Primary navigation

Best practices

Starting out with ZeroTwo, plus tested habits that improve your results

Treat ZeroTwo as a teammate you tune over time: give each task the right context, keep lasting guidance in a project instructions file, pick sensible permissions, connect outside systems, turn repeated work into skills, and schedule it once it is predictable. These habits apply to Work and Code on the desktop app, and mostly to the web app too.

Strong first use: Context and prompts

A good prompt names the goal, the context that matters, any constraints, and what done looks like. In the composer, type @ to mention files, folders, Library items, pages, or connectors so the agent starts from the right material. Match reasoning effort to the job: a lower setting for scoped edits, a higher one for debugging or long runs. Change it from the reasoning dropdown in the composer or with /reasoning. You can also dictate a prompt with the composer's mic button.

Plan first for difficult tasks

When a task is ambiguous or touches many files, turn on Plan mode with the "Plan mode" chip or /plan-mode. For that turn the agent works read-only and returns a plan for you to approve before anything changes. You can also simply ask it to question you about requirements before it starts.

Make guidance reusable with AGENTS.md

A project instructions file is a README the agent reads automatically: repo layout, how to build, test and lint, conventions, things to avoid, and what finished means. ZeroTwo loads one file per project, the first non-empty of ZEROTWO.md, AGENTS.md or CLAUDE.md, from the project root, and also reads a global one from ~/.zerotwo/. On the desktop app, /init has the agent draft an AGENTS.md for you. Edits are picked up mid-session. Keep the file short, add a rule only after the agent repeats a mistake, and use @include lines to pull in longer reference files. For preferences that belong to you rather than a project, use Settings > Personalization > Custom instructions.

Configure ZeroTwo for consistency

There is no config file; defaults live in the app. Pick the model and reasoning effort in the composer, and choose a permission preset: "Ask for approval", "Approve for me", or "Full access". Start with "Ask for approval" and loosen it only in repos you trust. On the desktop app, Settings > Coding holds Git, Worktrees, Hooks, and Code Cloud options. The OS-level sandbox exists only on macOS; on Windows and Linux, local commands rely on your approvals, so be more careful with "Full access" there. Fix environment problems early, since many poor results come from a broken setup rather than a weak prompt.

Improve reliability with testing and review

Ask the agent to write tests, run the checks, confirm the behavior, and review its own changes, and put your standard for each in the prompt or the instructions file. In Code, /diff opens the diff viewer. On the desktop app, /code-review reviews unstaged changes or compares against a branch, and the Code Review page offers "Review with ZeroTwo" for pull requests, with per-repository review instructions you can set. These reviews are started by you; ZeroTwo does not review every pull request automatically.

Use MCPs for external context

Reach for MCP when context lives outside the repo, changes often, or needs a repeatable tool. Check Customize > Connectors first, since many services are already available there. For anything else, use Add > Add MCP server and enter a remote URL with no authentication, a bearer key, or OAuth, then use Test connection. Local stdio servers run on the desktop app only. For each tool you can choose "Always allow", "Needs approval", or "Blocked". Begin with one or two tools that remove a real manual loop.

Turn repeatable work into skills

Package a recurring workflow as a skill: a SKILL.md focused on one job, with a description that says what it does and when to use it. Start from one task that already works and ask the agent to use the Skill Creator, or on the desktop app use Add > Record a skill. Your own skills live in ~/.zerotwo/skills/<slug>/, and the desktop app also finds skills in a project's .agents/skills or .claude/skills folder. To share a workflow with a team, bundle it in a plugin. Invoke a skill by typing $ or / in the composer.

Use scheduled tasks for repeated work

When a workflow is predictable, create a scheduled task under Scheduled in the sidebar. Set the prompt, how often it repeats (interval, daily, weekly, monthly, and so on), and whether it runs Local on the desktop app or in the Cloud. Local runs happen only while the app is open, and Cloud runs fire without it. A task can start a new chat each run or work against a GitHub repository. Scheduled runs use full access with no approval prompts, so schedule only workflows you already trust. Skills define the method; schedules define the timing.

Organize long-running chats

Keep one chat per coherent outcome and branch only when the work truly splits. Useful commands are /fork (listed as "Continue in new chat"), /compact to free context, /side for a throwaway side chat, and /worktree on the desktop app. The context ring in the composer shows how full the window is. Hand bounded exploration or triage to subagents, which appear in the Subagents panel.

Common mistakes

Avoid putting lasting rules in prompts, hiding how to build and test the project, skipping planning, granting full permissions too early, running parallel work without worktrees, scheduling workflows that are not yet reliable, babysitting the agent, and using a single chat for a whole project.