make-a-goal

/make-a-goal — autonomous-run mega-prompt builder

What it does

/make-a-goal [outcome line] interviews Stevie for the operating context an autonomous coding agent needs, then writes a filled mega-prompt to .agents/goals/<slug>.md. Stevie runs the goal by pasting the file contents into a fresh session — this skill never executes the run.

The mega-prompt template lives at assets/mega-prompt.md.tmpl. The stub plan handed to /interview lives at assets/stub-plan.md.tmpl.

Trigger

• /make-a-goal — no args, interview asks for the outcome line too
• /make-a-goal <outcome> — outcome line pre-fills {{OUTCOME}}, interview gathers the rest
• Natural language: "make a goal", "set a goal", "build me a goal prompt"

Flow

1. Capture the outcome line

If $ARGUMENTS is non-empty, treat it as the final-outcome line and slugify it. Else, use AskUserQuestion with a single open question: "What does done look like — in one line?"

The outcome line is the most important field. Push back on vague answers ("ship the app") and demand specificity ("deploy v1 of the dashboard at dashboard.example.com with login + 3 charts working"). One sentence, present-tense, observable.

2. Generate slug + paths

• Slug: kebab-case the outcome line, strip stopwords, truncate to 50 chars. Example: "deploy v1 of the dashboard with login" → deploy-v1-dashboard-with-login.
• Plan path: .agents/goals/<slug>.plan.md
• Spec path: .agents/goals/<slug>.plan-spec.md (what /interview writes)
• Goal path: .agents/goals/<slug>.md (the final mega-prompt)

Create .agents/goals/ if missing. If <slug>.md already exists, append a numeric suffix (<slug>-2.md).

3. Write the stub plan

Read assets/stub-plan.md.tmpl, substitute {{OUTCOME}}, and write to the plan path. The stub is deliberately prescriptive — it tells /interview exactly which fields to extract so the resulting spec maps back to the mega-prompt without guesswork.

4. Delegate to /interview

Invoke the /interview skill against <slug>.plan.md. /interview will use its mandatory structured-user-prompt tool (AskUserQuestion on Claude Code) and produce the spec at <slug>.plan-spec.md.

Per /interview's contract, it runs iteratively until "all major areas are covered". Trust it.

5. Map spec → mega-prompt fields, fill gaps

Read <slug>.plan-spec.md. Extract:

Mega-prompt field	Likely spec source
{{PROJECT}}	Executive Summary · Objectives → Primary Goals
{{STACK}}	Technical Requirements → Components / Architecture
{{CURRENT_STATE}}	Implementation Notes · context paragraphs
{{WORKING_DIR}}	Constraints section · or not present
{{CONSTRAINTS}}	Constraints & Dependencies → Technical / Timeline
{{AUDIENCE}}	User Experience intro · or not present
{{SUCCESS_1..3}}	Objectives → Success Metrics

Fields not reliably in /interview's spec template: working dir, audience, current state. If those are missing or thin in the spec, run one AskUserQuestion round with only the unfilled fields. Don't re-ask anything the spec already answers.

If the spec produced more than three success metrics, ask Stevie which three are the hard-gates for "done". If it produced fewer than three, ask for the remainder — three is the minimum the template needs.

6. Render the mega-prompt

Read assets/mega-prompt.md.tmpl. Substitute every {{FIELD}} token with the gathered value. Multi-line values: keep them single-line where the template uses a single line (Context block) — join with ; if needed. The Operating Rules / Quality Bar / Final Deliverable blocks are verbatim — never edit them.

Write to <slug>.md.

7. Report back

Tell Stevie:

• where the goal file landed (absolute path)
• where the spec landed (in case he wants to revise context without re-interviewing)
• one-line instruction to run it: paste the contents of <slug>.md into a fresh Claude Code or Codex session
• the slug (so he can refer to it later)

Caveman mode by default. Address as Stevie.

Non-goals

• Do not execute the goal. This skill only produces the artifact.
• Do not turn prerequisite access/setup into the goal. If Stevie gives secrets, tokens, project IDs, repo state, or provisioning context, treat those as context/constraints for the mega-prompt. The goal line should stay at the outcome/options level; Claude Code works out steps and phases.
• Do not create mini-goals by default. When Stevie says the ultimate goal came from another agent/person, recover that outcome and render one autonomous goal unless he explicitly asks to split it.
• Do not commit .agents/goals/. Leave that decision to Stevie — the directory may or may not be gitignored per repo.
• Do not edit the Operating Rules / Quality Bar / Final Deliverable blocks in the template. They're intentionally non-negotiable across every run.

Edge cases

• No /interview skill available (e.g. a host without it deployed): fall back to running the question set directly via AskUserQuestion using the field list from assets/stub-plan.md.tmpl. Skip the spec artifact in this mode and render straight to the mega-prompt.
• Outcome line is multi-sentence: ask Stevie to compress to one line before slugging. The mega-prompt's first line is the contract.
• Existing goal with same slug: append -2, -3, …
• Working dir is "this repo": capture the absolute path of $PWD at goal-generation time, not the literal string "this repo".