超级目标

Install

npx ultragoal

An interactive installer walks you through it: choose Claude Code, Codex, or both. Claude Code remains the default for non-interactive installs (--yes) and can install to this project by default (it lands in .claude/settings.json, so teammates get it through git) or machine-wide with --global. If you pick Claude Code, the installer can also pre-configure the repo: the seven working-style questions, .ultragoal/, and the managed CLAUDE.md block. --codex installs the Codex hook-backed goal loop plugin; --all installs both; uninstall removes the selected plugins and marketplace entries. Prefer the Claude route directly? Inside Claude Code:

/plugin marketplace add shamilkayal/ultragoal
/plugin install ultragoal@ultragoal

Want it available in every project on your machine instead of just this one?

npx ultragoal --global

FAQ

Fable 5 already handles long-horizon work — why add a harness? Not to make the model capable; the loop primitives are native and the model is built for them. Two things survive that fact. First, the published workflow still assumes expertise: in Anthropic's own experiments the engineer hand-writes the nine-criteria rubric, knows to spawn a fresh-context verifier, and runs a memory discipline — ultragoal does those for you. Second, one problem is structural, not a capability gap: a worker grading its own work fails in every model generation, which is why Anthropic's guidance reaches for an independent verifier with Fable 5 specifically. The gate makes that separation mechanical instead of habitual.

Do I need to know how to prompt? No — that's the point. You bring what only you know (what you want, who it's for, what must not break); ultragoal writes the expert-grade brief for itself. You review a plan in plain English, never author a prompt.

Versus ralph-loop? Ralph re-feeds the same prompt until a promise appears. Ultragoal adds the parts the article argues matter: a rubric with per-item check commands, an independent verifier, persistent cross-session goals, and enforced distillation into memory.

Versus Managed Agents "Outcomes"? Anthropic's API-side Managed Agents now offer Outcomes — a hosted define-rubric → iterate → independent-grader loop. Same thesis, different home: Outcomes runs server-side per agent request; ultragoal runs in your repo, on any Claude Code (or Codex) session, writes everything as diffable markdown you own, spans sessions and days, and compounds through git-shared memory — plus it authors the rubric for you from a ramble instead of asking you to supply one. If your workload lives in the Managed Agents API, use Outcomes; if it lives in your editor and CI, this is that loop, portable.

Does the verifier have its own context, or does it grade in the same conversation? Its own. The verifier is a separate subagent with a fresh context window and no access to the worker's reasoning — it only sees the goal file and what it learns by re-running the checks itself. (It does share the Claude Code process and permissions; for absolute isolation on high-stakes work, run /ultragoal:verify from a separate headless session as documented in that skill.)

How do I change my setup answers later? They're just markdown: edit .ultragoal/config.md directly (flip verification to off, change scope, anything), or re-run /ultragoal:setup to be re-asked interactively. Changes apply to the next goal you arm.

Does it spend a lot of tokens? The gate itself is free (no model call). The loop spends what the work needs — that's the point of goal-directed runs. Budgets cap the blast radius — pick the depth tier when arming (a quick pass for your first goal) to calibrate. And for execution-heavy goals there's an economy dial, offered beside depth and rigor when it can actually change the bill: the session model stays orchestrator and advisor — interview, spec, decisions, diff review, verification — while cheaper Sonnet executor subagents do the implementation, escalating up when stuck. It's Anthropic's published plan-big-execute-small pattern (their numbers: a Fable 5 orchestrator with Sonnet 5 workers keeps 96% of Fable's quality at 46% of the price; the inverse advisor shape, ~92% at ~63%), and ultragoal is unusually well-placed to run it: cheap executors are only safe when the checking around them is strong, and the rubric + independent verifier are exactly that checking. The verifier is never downgraded.

Can I run it unattended? Yes — that's the recommended mode: npx ultragoal run "<brief>" launches at full autonomy, and --headless runs the loop to completion with no UI at all. The discipline lives in the rubric, the verifier, and the budget — not in you approving each tool call.

Uninstall? npx ultragoal uninstall removes the Claude and/or Codex plugin plus marketplace entries it can find; your .ultragoal/ state stays — it's yours. Add --codex or --claude to target one surface, and --purge to delete a repo's state too.