Raytar (@Raytar) on X
A Twitter/X thread explaining Claude Code's /goal and /loop commands, which let Claude repeatedly prompt and check itself against a defined finish condition rather than requiring a human to babysit each turn. It contrasts single-prompt workflows (which can produce confidently wrong output, e.g. fabricated sources) with loop-based workflows that self-verify, and gives a copy-pasteable "charter" template covering goal, work discovery, self-checking, state persistence (LOOP-STATE.md), and stop conditions.
What it claims
The thread argues for "loop engineering" — writing loops that prompt Claude repeatedly and self-check until a job is done — over typing prompts one at a time. Its framing anecdote: Boris Cherny (credited as the person who built Claude Code) reportedly said in June 2026 "I don't prompt Claude anymore." The two Claude Code commands it names are /goal and /loop, and the summary already covers the charter template and the fabricated-sources example. What the summary does not carry is the distinction between the two commands and the explicit guidance on when a loop is the wrong tool.
How it works
The key distinction the summary flattens:
/goal— the loop with a finish line. You describe what "done" looks like; after every turn a second copy of Claude checks whether the goal is met and, if not, tells the first copy why. Stops itself when the condition is true./loop— the loop with no finish line, running on a rhythm. Example given verbatim:
/loop 30m check whether my live site is back up by loading the homepage.
The moment it returns a normal page, tell me and stop checking.The 30m means every 30 minutes; the page claims you can also phrase a schedule in plain language ("every morning, triage my inbox").
The page distinguishes a loop from a cron job: a cron job runs a fixed script; a loop puts Claude inside as the decision-maker that picks the next action, retries, undoes, or stops mid-job.
Limits
The page's own "when not to build a loop" section is the actionable part the summary omits:
- One-off tasks — a plain prompt is faster; loops only earn their setup cost on repeating or many-piece work.
- Cost — a self-checking, retrying loop runs Claude several times per item, so you hit usage limits faster on a Claude plan.
- Vague work — "think of a better product strategy" is not a loop; define the measurable goal first.
The charter also bounds retries: three tries per item, then log as blocked and move on.
Note the provenance caveat: the Boris Cherny quote and the "built Claude Code" attribution are the thread author's claims, dated June 2026, unverified here.
Organisations named here
Technologies named here
Techniques named here
- /goal
- /loop
- Loop engineering
- LOOP-STATE.md state file
Also named here
- Boris Cherny
Related in coding-agents
