271installs presentation-design presentation-design creates a 6-slide presentation storyboard as a single 3:2 composite image. Useful for concept validation and pitch deck visuals via image generation models. CX Codex +1 250installs article-to-html article-to-html turns Markdown or a document from the conversation into one self-contained HTML file. Use it for a static article page, not slides or a backend. CX Codex +1 220installs doctor-strange doctor-strange runs multi-branch causal scenario simulations using subagents and records synthetic projections into persistent assistant memory. CX Codex +1 212installs soviet-storybook-grotesque soviet-storybook-grotesque transforms photos into strange retro storybook drawings with shaky ink lines, faded watercolor, and nonsensical English rhymes. CX Codex +1 190installs feature-dev-loop feature-dev-loop orchestrates multi-step feature development: planning, subagent reviews, TDD execution, and HTML acceptance reports. CX Codex +1 187installs publish-research-site The publish-research-site skill turns a topic or thesis into an interactive, image-rich site backed by citations, then deploys it with Vercel CLI. Use it to publish a research report or an interactive story. CX Codex +1 142installs torn-paper-collage-poster Generates image-generation prompts and workflows for editorial torn-paper collage posters with ripped edges, tape, and stamps. CX Codex +1 135installs

Project sponsors

Companies that keep this analytics open