jackin-goal-prompt

Distill a roadmap item (and any plan files the operator hands you) into a single self-contained /goal prompt — one an executor agent with zero prior context can run to completion. The prompt is the product; its tightness decides whether the goal succeeds.

Capped at 4000 characters, hard. Structure borrowed from /improve's plans: self-contained, verification-gated, hard-boundaried. The skeleton that carries execution — objective, scope, verify gates, done criteria, STOP conditions — stays; everything the executor can read from files (code excerpts) or inherits from jackin❯'s auto-loaded rules (AGENTS.md, COMMITS.md, ENGINEERING.md) is pointed at, never restated.

Distinct from jackin-research: that authors a dossier brief (a multi-page prompt.mdx file run via /goal Follow). This produces a compact inline prompt run via /goal Implement or /goal Run.

When to use

• Operator asks for a goal prompt / wants to kick off /goal on a roadmap item.
• Operator has a roadmap item plus /improve plan files and wants them fused into one runnable prompt.

When NOT to use

• Authoring a research dossier brief → jackin-research.
• Designing the roadmap item itself → jackin-brainstorm.
• Opening the PR → jackin-create-pr / jackin-propose.

Arguments

• <slug> — the roadmap item (slug, path, or the .mdx). Always required.
• --plan <path|glob> — plan markdown to fold in (a folder, a file, or a glob). Usually /improve output.

Process

1. Gather the change. Read the roadmap .mdx. Anchor on title: and **Status**: only — real items drift from the Problem/Why/Design/Tasks scaffold, so tolerate any heading set. If --plan is given, read every plan file and lift its scope, steps, verify commands, done criteria, and STOP conditions.

   Done when you can state, in two sentences, what this change builds and where its code lives.

2. Read the repo's real verification gates. The verify commands come from the repo, never invented. Pull them from CONTRIBUTING.md (the merge-readiness block), .github/workflows/ci.yml, and mise.toml — for jackin❯ that is cargo fmt --check, cargo clippy --all-targets --all-features -- -D warnings, cargo nextest run --all-features, the e2e profile, plus the bun docs gate when docs/ is touched.

   Done when every gate you cite is a command the repo actually runs, with its expected result.

3. Compose the prompt — the compressed /improve skeleton:

   • Objective — one or two lines: what to build, pulled from the roadmap/plan.
   • Scope — in-scope paths; out-of-scope boundaries (what looks related but must not be touched).
   • Verify gates — the commands from step 2, each with its expected result.
   • Done criteria — machine-checkable, one line each (a command + result, or a greppable absence), never prose like "works correctly."
   • STOP conditions — drift vs the plan, a scope boundary that won't hold, a gate that won't go green honestly.
   • Conventions — one line pointing at jackin❯'s auto-loaded rule files; do not copy their content.

4. Enforce the 4000-character cap. Write the prompt to a temp file and wc -c it. Over the cap, compress in this priority order until it fits: drop code excerpts (name the file + the contract, not the code), collapse conventions to a pointer, tighten the objective, cut the "why" narrative, merge done criteria into the verify gates. Re-count. The cap is hard — never emit over 4000.

   Done when wc -c reports ≤ 4000 and the prompt still carries objective + scope + verify gates + done criteria + STOP conditions.

5. Emit + report. Print the prompt in a single code block, ready to paste after /goal Implement <slug> or /goal Run. State the character count. Name what you compressed or dropped to fit, so the operator can ask for it back inline if they'd rather raise the cap.

Common mistakes

• Exceeding 4000 characters — the one hard constraint.
• Inventing verify commands instead of reading the repo's real ones — a gate the repo doesn't run is noise.
• Restating jackin❯'s auto-loaded rules (AGENTS.md, COMMITS.md, ENGINEERING.md) — the executor auto-loads them; point, don't copy.
• Inlining code excerpts — the executor reads the files; name the file and the contract it must satisfy.
• Producing a prompt that leans on chat context ("as discussed", "the approach we chose") — not self-contained, breaks on a fresh executor.
• Ignoring --plan files when given — they carry the scope and done criteria; folding them in is the point.
• Dropping the STOP conditions — they are the executor's only honest escape hatch.

Tooling

No xtask owns this. The skill orchestrates reads (the roadmap .mdx, the plan files, the repo's verify commands) and emits a single ≤ 4000-char string. The structure borrows from /improve's plan anatomy (self-contained, verification-gated, hard-boundaried) and the /goal Run brief shape jackin❯ already uses for gap-closure.