15 Best GPT-6.1 Terra Prompts for Visual Analysis and Computer Use

Good GPT-6.1 Terra prompts define the decision, the available evidence, and the required output. They do not ask the model to “look carefully” and hope for a reliable result. Visual tasks improve when the prompt separates observation from inference, names the review criteria, and explains what to do when the image is inconclusive.

The templates below are starting points. Replace bracketed text, test them on representative examples, and preserve human approval for consequential computer actions.

A Prompt Pattern That Scales

Use five components:

1. Goal: the decision the output should support.
2. Evidence: the images, files, or current screen the model may use.
3. Criteria: the exact features to inspect.
4. Boundaries: prohibited assumptions and actions.
5. Output contract: fields, order, and evidence requirements.

This structure is more dependable than adding many adjectives or requesting hidden reasoning.

1. Compare Two Images

Compare Image A with Image B for [identity, layout, color, visible text, and missing elements].
Create three sections: confirmed matches, confirmed differences, and uncertain observations.
For every difference, name the visible region that supports it. Do not infer causes that are not visible.

Use this for design revisions, product photos, storyboard continuity, and before-and-after checks.

2. Review a Campaign Asset

Review this asset for publication in [channel and dimensions]. Check crop safety, text readability,
logo clearance, contrast, visual hierarchy, and required disclaimer placement.
Return a table with issue, visible evidence, severity, and a specific correction.
If a criterion cannot be verified from the image, mark it Not Verifiable.

3. Match a Product Photo to a Catalog Record

Determine whether the supplied product photo is consistent with this catalog record: [record].
Compare visible model name, color, shape, included accessories, packaging text, and distinguishing marks.
Return Match, Mismatch, or Needs Human Review. Cite evidence and do not guess hidden specifications.

4. Extract Visible Text With Context

Transcribe all legible text in reading order. Preserve capitalization where visible.
For each text block, describe its approximate position and purpose.
Use [unclear] for characters you cannot read; do not silently repair uncertain text.

OCR-style output should be checked when exact text carries legal, financial, or safety consequences.

5. Interpret a Technical Drawing

Summarize only the information visible in this technical drawing.
List labels, dimensions, callouts, and apparent relationships. Separate direct observations from interpretation.
Identify any area where resolution, cropping, or notation prevents a confident reading.
Do not calculate or approve a design unless every required value is visible.

6. Find UI Inconsistencies

Inspect this interface screenshot against the supplied UI requirements.
Check alignment, spacing, truncation, control states, error messages, and keyboard-focus visibility.
Return reproducible observations, not aesthetic preferences. Rank issues by user impact.

7. Compare Desktop and Mobile Layouts

Compare the desktop and mobile screenshots for information parity and responsive behavior.
Identify content that is missing, reordered, clipped, too small, or difficult to operate on mobile.
Do not treat intentional layout changes as defects unless they block the stated user task.

8. Review a Storyboard Contact Sheet

Review these storyboard frames in numerical order. Track character identity, wardrobe, props,
screen direction, location geography, time of day, and shot-to-shot progression.
For each continuity risk, cite the affected frame numbers and propose the smallest corrective action.
Do not invent motion or dialogue that is not represented in the frames or supplied script excerpt.

This prompt can produce a revision brief before a creator generates or animates new assets in a dedicated platform such as Elser AI.

9. Check Character Consistency

Treat the first image as the approved character reference. Compare every later image against it.
Check face shape, hair, eye color, costume pieces, body proportions, accessories, and age cues.
Distinguish acceptable pose or lighting variation from an identity change.
Return findings by image number with visible evidence.

10. Evaluate Composition and Crop Risk

Evaluate this image for [16:9, 9:16, and 1:1] delivery.
Identify subjects, text, logos, and story information likely to be lost in each crop.
Suggest crop windows or layout changes. Do not claim a crop is safe when required margins are unknown.

11. Turn Visual Findings Into Revision Notes

Convert the supplied visual-QA findings into revision notes for a [designer/editor/animator].
Group notes by asset, state the requested change first, and include the evidence and acceptance test.
Remove speculation, duplicates, and vague phrases such as “make it better.”

12. Plan a Safe Computer-Use Task

You are working in an isolated test environment. Review the current screen and propose the minimum
sequence of actions needed to [goal]. Do not perform purchases, submit forms, transmit personal data,
delete content, or change account settings without explicit approval. Stop if the screen differs from the expected state.

Planning and execution should remain separate when an action carries risk.

13. Verify a Completed UI Action

Inspect the current screen after the attempted action. Verify the result using visible evidence.
Report Completed, Not Completed, or Uncertain. Cite the confirmation state, saved value, or error shown.
Do not infer success from the absence of an error message.

14. Triage a Visual Asset Library

Classify each supplied asset by [campaign, product, orientation, visual style, and review status].
Use only the allowed labels. Add a short evidence field and route ambiguous or duplicate-looking items
to Human Review rather than forcing a category.

15. Produce a Structured Visual-QA Report

Audit the supplied images against this rubric: [rubric].
Return JSON with asset_id, issue_type, evidence, severity, recommended_fix, and confidence_basis.
Use an empty issues array when no issue is visible. Never omit an asset. Mark unavailable evidence explicitly.

Use Structured Outputs in the API when the application requires a guaranteed schema; a prompt requesting JSON alone is not the same guarantee.

How to Improve These Prompts

Build an evaluation set with accepted examples, known defects, ambiguous cases, and adversarial screens. Change one prompt component at a time. Measure issue recall, false positives, schema validity, reviewer corrections, latency, and cost.

For visual work, improve the evidence before increasing prompt complexity. Supply the correct crop, legible resolution, reference image, frame identifiers, and the exact policy or creative brief. No wording can recover details that are absent from the input.

For computer use, restrict the environment and tools outside the prompt. A sentence saying “be safe” is not an access-control system.

Frequently Asked Questions

Should I ask Terra to explain its chain of thought?

No. Ask for concise evidence, assumptions, checks, and the final decision. OpenAI does not expose raw reasoning tokens. An auditable answer comes from visible evidence and verifiable outputs, not a request for private reasoning.

Should every task use high reasoning effort?

No. Medium is Terra's default. Compare low, medium, and high on the same evaluation set, and use higher settings only when the quality gain justifies the additional latency and cost.

Can a prompt make computer use safe?

A prompt helps define behavior but cannot replace isolation, action restrictions, approvals, time and cost limits, and outcome verification.

Closing Guidance

The best GPT-6.1 Terra prompt acts like a review specification. It tells the model what decision matters, what evidence is allowed, what uncertainty looks like, and how another person or program will verify the result. Start there, then improve the workflow with evaluations rather than accumulating instructions.