Grok Bot 省 token 的分工与沙箱做法

Solomon Rojie @Solomonrojie原推发布于 2026.10.25

在 X 上打开原推

Official SpaceXAI tip: stop burning Grok Bot tokens stop making your bot reason through the same commands, inspect the same files, and explain the same process every time a routine runs give Grok Bot the decisions. give scripts the repetition. enforce permissions outside the prompt here’s the split: → the lead Bot defines the task, approves the approach, and handles exceptions → scripts collect data, transform files, and run repeatable checks → a separate verifier checks the deliverable against the brief and test results three checkpoints need judgment: → before the plan becomes execution → when a check fails and the cause is unclear → before the result is accepted or an external action is approved everything between those checkpoints should produce files, exit codes, and a short report. your bot shouldn’t need to narrate every successful command the security part matters just as much “only touch this folder” is an instruction. a restricted executor can enforce which files a process can read or write and which domains it can reach Anthropic’s open-source sandbox-runtime can wrap commands and self-hosted MCP servers with those restrictions. it’s a separate tool you configure on a compatible executor, not a switch that sandboxes every Grok Bot action your Bots also share one cloud computer. creating another Bot doesn’t create a separate vault for credentials or files Grok Bot plans. scripts execute. the verifier checks. you approve consequential actions the full workflow: → keep the goal, acceptance criteria, and approved scope in BRIEF.md → save repeatable commands in scripts instead of rebuilding them in chat → write detailed output to files and return only a summary with links → pass the verifier the brief, the deliverable, and the check results → repair the failed step without rerunning successful work → require approval before sending, publishing, deleting, or changing production → save the proven process as a skill before putting it on a schedule start with one workflow and compare accepted results, elapsed time, and usage. another Bot earns its place only if its contribution justifies the extra work paste this into Grok Bot: “Refactor this recurring workflow so scripts handle repeatable operations and Bots handle decisions: 1. Inspect the current process. Identify deterministic steps, judgment calls, required access, and the finished result. Show me the proposed split before changing anything. 2. Create BRIEF.md with the goal, inputs, acceptance checks, and approval boundaries. Turn stable operations into reusable scripts. Save detailed logs to files and return concise summaries. 3. Where I control the executor, propose filesystem and network restrictions for commands and self-hosted MCP servers. Test an allowed action and a deliberately blocked action. Report anything the environment cannot enforce. 4. Use a separate verifier when an independent check is useful. Give it the brief and evidence. Preserve successful steps, stop after two failed repair attempts, and report what remains unresolved. Show me the changes and one completed test run before enabling a routine.” save this. your bot should spend its allowance solving the exceptions, not rehearsing the routine the article below explains 7 more changes for reducing repeated context, unnecessary checks, and expensive retries ↓

这条是引用着另一条写的：x.com/greshamtrade/status/2112214741103825314