Ts Floyd@ts_floyd

Segui

20 GrokBot Tips from poteto Founder Session

If you missed Lauren Tan's founder session on Grok Bot Galaxy, here is the breakdown of twenty specific tips for using Grok bots effectively. The core philosophy starts with a mission rather than defined roles; a prompt like compare 3 competitors' pricing by friday works better than make a research bot because it gives direction. The first hire should be a chief of staff bot to distribute tasks and maintain context across the team. A starter team consists of a coordinator, an engineering owner, and an independent reviewer, adding more only once those three are functioning. On day one, teach the bots your top three priorities while strictly restricting them from sending, paying, or publishing anything. Day two involves handing off one repetitive task, which becomes a skill by day three, with specialists added in week one. Demonstrating a workflow once allows it to become a skill without requiring perfect prompts. Trust in bots should be established in stages, starting with drafting and approval before granting full autonomy. External emails and calendar changes require explicit human approval, waiting for a simple yes. Completion requires verification through screenshots, videos, or data checks rather than just a status of done. Verification must confirm actual behavior, such as the absence of duplicates, rather than relying solely on passed tests. Bots cannot approve their own work; a separate reviewer bot must check the final version. Overnight jobs require a finish condition, an isolated worktree, a decision log, and manual merging instead of auto-merging. Briefs should follow a ticket format including problem, repro steps, owner, and acceptance criteria. Bots should utilize app traces and simulators instead of guessing what to do. Providing bots with a feature map containing tabs, selectors, and shortcuts reduces token usage significantly. Evaluations function as unit tests where a coordinator writes rubrics and sub-agents run blind tests. Coding preferences like banning useEffect and code comments should be enforced in CI rather than prompts. Methods should be packaged as skills, illustrated by a plugin setup command like /add-plugin pstack followed by /setup-pstack. Verify locally before scaling in the cloud because fifty agents without checks equals fifty expensive wrong answers. Count accepted work, not PRs, tracking fixes accepted, review time, rework, escaped bugs, and cost per change.

4h

Ts Floyd@ts_floyd

Segui

20 GrokBot Tips from poteto Founder Session

If you missed Lauren Tan's founder session on Grok Bot Galaxy, here is the breakdown of twenty specific tips for using Grok bots effectively. The core philosophy starts with a mission rather than defined roles; a prompt like compare 3 competitors' pricing by friday works better than make a research bot because it gives direction. The first hire should be a chief of staff bot to distribute tasks and maintain context across the team. A starter team consists of a coordinator, an engineering owner, and an independent reviewer, adding more only once those three are functioning. On day one, teach the bots your top three priorities while strictly restricting them from sending, paying, or publishing anything. Day two involves handing off one repetitive task, which becomes a skill by day three, with specialists added in week one. Demonstrating a workflow once allows it to become a skill without requiring perfect prompts. Trust in bots should be established in stages, starting with drafting and approval before granting full autonomy. External emails and calendar changes require explicit human approval, waiting for a simple yes. Completion requires verification through screenshots, videos, or data checks rather than just a status of done. Verification must confirm actual behavior, such as the absence of duplicates, rather than relying solely on passed tests. Bots cannot approve their own work; a separate reviewer bot must check the final version. Overnight jobs require a finish condition, an isolated worktree, a decision log, and manual merging instead of auto-merging. Briefs should follow a ticket format including problem, repro steps, owner, and acceptance criteria. Bots should utilize app traces and simulators instead of guessing what to do. Providing bots with a feature map containing tabs, selectors, and shortcuts reduces token usage significantly. Evaluations function as unit tests where a coordinator writes rubrics and sub-agents run blind tests. Coding preferences like banning useEffect and code comments should be enforced in CI rather than prompts. Methods should be packaged as skills, illustrated by a plugin setup command like /add-plugin pstack followed by /setup-pstack. Verify locally before scaling in the cloud because fifty agents without checks equals fifty expensive wrong answers. Count accepted work, not PRs, tracking fixes accepted, review time, rework, escaped bugs, and cost per change.

4h