Build a repeatable workflow

Use this topic-specific sequence: sanitize → orient → observe → infer → request evidence → verify. Give each stage one input, one output, and one gate. The first run should be narrow and reversible. Later automation is earned by measured reliability, not by how easy it is to connect tools.

For diagnose a confusing dashboard screenshot using image evidence plus written context, begin with the job card and sanitized packet. Run the constrained prompt:

Inspect this redacted checkout dashboard screenshot. Goal: explain why the conversion total appears inconsistent. First inventory only visible labels and values with locations. Separate observation from inference. Then propose three hypotheses and the next screenshot or log needed to test each. Do not identify people, infer hidden fields, or invent unreadable text.

Save the response beside its prompt and input version. Then apply the quality rubric and zoom and compare source pixels, transcribe critical values manually, reproduce in the product, inspect logs, and reject claims unsupported by visible or supplied evidence. A failed check returns to the smallest responsible stage; do not regenerate everything. If the source was missing, repair context. If the instruction was ambiguous, repair the prompt. If the candidate violates policy, stop and escalate rather than prompt around the policy.

Name an owner for source approval, generation, verification, and release. One person may hold several roles on a small project, but the role changes should remain visible. The reviewer needs the evidence packet, not merely the final artifact.

Define operational states: draft, needs evidence, blocked, approved, released, and rolled back. This vocabulary prevents a plausible draft from being mistaken for an approved result. Attach timeouts, retry limits, and an off switch to any automated stage.

Log the defect category rather than just “bad output.” This chapter's recurring defects are hallucinated small text; mixing observation with inference; ignoring chart scale; accidental exposure around the crop; treating an image as current truth. Track their rate on representative cases. Review false positives and false negatives separately when classification is involved; track factual, continuity, or rights defects when producing media.

The end product is a multimodal evidence packet with redaction checklist, annotated image, observation/inference table, follow-up requests, and verification result. Point to regions, labels, frames, or timestamps. When the model cannot read a detail reliably, provide a crop or transcription instead of asking it to guess harder. Periodically rerun a stable set of cases after changing models, prompts, source material, formulas, or settings.

Imagine the independent check fails after release. Identify how to stop distribution, identify affected outputs, restore the last approved version, notify the owner, and preserve enough evidence to learn. A workflow without rollback is only a happy-path demo.