Grok Bot 官方文档的省 token 工作流：文件记、代码筛、代理收尾

Solomon Rojie @Solomonrojie原推发布于 2026.10.21

在 X 上打开原推

THIS IS A PURE F*CKING TREASURE SpaceXAI just documented one of the best setups for Grok Bot the idea: keep the working chat short, save useful research in files, and let code handle routine checks before the bot starts thinking 100,000 → 12,500 tokens per task. the same budget goes from 7 usable results to 70 because less of it disappears into repeated work you start Grok Bot like this: “Set up this workflow to reuse completed work, read only relevant sources, and check every result before calling it done.” here’s how it splits: → Grok Bot defines the task, makes the decisions, and produces the finished result → saved files hold confirmed facts, previous findings, and the next action → scripts search, filter, compare versions, and check whether anything changed → verification catches failed requirements and returns specific corrections before you see the result fresh research only starts in three moments: → no saved result answers the question → a relevant source or input changed → an important fact has expired or can’t be verified files remember. code filters. Grok Bot finishes the same principle applies to a whole team: one owner per task, short handoffs, and no second bot rebuilding research the first one already completed the full tree: → save the goal, decisions, and next action in HANDOFF.md → start the next task in a fresh chat using that file → search sources before opening entire documents → check for changes before running a full investigation → reuse valid results and refresh only affected findings → stop failed operations after 2 attempts and preserve completed work → check the output against the brief, then fix only what failed paste the tree and this prompt into Grok Bot ↓ “Reconfigure my workflow around this setup: 1/ Review recent runs and identify the biggest source of repeated work. 2/ Create BRIEF.md, HANDOFF.md, sources/, results/, and a run log. Keep the working context short. 3/ Search saved results first. Retrieve relevant passages, check freshness, and update only what changed. 4/ Give each task one owner. Save progress and stop the same failing operation after 2 attempts. 5/ Check every result against the brief. Allow one focused correction pass, then flag anything unresolved. Show me the proposed changes and test one real task before scheduling it.” save it before your next upgrade the full course below walks through all 7 fixes, with tools, ready-to-copy files, and the complete workflow ↓

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