Gemini 4 Hydrogen: test a client caption fact check before using it
Google introduced Gemini 4 Hydrogen on November 12. If it is available to your account, try a text-only caption check with fictional inputs. Use the same prompt and captions in a fresh chat for each assistant you compare, check which model generated each response where the app shows it, and score the decisions against the known answers. Keep all captions with unsupported facts or unclear client scope out of the publishing handoff.[1][2]
Choose one job to evaluate after the announcement
Google introduced Gemini 4 Hydrogen on November 12 and listed the Gemini app among the places it is available. The Gemini help explains that usable models vary with the plan and remaining usage, and shows how to check which model produced a response. Record the model actually used for this exercise. If you cannot verify Hydrogen, report the run under the verified model or as model unknown.[1][2]
An agency can use a synthetic caption audit as a small acceptance test for new assistants. The job is to check whether a draft preserves approved facts, belongs to the requested client and avoids prohibited commitments. Use the browser chat you already have access to. The exercise needs text only and calls for no product connection or posting permission.
Give each assistant the same brief and six captions
The fictional client is Juniper Repair Studio, workspace JR. The approved facts are: the studio offers zipper repairs; staff review each garment at the counter before quoting; estimates depend on what the team sees; and the only approved invitation is to bring a garment to the studio for an assessment. The client authorizes no quoted prices, discounts, rankings, completion times, opening hours, pickup promises or repair guarantees. Oak Tailoring, workspace OT, is a different fictional client outside this task.
The worksheet supplies all six complete draft captions, their stated workspace and their reference dispositions. For the actual exercise, give each assistant only the brief and the first two worksheet columns. Keep the decision, reason and next-step columns out of its input. Ask it to leave the wording and record IDs intact so a reviewer can line up the outputs.
Use the same prompt and a separate answer key
Reusable prompt: Review each supplied caption against the Juniper brief and its stated workspace. Return exactly one row per ID with its original caption, disposition and a short reason tied to the brief or workspace. Quote a conflicting phrase for each issue in a held caption. Allowed dispositions are Ready for human review, Hold and Exclude. Exclude a row outside JR before judging its wording. Hold a JR row if it adds any unsupported or prohibited detail. Otherwise return Ready for human review. Identify every issue in each held row. Do not fix copy, infer missing facts, open accounts, schedule or publish.
Start a fresh chat for each assistant and model you compare and keep the brief, prompt and captions identical. Record the date, displayed response model where available and any setting you changed. Check each of the six IDs against the worksheet, counting missing or repeated IDs as errors. Give one decision point per ID listed exactly once with the correct disposition. Separately check whether the reasons catch all four held-caption issues: price, turnaround, pickup and ranking.
The reference result is two ready for human review, three held and one excluded. For example, if an assistant marks R3 ready without flagging the $12 claim, catches only pickup in R5 while correctly holding it, and gets everything else right, it scores five of six decisions and two of four issues. It has one false release because R3 should have been held. For this exercise, require six correct decisions, all four issues and zero false releases before trying the same process on separately approved material.
Keep the caption review tied to the client workflow
Read the approved brief and full caption yourself before marking a real draft approved. Record the exact client workspace, owner and version after the human review. If a model misses an issue, revise the instructions or stop using it for this job. Save that first result and use a fresh set of deliberately difficult captions for a follow-up check.
AgencyPro’s API documentation describes social publishing in separate client workspaces and optional approval routing for API-created posts. The caption check can precede your existing review and publishing steps. Use the AgencyPro overview if you need client workspace separation and a repeatable handoff. This example makes no claim that Gemini connects to Sociamonials or runs its approval flow.[3]
Six fictional captions and the reference answer key
| ID and supplied workspace | Full candidate caption | Reference decision | Evidence for the decision | Next step |
|---|---|---|---|---|
| R1 / JR | Need a zipper repair? Bring your garment to Juniper Repair Studio for an assessment. | Ready for human review | The service and assessment invitation match the brief. | Reviewer checks the exact copy and intended workspace. |
| R2 / JR | Juniper Repair Studio reviews each garment at the counter before quoting. Bring yours to the studio for a zipper repair assessment. | Ready for human review | Counter review precedes quoting and the invitation is approved. | Reviewer confirms the current brief before deciding whether to approve. |
| R3 / JR | Bring your garment to Juniper Repair Studio for a zipper repair assessment. Repairs start at $12. | Hold | The $12 starting price is unsupported. | Ask the client to confirm an approved price or request copy without a price claim. |
| R4 / JR | Juniper Repair Studio offers zipper repairs and will finish your garment the same day. Bring it to the studio for an assessment. | Hold | Same-day completion adds an unsupported turnaround promise. | Ask the client reviewer for accurate timing or remove the promise in a separate revision. |
| R5 / JR | Juniper Repair Studio offers the city’s highest-rated zipper repairs. We’ll pick up your garment tomorrow. | Hold | Highest-rated and pickup are both unsupported. | Keep the draft out of review until both statements have been addressed. |
| R6 / OT | Need a zipper repair? Bring your garment to Juniper Repair Studio for an assessment. | Exclude | The supplied workspace is a different client even though the caption matches R1. | Leave OT untouched. Verify the actual workspace associated with the request. |
When this approach needs adjusting
- The clients, captions, example score and expected answers are fictional teaching materials prepared with Codex. No Gemini or other assistant run, real account check or publishing test was performed.
- The scoring example illustrates the rubric. Actual runs of this exercise would evaluate this caption set only, without establishing performance on other work or customer outcomes.
- Model availability and actual model used depend on your account and usage. Verify the response details and human approval before using any proposed draft.
Return to this next week
Take a fresh set of fictional cases with the same rules. Record the model, six decisions and every supporting phrase. Track false releases separately and compare the exact drafts with their client briefs before authorizing real publishing.
See how AgencyPro fits your agency