This is the exact wording behind [Talking to Your Agent](/CoDev/working-with-your-agent/)'s "What you actually say" section — copy-paste it rather than reconstructing it from memory. Two prompts you'll type often: kicking off the next task, and starting outer-loop review on an open pull request. Both assume CoDev is installed -- there is no agent to start or select first, just the conversation you're already having. The agent loads the `outer-loop-review` skill for the second prompt itself, so it asks you for a specialist selection rather than sending you to a different session. Fill in the bracketed parts. Everything else is meant to be typed as-is — these are deliberately short. The agent grounds itself in the repository and asks you only for decisions that are actually yours to make; a longer prompt does not make it safer, and mostly just repeats what it already does by default. ## Starting the next task ```text Take task from the wave plan at . GitHub user for ownership/assignment is . If the plan still has placeholder owner/reviewer names that haven't been assigned yet, disregard them and use the login above instead. Create the task's own branch, then stop and show me the implementation plan before any code changes — I want to discuss and accept it first, not after the fact. ``` Why this is enough, and no more: - **You don't need to ask for a GitHub issue explicitly.** `codev slice begin` checks for one itself before opening round state, creates it if this repository tracks issues on GitHub and none exists yet, and `codev task start` refuses to proceed silently without one — you'll be told directly if it can't resolve this on its own instead of finding out later that a pull request never linked back to anything. - **Name the real GitHub user up front.** Ownership/reviewer fields in a wave plan are often still placeholders (`Developer 1`, a role name, whatever the plan template shipped with) — naming the actual login once here means it's the only one used, everywhere it matters, without you having to repeat the correction later. - **"Show me the plan before any code changes" is the one instruction worth always including.** It's already the default for delegated, multi-session, or higher-risk work, but stating it removes any doubt for a borderline case and guarantees a checkpoint even for a smaller item. - **If this task's pull request needs a specific integration branch** (not the repository's default — a `plugins` line, a `develop` branch, whatever this repository actually uses), say so before the branch is pushed: "target ``, not ``." Set it once, repository-wide, with `codev config set git.pr_base ` instead of repeating it on every task. ## Starting outer-loop review ```text Start outer-loop review for PR #. Present the five specialists as a numbered menu and wait for my selection before dispatching anything — don't run a fresh five-specialist pass on your own judgment. ``` Why the second sentence is there: this is a known, actively mitigated gap. The `outer-loop-review` skill already instructs the agent to present the menu, but nothing can force a model to render it before invoking a specialist, and doing exactly that — skipping straight to a full five-specialist pass — has been observed in real sessions. That menu question is the actual guarantee that five real model calls only happen after you've answered it, on every platform. Restating the expectation here costs one sentence and closes the gap the rest of the time. If you'd rather run every dimension without narrowing it down, say `all` when asked, or state it in the same message: "...run all five." Either way, CoDev records exactly which specialists actually ran, distinctly from what was merely asked, in `codev task log`.