Stop (topic)

A "Stop" in AI development is a stopping condition (a rule that tells the AI when to pause and get human help). It matters because AI doesn't naturally know when to stop — it will keep running continuously until it thinks it's done, which can waste time and money. You need to put guardrails (boundaries that prevent the AI from doing the wrong thing) around long tasks.

There are four key stopping conditions to set. First, cap the number of attempts — limit the AI to three, ten, or twenty tries depending on the task, so it doesn't burn resources. Second, tell the AI to stop if it's missing context or critical access to another system, so it can ask you for that file or connection. Third, use stopping conditions for high-stakes tasks where the work involves legal, financial, or reputational risk — the AI completes the task but flags it for your review before anything gets sent to a client. Fourth, use success criteria (binary checks that define what "done" looks like) so the AI can verify it met the goal and stop on its own.

When you build these stopping conditions into your prompt, the AI knows to seek your approval or assistance at the right moments. This prevents it from making bad assumptions, going down wrong paths, and burning money on endless attempts.

To use the Stop feature effectively, you must understand two roles: stop sequences in an API call and stop hooks in a coding agent like Claude Code.

A stop sequence is a specific phrase you set in your API request that tells the model to end generation when it reaches that phrase. For example, if you send a prompt and include "---END---" as a stop sequence, the model will stop writing as soon as it outputs that string. The API then returns a response with a "stop reason" field (the signal telling you why generation ended). This field is crucial: it tells you whether the model stopped because it hit your custom phrase, reached its max token limit (the maximum number of words it's allowed to generate), or produced a natural ending. Checking the stop reason helps you debug prompts, understand cost, and decide if a short answer ended naturally or was cut off by a limit.

A stop hook (a script that runs when generation finishes) is used in Claude Code to prevent infinite loops and burning cash (wasting money on repeated calls). The key rule: keep stop hooks passive. If your stop hook runs a command that causes Claude to start generating again, you create an infinite loop. The official Anthropic guidance says to use a "stop hook active field" to break out of such loops. A safe use of a stop hook is to reflect on what happened during the session and propose updates to your rules (like Claude.md) while the context is fresh. This lets your codebase rules evolve automatically. For example, with a Gemini model, you can use a stop hook to check if the output is complete before letting the AI continue, ensuring it doesn't keep trying continuously and burning money and time. Always inspect the stop reason in your application to maintain control over the conversation flow.

To avoid burning cash when using Gemini or other AI tools, you must set clear stopping conditions (rules that tell the AI when to halt) before any session. Without them, the AI runs continuously, wasting money and time. Always add a safety cap—for example, tell the AI to stop after 20 or 30 turns (conversation steps). This prevents runaway costs.

The most effective approach uses a three-part structure: outcome, check, and stop. First, define the specific outcome you want. Second, build in a check so the AI evaluates whether the task meets your requirements. Third, set a hard stop condition. This skips most common mistakes before they happen. For high-stakes tasks—like sending money, touching your live product, or emailing a client—never let the AI run unattended. Keep those jobs for yourself.

Another critical practice is using stop hooks (scripts that run after each turn). After every AI action, a stop hook can reflect on what happened and propose updates to your rules while the context is fresh. This evolves your guidelines automatically, so you do not maintain them by hand. Start hooks (scripts that fire when a session begins) can load context upfront.

Also, instruct the AI to flag high-risk tasks it identifies—anything with legal, financial, or reputational stakes. The AI completes the task but does not send anything until you review it. This prevents damaging outputs.

Finally, know your burn rate (the gap between cash going out and revenue coming in). Track this monthly. If you do not set stopping conditions, you will burn through your budget fast. Stop conditions let the AI work unattended safely, but only for low-risk, repeatable tasks that do not require real judgment. Use this pattern to stretch your resources.