Plan For Goal

Plan for Goal

Takes any brief - a game idea, a refactor, a client deliverable, a content batch - and turns it into a structured plan you can feed straight into /goal. Works for Claude Code /goal (v2.1.139+) and Codex /goal.

What this does

Strips fluffy intent down to a single verifiable Definition of Done and the sub-checks that prove it. Locks scope (in + out + constraints) so /goal doesn't spiral. Names the verification path explicitly. Sets a turn ceiling.

When to invoke

• You have a brief and want to set it up for autonomous execution
• Before pasting work into /goal, /loop, or a long agent run
• Triggers: /plan-for-goal, plan-for-goal, plan for goal, build me a /goal plan, turn this into a plan

Output format (LOCKED)

Two things every run:

1. A markdown file at ./plans/{slug}_{YYYY-MM-DD}.md (or wherever your project keeps planning docs)
2. A copy-paste-ready /goal one-liner rendered inline in chat

Section order is locked. 9 core sections, 2 optional. Drop a section only when it's empty. Don't add new sections.

# Plan · {Title}

## Brief
[1-2 sentences. Context + outcome. State the goal, not the steps.]

## Stack
- Tool, framework, API, or model that will be used
- One bullet each, name version when load-bearing
- Flag paid APIs that need cost approval

## Scope
**Visuals**
- [Bullets describing what it looks like - only when the task is visual]

**Functionality**
- [Bullets describing what it does]

For non-visual tasks (refactor, content batch, ops), drop the Visuals sub-header and flatten to bullets under Scope.

## Out of Scope
- Explicit non-goals
- At least 2 bullets - empty negative space is where /goal burns tokens

## Constraints
- What must hold throughout the work
- e.g. single file, no build step, no new deps, brand rules, perf budget

## Definition of Done
[One verifiable sentence. The whole plan funnels here. Phrased so a small evaluator can judge it from the transcript. No vibe words.]

## Acceptance Criteria
- Bulleted sub-checks flowing from DoD
- Each independently verifiable as true / false
- Aim for 5-10 bullets

## Verification
- Exact commands or visual checks /goal will run each turn to produce evidence
- Skip ONLY when DoD is already a literal runnable command

## Turn Budget
Stop after {N} turns, or sooner once the DoD condition holds.

## References
- Links, files, screenshots, prior art (optional - skip if empty)

## Risks / Open Questions
- Known failure modes, judgment calls flagged for the user (optional - skip if empty)

The /goal one-liner

After saving the plan, emit a paste-ready /goal line shaped like:

/goal {DoD sentence}, with {top 2-3 acceptance criteria inline}, stop after {N} turns

Engine-agnostic. stop after N turns reads cleanly for both Claude Code and Codex evaluators.

Example (from examples/catan-ai-wars.md):

/goal game.html opens in browser, plays a complete Catan match vs 3 AI bots, declares the winning house at 10 VP, zero console errors during full play, stop after 60 turns

Section rules

Brief - 1-2 sentences. State the goal, not the steps. No verbs of doing ("we will build...").

Stack - bullet per tool. Name version when load-bearing (Node 20, Godot 4, Python 3.11). Flag any paid API (image gen, video gen, eval API) that triggers a cost-approval gate in your workflow.

Scope - Visuals + Functionality sub-headers when the task is visual (game, UI, deck, site, animation, thumbnail batch). Flat bullets for backend / script / content work. Don't force the split.

Out of Scope - at least 2 bullets, always. Most spiraling /goal runs come from undefined negative space. Common cuts: multiplayer, mobile, save-load, accessibility pass, internationalisation, expansions, advanced AI.

Constraints - what must hold throughout. Different from Out of Scope - these are rules during the work, not boundaries of the work. "Single file, no build step" is a constraint. "No multiplayer" is out of scope.

Definition of Done - ONE sentence. Must be judgeable from the transcript. Avoid vibe words ("clean", "good", "polished", "great UX"). Use concrete proof: file exists at path X, command exits 0, count equals N, screenshot matches reference, all rows in table Y are filled.

Acceptance Criteria - 5-10 bullets. Each independently true/false. If a criterion is squishy, fix it now. Bad: "looks good on mobile". Good: "renders at 360px width with no horizontal scroll".

Verification - skip when DoD IS a runnable command (npm test exits 0). Include when DoD is a state ("game plays end-to-end with no errors") that needs a recipe to prove. Spell out the recipe step by step. This is where the agent's eyes go each turn.

Turn Budget - pick a number. Light task 15-25. Medium 30-50. Heavy 60-100. Don't go above 100 without a strong reason. The 500-turn Claude Code safety cap is not a budget.

References / Risks - only include when there's something real. Empty sections are noise. Risks should flag known failure modes and judgment calls worth surfacing before the run.

Voice rules

• No em-dashes. Use - with spaces, or a comma
• Plain language, no marketing voice
• Bullets are short noun phrases or short imperatives, not paragraphs
• Match the brief's domain - game / content / code / ops. Don't reflexively use dev framing for content tasks.

Ambiguity handling

If the brief is missing critical pieces (no clear DoD, no scope boundary, no verification path), ask ONE tight question before drafting. Common gaps:

• DoD vague - ask "what does done look like in one sentence?"
• Scope sprawling - ask "what's explicitly NOT in scope?"
• No verification path - ask "how will we know it works?"

One question max, then draft. Never ask for confirmation on minor details the user can redirect in their next turn.

File operations

• Slug = lowercase, dashes, first 3-5 words of the brief
• Date = today's date, YYYY-MM-DD
• Plan path: ./plans/{slug}_{YYYY-MM-DD}.md (create the folder if missing)
• Report the file path back to the user when done

Anti-patterns

• Don't pad sections with filler. Empty section is fine, weak section is not.
• Don't add new sections beyond the locked set
• Don't write DoDs with vibe words ("clean", "polished", "great UX", "feels nice")
• Don't skip Out of Scope - it's load-bearing
• Don't promise external system interactions unless they were explicitly named in the brief
• Don't ask multiple clarifying questions - one or none

Reference example

See examples/catan-ai-wars.md for a complete worked example: Catan game with Game-of-Thrones-themed AI-company houses, 32-bit pixel art, single-file web.