keeps the agent working until a completion condition is met. Both Claude Code and Codex take it as one line, max 4,000 characters. In Claude Code a small model re-judges the condition after each turn from the transcript alone — it cannot run commands.

Draft a condition that can terminate, then format it. A goal fits work bigger than one turn with a checkable finish line; chain small goals with review between them rather than writing one giant goal.

Include, joined with AND — never "or", the loop takes the cheaper branch:

Keep it small. Every constraint narrows the state space the model can explore. Collapse to one terminating criterion when possible, move scope and definitions into a referenced file, and drop non-goals — a constraint earns its place only by closing a real easy-out.

For long goals, also name the final evidence (diff, report, artifact) and require a progress log file — durable state across compaction and resume. If the brief exceeds 4,000 characters, put the details in a and reference that file from the objective.

Never invent missing elements. Ground every element in the user's request, the conversation, or the repository — look things up rather than guessing. If an element cannot be filled from available information, still optimize and format what the user provided, leave the element out, and flag it as missing (see Format). A goal with an invented success condition terminates on the wrong contract.

Before formatting, reread the drafted condition as a lazy model would: what is the cheapest way to make every check pass without doing the intended work? Close the cheapest ones — prefer pairing checks you already have over adding constraints, and do not enumerate every conceivable out into a non-goal list. The recurring outs:

Same discipline as above: an out you cannot close from available information goes in the list as a warning — an invented or absurd constraint is worse than a flagged gap.

Collapse audit goals to one terminating criterion, such as identifying, triggering, and validating one high-severity vulnerability valid under a referenced threat-model file. That file, not the goal, carries scope, attacker powers, severity baseline, and known findings to skip. Use neutral wording ("trigger and validate", not "prove this is exploitable"), require demonstrated preconditions — assumed attacker access is the most common false positive — and stop for human review after each finding rather than piling up untriaged reports. Validate findings with a second pass by a fresh agent, never the finder alone.

Run on the draft (file or stdin). It collapses whitespace to one line, strips prefixes, quotes, and fences, warns on a missing stop clause, and rejects output over 4,000 characters — shorten or move detail to a file and rerun.

Return exactly one fenced block, one line:

/goal <single normalized objective>

Add no prose around it — except when checklist elements could not be grounded: then follow the block with a list, one line per gap, telling the user what to supply.

Draft:

/goal Migrate the auth module:
  - replace legacyAuth() with auth.verify()
  - make sure the tests still work

Redrafted and formatted:

/goal All legacyAuth() call sites use auth.verify(): `rg "legacyAuth\(" -t ts` prints nothing AND `npm test` exits 0 (run both, show the output), without modifying vendor/ or weakening any test. If blocked, stop and report attempted paths and the blocker, or stop after 20 turns.

Here came from the repo's package.json — not a guess — and pairing it with the zero-matches check closes the cheapest out: deleting the call sites instead of migrating them. When nothing grounds an element, format what exists and flag the gaps:

Draft: , with no metric or benchmark anywhere in context:

/goal Make checkout faster

Missing: