Mastering the /goal Command in Codex: Practical Tips for Effective AI Agents

Hello, I’m Lu Gong. I recently read Dominik Kundel’s short guide on using the /goal command in Codex and found it valuable enough to share.

The /goal command puts Codex into a goal‑oriented mode: you give it a target, and it iteratively decomposes, executes, and judges completion, running for hours or even days. The feature is now available in Codex’s app, IDE plugins, and CLI.

Reading Kundel’s guide reminded me of OKR (Objectives and Key Results) in workplace management: setting a clear, quantifiable, and slightly challenging goal for an agent mirrors setting an OKR for a team.

1. Goals must be verifiable, preferably quantified

The prompt you write when launching a goal serves both as a task description and as the exit criteria. Codex compares its progress against this text after each iteration, so the prompt should be concise yet explicit about what counts as “done.” Vague goals like “optimize performance” lead to arbitrary judgments, whereas numeric targets such as “reduce build‑deployment time by 30%” or “raise Dice score above 0.9” give the agent a clear benchmark.

You can also start with a discussion to clarify requirements before invoking /goal, and you can modify the goal later as needed.

2. Provide ample context

Supplying only a high‑level target (e.g., “reduce build time 30%”) may cause the agent to wander or produce irrelevant ideas. If you know the bottleneck, mention it, along with usable tools and potential pitfalls. Kundel’s example tells Codex it may log into Google Colab via Chrome and generate its own dataset. When the context is unclear, you can first run a “plan” mode to produce a plan.md and then feed that file to /goal.

3. Make progress measurable

For distant or multi‑path goals, the agent needs a way to gauge remaining distance. Some tasks naturally expose metrics (build time, test coverage). Others require you to create measurement utilities, such as a screenshot‑diff tool or an evaluation suite. Kundel noted that when replicating a component from video, Codex built a screenshot‑diff tool and later added several comparison modes.

Guard against “cheating” – e.g., achieving 100% test pass by deleting tests or meeting pixel‑level similarity by copying the design image. Include safeguards in the acceptance criteria.

4. Run in a production‑like environment

The agent must execute in an environment that mirrors production: same stack, flags, and databases. Kundel reported a failure when a preview environment omitted certain build paths, forcing a manual deployment to a near‑production configuration. For algorithmic work, provide a real deep‑learning setup with CUDA, GPUs, and PyTorch; otherwise the agent cannot run or evaluate correctly.

5. Be cautious with visual‑type goals

Goals such as “match this image pixel‑perfectly” are tempting but prone to drift. Without tight constraints, the agent may obsess over a single icon and lose sight of the overall objective, and token limits make repeated image reading inefficient. A safer approach is to treat the image as a reference and verify against concrete specs, feature lists, or design guidelines.

6. Track long‑running tasks

When the agent runs for hours or on another machine, users can lose visibility. Useful tactics include having the agent commit at key milestones, push draft PRs with preview deployments, write progress to a markdown or HTML file, post updates to Slack, or open a side‑branch conversation (/side) to query the current state without interrupting the main flow.

7. Review and clean up after goal completion

Reaching the goal does not mean immediate hand‑off. Especially for optimization work, let the agent review the entire process, run a local code review (/review), and reflect on alternative solutions tried. Since /goal does not automatically backtrack, the workspace may contain failed or half‑finished attempts; prune these artifacts before delivering the final output.

Conclusion

The /goal command embodies the expectation that an agent can autonomously achieve a predefined objective. In practice, combining /goal with a “yolo” mode lets both Claude Code and Codex complete a wide range of tasks in a single run.