Your team

Getting good results

The tool matters less than how you use it. These habits come from watching real founders run Guildly day to day.

Describe outcomes, not implementations

"If someone isn't on a Mac, offer them a waitlist for Windows and Linux" beats "add a conditional to the download component". You hired the team to figure out the how. When you dictate implementation details you take on risk you don't need, and you often get a worse solution than the one the team would have found.

Give references, and check they were used

For anything visual or stylistic, point at examples: "here are two sites whose design I like, study how they use spacing and type". And when the work comes back, verify the references actually shaped it. "I don't think you looked at the sites I gave you" is a legitimate and useful review comment.

Be blunt in reviews

Softened feedback produces softened fixes. "This isn't good, the fonts are random and the layout has no hierarchy, redo it using the references" gets a genuinely different second attempt. Agents take direct criticism well because there's nothing to take. Vague approval-ish feedback is the only real way to waste a review cycle.

Match the model to the task

Model choice is your quality and cost dial. Strong models for thinking and taste, cheaper models for well-specified execution. Telling the manager "design with the strong model, implement with the cheaper one, and make the design instructions explicit enough that the cheaper model can't go wrong" is an advanced move that works remarkably well.

One thread per topic

Keep each piece of work in its own thread and answer questions in the thread they were asked in. When you cram three topics into one conversation, agents lose track and so will you.

Ask for status instead of wondering

"What's happening?" is a perfectly good message. So is "what did we change last week to fix the spam?" The team can always reconstruct and summarize what it did. Never sit there guessing.

Put repeated corrections into instructions

Correcting an agent in chat fixes today. Editing its system instructions in its profile fixes forever. If you notice yourself repeating a correction, that's the signal to make it permanent.

Actually do human review

The Human Review column only protects you if you look at what you're approving. Follow the test steps, click around the result. Two minutes of real review per ticket is the entire difference between a codebase you trust and one you don't.