Instead of receiving a wall of text from an agent, prompt it to generate an interactive HTML artifact using a tool like Lavish. This makes plans easier to skim, critique, and annotate, enabling a much richer and faster feedback loop with the agent.
Manually managing git worktrees for parallel agents creates high cognitive load (naming, dependency installs). Kun Chen built 'Treehouse', a simple tool that automates worktree creation from a managed pool, pre-installing dependencies and eliminating manual overhead.
When a task involves extensive exploration or running many parallel experiments, spawn sub-agents. This prevents clogging the main agent's context window with potentially irrelevant information. The sub-agents perform the work in isolation and return only their final conclusions.
Create a project-specific `agents.md` file to provide agents with high-level context, key file structures, and explicit instructions for tasks like end-to-end testing. This ensures agents perform comprehensive, project-appropriate validation beyond generic unit tests.
The agent that writes code is biased and may miss its own errors. Kun Chen uses a tool spawning a new agent with a fresh context. This 'reviewer' agent analyzes the original intent and scrutinizes the code for bugs, catching far more edge cases than self-review.
Treat AI agents like a team you oversee. By running multiple concurrent sessions, even within the same project, you can stay productive while agents complete long-running tasks. This mirrors a manager's context-switching between different high-level workstreams.
Traditional software development processes, like peer code reviews, were built for a cadence of 10-15 PRs per month. When AI agents enable a 10x increase in output, the human team becomes the bottleneck, forcing a shift towards AI-driven review and validation.
Kun Chen's 'no mistakes' pipeline includes a testing phase where agents run comprehensive end-to-end tests to check for regressions. Crucially, the agent captures and embeds evidence, like screenshots or videos of the working feature, directly into the PR description for easy human verification.
Kun Chen spends most of his time in the planning phase, creating detailed specs for AI agents. The coding is almost entirely delegated, and validation is also agent-led. This shift allows him to scale his output by focusing on high-level direction, not implementation details.
Kun Chen doesn't just use off-the-shelf agents; he builds his own tools like Lavish (visual planning) and 'no mistakes' (validation). This highlights a key trait of power users: identifying and solving personal workflow frictions with custom software instead of waiting for public solutions.