Twelve months ago I set up a dumb JSON file to remember which AI image prompts I had approved and which I had rejected. It became the most valuable piece of creative infrastructure in the whole image pipeline. After 21 review rounds and close to 500 generated images, the file stopped being a memory aid and started acting like a taxonomy: here are the prompts my model can actually render, here are the ones it cannot, and here are the cues that turn a MAYBE into a GOOD on the reroll.
This is the retrospective. What I shipped, what worked, what didn't, and where the whole thing is still wrong.
What I shipped (the ledger and the tooling around it)
The ledger is a single JSON file at .image-feedback-ledger.json. One top-level slugs object keyed by article slug, each with a history array of {round, verdict, note} entries plus a cues_to_preserve list and a status field. As of this writing it holds 106 slugs and 142 verdict entries across 21 review rounds. The ledger is one component in a larger pipeline; the full wiring between the Mac prompt desk and the Windows GPU box is covered in the hub post.
Three functions do most of the work. record_verdict writes a new entry and runs the note through a regex cue extractor that derives short reusable prompt fragments. apply_cues takes the base prompt for a slug and appends the preserved cues before the next render. parse_verdict_paste reads a block I paste from my phone and returns a list of structured entries ready to record.
The loop runs like this. On my Windows box the dGPU renders a round of heroes overnight. In the morning I open a review page on my phone, tap through each image, and dictate a one-line verdict per slug into the paste format: - slug [GOOD|MAYBE|BAD] -- free text. Back at the Mac I paste the block into the tool, it parses, records, extracts cues, saves. The next generation cycle picks up the cues and reshapes the prompt. The Mac side of this loop is the same desk I run a short AI review block first thing every day from, which is why adding image triage cost me no new habit.
Why a ledger instead of vibes. A few hundred images is already too many to hold in working memory. My MAYBE verdict on round 3 of capi-purchase-event-hash said "I want these shapes to take up more of the frame." By round 7 I had forgotten I ever said that. The JSON remembered. The cue extractor tagged the slug with "tight macro zoom so the hero fills at least two-thirds of the frame" and the next generation carried that note forward automatically. The generation itself runs through a safety wrapper that keeps the Mac from swap-thrashing under concurrent Flux loads, so rounds can queue deep without me babysitting the laptop.
What the AI image prompt taxonomy says works
Across the ledger, the categories that landed at a usable hit rate all share a shape: one subject, clear placement, abstract enough that the model does not have to interpret a metaphor.
Atmospheric single-subject fantasy landscapes hit 77 of 150, a 51% rate. The winners were things like aurora tundra, salt flats mirror, slot canyon, obsidian lava field, ice chasm depth, valley of standing stones, red mesa citadel. No people, no vehicles, no architecture beyond a dominant natural form.
Monoliths landed at 27 of 110, a 25% rate overall but tightly clustered. Monoliths in nature with a single earthly placement worked (mesa top, cliff edge, glacier peak, mirror lake). Multi-monolith formations worked (triangle, ring, parade line, pyramidal stack). Abstract voids worked (aurora void, wormhole opening, liquid mirror plane). Everything else in the monolith category underperformed.
Dune came in at 26 of 66, a 39% rate, splitting cleanly into three buckets: desert landscapes, architectural interiors like atreides palace and sietch interior pool, and planetary orbital views. None of the Dune winners included vehicles. None had more than one subject.
Iridescent glass macros (the brand aesthetic I use for hero images on this site) sat at 7 of 30, a 23% hit rate, with the same pattern underneath. Simple geometric primitives won: single cube, triangular prism, flat horizontal slab, thin layer stack, grove of spikes. Complex or organic forms lost. The glass-macro result echoes something I found when comparing the prompt behaviors of Z-Image Turbo and Flux Schnell directly: Z-Image prefers the primitive, Flux tolerates more surface drama.
The common thread, stated as bluntly as I can state it. The model renders one thing well. When I give it one thing to render, it renders. When I give it a metaphor or a diorama or a small-human-in-a-big-scene, it renders something that looks AI-generated in the bad way.
What the taxonomy says never works
The negative side of the taxonomy is more useful than the positive side. Knowing what to avoid saves batch time.
Spaceships: 0 for 14. Absolute fail across every variant I tried: starships in orbit, dropships in atmosphere, frigates in dock, abandoned generation ships. Never prompt Z-Image Turbo for a spacecraft as the primary subject.
Hooded human figures in portals or caves: 2 for 15. The model cannot hold a small human silhouette and complex portal geometry at the same time; one of them always comes out warped.
Dune vehicles: 0 for 16. Ornithopters, frigates, dropships, heighliners, harvesters, convoys. All bad. The moment the prompt asks for a vehicle the render loses cohesion.
"Looks-like" metaphor prompts: dragon-spine ridge, kraken island, serpentine reef, sky whale skeleton, leviathan skeleton. The model reads the animal name and renders the literal animal instead of a ridge or island that evokes it. Banned until a different model handles it better.
Small architectural specifics on distant cliffs: cliffside monastery network, bridged monastery peaks, hidden mountain temple, glacier temple ruins. Small buildings on large geography do not resolve at 1024px. Either the buildings get smeared or the geography does.
What worked in the ledger itself, and what didn't
The review-on-the-phone flow worked. The format is - slug-name [GOOD|MAYBE|BAD] -- free text note. I can tap through 40 images in five minutes while waiting for coffee, and the paste block parses clean because the regex is strict about the brackets and the double dash.
The regex cue extractor worked better than I expected. The rules started at eight patterns in round 1 and grew to about thirty by round 6 as I noticed recurring notes. "Doesn't take up enough of the frame" became a cue. "It's too centered" became a cue. When I see a new phrase repeatedly in MAYBE notes I add a rule, and the next generation carries the cue forward on every affected slug without me re-editing prompts.
The saved capi-purchase-event-hash entry has a round 3 MAYBE note asking for the shapes to fill more of the frame. Its preserved cue calls for a tighter macro view. The snapshot does not contain a later GOOD verdict for that slug, so it supports the captured correction rather than a claimed round 7 result.
What did not work: encoding the judgment by verdict alone. The MAYBE notes carry most of the signal because they are the rerolls I can rescue. The BAD notes often said things like "hey guys" or "nope" because I was triaging fast on the phone.
The first cue extractor pass also missed most of the patterns. It had eight rules. I spent rounds 4 through 6 expanding the rule set against what I was actually saying on the phone, and by round 10 it stabilized.
“The MAYBE notes carry most of the signal because they are the rerolls I can rescue.
”
What I would do differently now
Capture the round number automatically instead of making me type it at the top of the paste. I have typed the wrong round twice, and having the tooling own that removes a class of error.
Split cues into namespaces: style, composition, banned-concept. Right now they are a flat list, which works but does not help me see at a glance whether a slug's problems are about the aesthetic or the framing.
Add a pre-generation filter that hard-rejects known-bad categories before they enter the batch. If a prompt contains "spaceship" or "ornithopter" or "dragon-spine ridge," the generator should refuse and ask me to reword. Nothing currently prevents me from waste-rendering the same failure again.
Keep the free-text notes above everything. The regex taxonomy is second-order; the notes are the actual data. If I had to rebuild the tooling from scratch the notes are what I would save and the rules are what I would regenerate.
What the next six months will test
Does the taxonomy transfer to a different model. Most of what I have learned is Z-Image Turbo-specific. When I move a batch to Qwen Image or to a newer Flux release, the hit rates will shift and some of the banned categories may open up. The ledger shape is portable; the cue rules are probably not.
Does cluster-uniform styling beat per-article bespoke prompts. I have been generating one bespoke hero per article, which is expensive in both render time and review time. Moving to one uniform aesthetic per cluster of articles would let me trade bespoke variety for consistent throughput. I want to see whether the reader actually prefers the variety or whether cluster-level consistency reads as more professional.
Does a library-first strategy eat the problem. If I keep building the approved pool, eventually 80 to 90% of new articles can pull a hero from the existing library and only 10 to 20% need bespoke rerolls. At that point the ledger stops being a prompt-iteration tool and starts being a curation tool. Different UX, different workflow, probably better economics. I have already run the library-first version of this once, generating a 174-image brand library for a pre-launch legaltech startup.
At what point does the feedback loop's value curve bend. Round 1 moved the hit rate from unknown to roughly 40%. Round 10 was north of 60%. I do not expect round 30 to be at 90%; diminishing returns are real and at some point the remaining failures are model-capability failures, not prompt failures. I would like to know where that floor is. If you want to see the full creative-tech stack around this loop, the live case studies and the product suite both show how the image pipeline ties into client work and into the shelf of productized offers.
Comparing models without mixing the evidence
The larger taxonomy notes describe roughly 500 generations across 21 review rounds. Its saved ledger snapshot contains 106 slugs and 142 verdict entries. Generations, rounds, and recorded verdicts count different things. The sanitized ledger excerpt preserves one real review note and its extracted cue, with snapshot counts checked in October 2026.
Earlier category totals remain observations from that workflow: atmospheric landscapes at 77 of 150, monoliths at 27 of 110, and the Dune category at 26 of 66. They describe the operator's selected batches and review criteria. Treat another model configuration as a fresh measurement problem.
A separate set of cross model notes covered about 40 fantasy prompts. Animal named geography failed in both pipelines in different ways: Z Image tended toward the literal animal, while Flux could retain the geography and still introduce unwanted animal forms. Hooded figures were weaker in the Z Image batch and marginally usable in some Flux outputs. Layered scenes and ornate surface detail often routed to Flux; isolated geometric macros often routed to Z Image. Those are qualitative routing observations from the notes, rather than matched aggregate scores.
A schema for the next ledger revision should record the model and configuration on each generation. For example:
type Review = {
slug: string;
round: number;
model: string;
configuration: string;
verdict: "GOOD" | "MAYBE" | "BAD";
note: string;
cuesToPreserve: string[];
};
That is a proposed extension. The preserved snapshot uses slug history entries with round, verdict, and note, plus cues stored at the slug level. Adding model metadata prevents two pipelines from appearing to support the same rule when the notes came from different configurations.
Regenerate cue rules by collecting the repeated corrections in MAYBE notes, grouping them by model, and testing a candidate rule on a held out set of prompts. Keep the original note beside the derived cue. After a model swap, retain history and rerun the rule review before applying the old cue pack to a full batch.
Sources and specifics
- Ledger location:
/.image-feedback-ledger.jsonin the solo operator's workspace root. 106 slugs, 142 verdict entries, 21 rounds captured as of this writing. - Tooling:
.image-work/image_tooling.pyexportsload_ledger,save_ledger,apply_cues,record_verdict,parse_verdict_paste,slugs_needing_reroll. Approximately 30 regex cue rules as of round 11. - Category hit rates are from a single operator's review of ~500 generations on Z-Image Turbo, rounds 8 through 18: fantasy landscapes 77/150 (51%), monoliths 27/110 (25%), Dune 26/66 (39%), portals 10/45 (22%), iridescent glass 7/30 (23%), sci-fi 2/16 (12%), spaceships 0/14 (0%), epic fantasy 6/20 (30%).
- All verdict notes are free-text dictated on a phone during morning review; the regex extractor is pattern-based and lossy by design.
- The pipeline runs Mac-for-prompts and Windows-for-GPU; see the hub article on the cross-machine setup for the full wiring.