When assigning Gemini 4 Carbon a complex task, align on the goal, boundaries, and measurable completion criteria first, then confirm that the necessary tools and materials are available. During execution, provide real feedback signals so the model can verify and iterate against a plan. Avoid replacing a clear task definition with instructions added one turn at a time.
Suitable tasks: Engineering work spanning multiple files or steps that can be verified through tests or runtime results, such as auditing accessibility across several pages and fixing issues in batches.
Unsuitable tasks: Routine changes with a clear goal that can be completed quickly. These usually do not require a long autonomous run.
Applicable model version: Gemini 4 Carbon. The guide recommends Gemini 4 Argon for most day-to-day engineering and knowledge work, reserving Carbon for harder tasks that require deeper exploration and repeated verification.
Applicable clients, agents, or APIs: Google AI Studio, Gemini API, and Google Antigravity. This article covers task-prompting methods and does not discuss API parameter configuration.
Recommended reasoning level and parameters: None specified for the prompt template. The source only shows API configuration for controlling the thinking level by task in an example. See the model specifications document for details.
Google's guide gives an accessibility-fix example: first have the model write a script to crawl project routes and record the existing violation baseline, then group fixes by shared root cause, and rerun the script after each change until every route passes with no regressions. The framework below is an editorial adaptation of that example, not a verbatim prompt published by Google:
Goal: <the outcome to achieve>
Scope: <the project areas, files, or routes to inspect or change>
Constraints: <behaviors, compatibility requirements, or design requirements that must not change>
Completion criteria: <observable, decidable pass conditions>
Verification: Record the current baseline first, then verify with <tests, scripts, logs, or screenshots>.
Before starting, restate your understanding of the goal and boundaries, raise any clarifying questions that could affect the implementation, and confirm that the required tools, materials, and environment are available. Then provide a brief plan and follow it. Run verification and check for regressions after each change. If verification fails, describe the observed result, likely causes, and the next check, then continue iterating. Pause and ask for confirmation before taking an irreversible action. At the end, report the scope of changes, verification commands and results, and any unresolved items.For an accessibility fix, for example, completion criteria could be “zero instances of the target violations across project routes, with no new failures in existing tests.” The verification signals would be the route-by-route scanning script and the project's test results. Choose pass conditions for the actual project rather than applying this example to every task.
Align on the task: State the goal, scope, and constraints, and let Carbon raise questions about edge cases or success criteria before a long run.
Confirm execution prerequisites: Check whether it can access the project, required documents, tools, and test environment. Fill any gaps first.
Set decidable success criteria: Turn “fix it” into an observable result, such as passing the target checks, avoiding regressions in specified tests, or meeting a benchmark threshold.
Set up a feedback loop: Provide a test suite, runtime logs, browser screenshots, or a reproducible script. If no metric exists, first ask Carbon to analyze the project and establish a baseline.
Allow it to execute the plan: Once the goal is clear, let the model work across files and iterate based on feedback. Pause for confirmation before irreversible steps, as agreed for the task.
Review the delivery: Inspect the final diff and actual verification results, and confirm that each completion criterion is met. After a long task, summarize repository conventions encountered and capture them in reusable guidance.
The guide's “Align on intent before long runs” section recommends sharing the goal first, letting Carbon raise questions about boundaries and success criteria, and confirming the tools, documents, and intent before starting.
The guide's “Give the model a way to verify its work” section recommends using tests, runtime logs, browser screenshots, or verification scripts to create a feedback loop. When there are clear pass/fail criteria, the model can record a baseline and keep verifying its work.
The official accessibility example asks the model to scan src/routes/, record the violation baseline, group fixes by shared-component root cause, and rerun checks after each change to confirm fixes and check for regressions. This article summarizes that example and does not misrepresent it as a new template.
The guide also recommends using concrete visual references or DESIGN.md to keep the interface style consistent. After the task, review the execution and capture useful repository knowledge in AGENTS.md or a reusable skill.
These are Google's recommendations for using Carbon, not a guarantee that they will work for every task. Results depend on task boundaries, tool permissions, verification quality, and project test coverage.
“All checks pass” only means the specified checks passed. It does not prove that behavior outside the requirements is correct. Review the diff and confirm that verification covers the intended scope.
Before a long autonomous run, confirm the goal, verification signals, and permission boundaries. Pause for confirmation before irreversible actions or actions outside the agreed scope.
Google's automation example assumes a particular repository structure and accessibility checks. Adapt its routes, check rules, and pass conditions to the target project.
In Google AI Studio or a development environment configured for Gemini 4 Carbon, choose a routine engineering task with a clear scope and runnable feedback. Use the template to provide the goal, constraints, completion criteria, and verification method. Observe whether the model first clarifies the goal and checks tool availability, then makes changes, runs verification, and reports the results. The accessibility example can serve as a reference, but replace it with the actual scanning script and pass criteria for the current project.
Gemini 4 Carbon