Turn photo feedback into a checklist for your next shoot

"Too cluttered" is not an action. "Cut the bright sign at the edge" is. "Lacks depth" is not an action. "Add a foreground layer and increase subject-background distance" is.

Most AI critiques and human mentors give descriptive feedback — they tell you "what" but rarely "how." The key to translation is converting every descriptive comment into a specific operation you can perform on location.

"Cluttered" → Next time scout the background first, move to avoid distractions, or use a longer focal length to compress.

"Exposure is off" → Meter on highlights instead of average, or enable highlight alert.

"Colors are flat" → Try shooting during golden hour or blue hour, or add local contrast in post.

"Subject doesn't stand out" → Use a wider aperture, foreground/background separation, or find color contrast to emphasize the subject.

Compress the last review into three reminders or fewer, then give each a condition you can directly see in the next frame. For example, "move to avoid the background sign" becomes "no bright text or intersecting line touches the subject's outline."

Why three or fewer? On location you are already processing light, scene, subject, composition, and camera settings. Working memory is limited. Ten checklist items will probably become zero remembered items; without success conditions, the post-shoot verdict also becomes a vague "it feels better."

Different genres have very different priorities. Here are some starting templates you can customize based on your weaknesses:

Street: ① Clean background ② Clear subject hierarchy ③ Decisive moment timing.

Portrait: ① Catchlight and face lighting direction ② Subject-background separation ③ Pose and expression guidance.

Landscape: ① Foreground structure ② Timing window (golden/blue hour) ③ Sky-to-ground ratio.

Food: ① Light angle (top or side) ② Props shouldn't upstage the dish ③ Warm, appetizing color palette.

Architecture: ① Perspective correction and verticals ② Light emphasizing form ③ Simplify surroundings.

Suppose the last street critique says: "The background is too cluttered and the subject is lost in the environment." Translate it into: ① Find a solid wall or open sky ② Use 50mm+ to compress the background ③ Wait until the subject enters a clean zone. The success condition is an intact subject outline with no bright sign or overlapping person around it.

After the retake, select the original review in PicSpeak Retake Coach and upload the new frame. One GPT-5.6 Terra comparison rescores both images, shows before/after changes across five dimensions, cites visible evidence, and proposes the next action. The server calculates the deltas rather than asking the model to do arithmetic.

If the images pursue different goals or the scenes are too different, the analysis remains available but is marked non-comparable and excluded from the progress curve. This feedback → translate → execute → evidence cycle is what turns critique into training.

First check whether the stated success condition appears, then see whether the relevant score moved in the same direction, and finally read the visible evidence. When all three agree, keep the action. If only the number rises, do not call the experiment successful yet.

A flat result is still useful. The action may have been too subtle, the pair may not be comparable, or the strategy may be wrong. Change one variable in the next round and keep using the same source-review chain so the useful signal remains legible.