goal-writer

A portable agent skill that drafts goal+rider document pairs — the two-file briefing system for autonomous coding agents — fully aligned with the official /goal best practices of Claude Code and OpenAI Codex CLI.

Invoke it as /goal-writer in Claude Code, or $goal-writer in Codex.

Documentation site: https://nicholasspisak.github.io/goal-writer/ — a plain-language, accessible walkthrough of the two-file system and the four loop types.

Learn to run agents like this: join the AI Operator Academy community — designing loops, briefing agents, and shipping real work.

The two-file system

One round of agent work gets two committed markdown files in your project's docs/goals/:

The 4,000-character cap is not arbitrary: it is Codex CLI's MAX_THREAD_GOAL_OBJECTIVE_CHARS, matched by Claude Code's /goal condition limit. The goal file body pastes directly into /goal in either harness; the rider stays on disk, referenced by absolute path. The same pair briefs Claude Code, Codex, Cursor, or a human reviewer without modification.

The pattern is Greg Ceccarelli's goal engineering: "The cap forces decisions. The rider forces precision. The phases force sequencing. The depth tests force testability."

One pair, four loop types

A loop is an agent repeating cycles of work until a stop condition is met, and /goal is one of four types (turn-based, goal-based, time-based, proactive). The pair is primitive-agnostic: /goal is its primary runtime, but the same two files feed a turn-based verification skill, a scheduled /loop, or a proactive routine without a rewrite. /loop, /schedule, and dynamic workflows are Claude Code primitives; Codex does the same recurring work by wrapping headless codex exec in a scheduler. The skill also recommends encoding "what good looks like" as a companion verify skill so the agent self-verifies end-to-end. See references/harness-goal-commands.md.

What the skill does

When you ask for a goal (/goal-writer add rate limiting to the API), it:

1. Reads your project first — architecture doc, prior goal+rider pairs, recent commits, source at HEAD. It asks rather than invents.
2. Scopes one round — one feature, 2–5 files, half a day to a day of agent runtime, named by a single headline word.
3. Drafts the goal under the 4,000-char cap, with verification written to satisfy both harnesses' evaluators: Claude Code's tool-less transcript evaluator and Codex's evidence-based completion audit.
4. Drafts the rider — phased plan with named depth tests written before implementation, posture invariants, error footers, out-of-scope list.
5. Validates mechanically with the bundled validate_pair.sh (size cap, phase count, required sections, mutual citation, shared timestamp) and commits the pair.

Install

Preferred — the Vercel skills CLI (npx skills):

npx skills add NicholasSpisak/goal-writer

That's the whole onboarding. The CLI detects which agents you have (Claude Code, Codex, and others), then asks where to put the skill — this project (into each agent's skills dir, e.g. .claude/skills/) or globally for your local agent — and whether to symlink (auto-updates) or copy. It reads SKILL.md straight from this repo; nothing to clone.

Skip the prompts for a global, both-harness, non-interactive install:

npx skills add NicholasSpisak/goal-writer -g -a claude-code -a codex -y

Useful flags: -g global (user-level) instead of project · -a <agent> target a specific agent · --copy copy instead of symlink · -l list skills without installing · -y skip confirmation prompts. Full syntax: npx skills add --help.

git clone https://github.com/NicholasSpisak/goal-writer.git
cd goal-writer
./install.sh              # both harnesses, user-level

./install.sh --claude         # Claude Code only  → ~/.claude/skills/goal-writer
./install.sh --codex          # Codex only        → ~/.agents/skills/goal-writer
./install.sh --project        # current repo      → .claude/skills + .agents/skills
./install.sh --codex-prompt   # optional /prompts:goal-writer command for Codex

Use

Claude Code

/goal-writer harden the webhook handler against replay attacks

Codex CLI

$goal-writer harden the webhook handler against replay attacks

(or /skills to browse; or /prompts:goal-writer <topic> if you installed the prompt shim). Codex also auto-selects the skill when a request matches its description.

Then start the round in either harness by pasting the drafted goal file's body:

/goal <goal file body>

Claude Code keeps working until its evaluator confirms the condition from the transcript. Codex re-injects the objective every turn and runs a completion audit before accepting "done." (Codex: enable once with codex features enable goals if /goal isn't listed.)

Repository layout

skills/goal-writer/
├── SKILL.md                              # the skill (agentskills.io standard)
├── references/
│   ├── goal-template.md                  # goal skeleton + worked example
│   ├── rider-template.md                 # rider skeleton + depth-test rules
│   └── harness-goal-commands.md          # /goal mechanics in both harnesses
└── scripts/
    └── validate_pair.sh                  # 5 mechanical pair checks
codex/goal-writer.md                      # optional Codex custom-prompt shim
install.sh

When not to use it

A five-line CSS tweak doesn't need eleven phases. Skip the pair for hot-fixes and 15-minute polish; the overhead amortizes only against rounds that would otherwise run hours of agent time. The pattern also wants a real architecture doc, a CHANGELOG, and a fast (~30 s) test lane — the skill offers to bootstrap the missing pieces first.

Community

Goal Writer is one piece of a bigger practice — designing loops, briefing agents, and shipping real work. Come build with the community: Join the AI Operator Academy.

Credits

• Pattern & original skill: the goal-engineering method and the goal-rider-author skill are by Greg Ceccarelli / SpecStory, Inc., published under Apache-2.0. This skill is a derivative work — restructured for progressive disclosure and extended with the official Claude Code and Codex /goal documentation (evaluator constraints, completion-audit alignment, dual-harness install).
• This skill: Nick Spisak.

License

Apache-2.0. See NOTICE for attribution.