Build better prompts. Create better images.
Promgrammer is a free AI image prompt builder with 1,789+ curated tags across 18 categories and 165 sub-categories. Instead of typing prompts from scratch, you browse a visual tag library, click to assemble, drag to reorder, and copy the finished prompt into Stable Diffusion, Midjourney, NovelAI, or any comma-separated-tag AI model.
Writing good AI image prompts is harder than it looks. You need to know which tags actually work (most "beautiful" adjectives don't), in what order they should appear (earlier tokens carry more weight), and how to combine style, subject, and negative prompts without cancelling each other out. Promgrammer encodes all of that knowledge into a clickable interface so you can focus on the creative decisions, not the tag dictionary.
The tool is built for three audiences: newcomers who want to see what's possible, intermediate users tired of Googling "good anime hair tags" for the hundredth time, and power users who want a personal catalog of custom tags next to the curated library.
There is no canonical, maintained, multilingual catalog of AI image prompt tags. The closest equivalents are scattered across Danbooru wikis (focused on a narrow style), Reddit threads (unstructured), and paywalled prompt marketplaces (which sell whole prompts but don't teach the underlying tags).
That leaves three recurring friction points for prompt writers:
rim_lighting, side_lighting, or dramatic_lighting.AI image models don't "understand" prompts the way you do. They break the text into tokens, match those tokens against concepts learned during training, and steer the image toward the closest visual match. Three mechanics are worth knowing before you assemble your first prompt — and Promgrammer's interface is shaped by all three.
Each tag is converted to one or more tokens. Stable Diffusion uses CLIP's ~49,000-token vocabulary; NovelAI uses a similar table. Common tags like 1girl or masterpiece are single tokens. Rarer compounds — an unusual character name, a compound style — can split into three, four, or more tokens. SD 1.5 has a hard 75-token limit per prompt segment; SDXL allows longer but effective results usually cap around 150. Promgrammer's click-to-add approach helps you stay within these limits without counting manually.
Earlier tokens carry more influence on the generated image. The first 5–10 tokens dominate overall composition; tokens past position 40 behave like weak suggestions. This is exactly why the recommended order is Quality → Subject → Details → Background → Style. When you drag a tag in Promgrammer, you are literally changing how much weight the model gives it. Dragging golden_hour from position 18 to position 3 can change a dim indoor scene into a cinematic sunset portrait — same tags, different prompt.
SD WebUI and its forks support (tag:1.2) to boost a tag's influence by 1.2×, and [tag] to reduce it to 0.9×. NovelAI v4 uses {tag} and [tag] with stacking ({{tag}} is stronger). Promgrammer outputs plain tags on purpose — you add the weight syntax only when you need it, after deciding what matters most. A common pattern: boost the single most important detail (e.g. (golden_hour:1.3)) and leave the rest alone. Over-weighting three or four tags usually produces worse results than boosting one.
masterpiece and best_quality. Most fine-tuned models respond to these as signal boosters.1girl, landscape, portrait.golden_hour) can unify an entire image.low_quality, blurry, bad_anatomy, extra_fingers.After reviewing thousands of prompts — our own and shared from the community — the same handful of mistakes keep surfacing. Knowing them up front saves a lot of regeneration time and credit burn.
beautiful, pretty, cute, attractive all at once doesn't multiply the effect — it dilutes every other tag by spending token budget on noise. Pick one quality adjective and trust the model.low_quality, blurry, bad_anatomy, extra_fingers, watermark, text.anime, photorealistic, oil_painting in one prompt fights itself. The model averages contradictory style signals and the output looks muddy — neither crisp line art nor believable photography. Pick a style family and commit; use Promgrammer's Theme and Composition categories to stay consistent.long_hair. Natural-language models (Midjourney, DALL·E 3, Flux) expect long hair. Promgrammer's tags use underscores by default; swap them for spaces when pasting into natural-language models.full_body poses poorly — the character ends up cropped or squished. Portrait aspect is almost always better for standing characters; landscape for environments. Set the aspect ratio in your generator (SD WebUI "Width × Height", Midjourney --ar 2:3) before tuning tags. It's a bigger lever than most people realize.Rather than dump all 1,789+ tags on one screen, Promgrammer organizes them into 18 categories that map to the decisions a prompt writer actually makes. Each category is a different "axis" of the image. Here's what each one controls and when to reach for it.
Quality markers (masterpiece, best_quality), art styles (anime, realism, pixel art), and resolution hints. These tags tell the model the overall aesthetic direction before it worries about subject.
High-level genres and moods — fantasy, cyberpunk, watercolor, slice-of-life. One theme tag can set the entire palette of the image.
Depth-of-field controls — bokeh, shallow focus, tilt-shift. Useful for photographic realism and emphasizing a subject against a blurred background.
Environments — city street, enchanted forest, open ocean, space station, ancient ruins. Pairs naturally with Lighting and Theme.
Particle-level atmosphere around the subject — aura, sparkles, petals, fog. Adds visual richness without committing to a full environment.
Height, build, and proportions. Separated from Appearance so you can mix body types with any hairstyle or clothing.
Hairstyles, hair colors, eye colors, skin tones. The largest single-axis group — the combinatorial space is intentionally big.
Body position and gesture — standing, sitting, jumping, floating, cross-legged. Works together with Action for dynamic scenes.
Smile, wink, surprised, tears, neutral. A single expression tag can completely change the emotional tone of a portrait.
Verbs — running, dancing, fighting, reading, cooking. Action tags imply motion blur and dynamic composition; models respond well when you combine 1-2 action tags with a Pose.
Color grading — warm tones, cool tones, monochrome, pastel, neon, muted. These are global hints; the model biases the whole palette.
Things you do NOT want in the image — low_quality, blurry, bad_anatomy, extra_fingers. Negative prompts are as important as positive ones, especially for anatomy-sensitive models.
Dress, armor, school uniform, business suit, casual wear, swimsuit. Tag specificity (e.g. "pleated_skirt" vs "skirt") changes results significantly.
Glasses, earrings, crown, sword, backpack. Accessories are the fastest way to signal a character archetype (student, warrior, wizard) without over-describing clothing.
Camera angle and framing — front view, side view, from above, from below, close-up, full body. Angle is underused but has outsized impact.
Child, teen, young adult, middle-aged, elderly. Use cautiously — some models conflate age with body type.
Backlight, rim lighting, golden hour, neon, cinematic lighting, studio lighting. Lighting is the single biggest lever for atmosphere.
Smooth, metallic, glass, fabric, crystalline, rough. Texture tags work best when paired with material descriptors in the subject.
Promgrammer's tag vocabulary is model-agnostic in theory — all tags are plain English keywords separated by commas. In practice, different models respond to the same tags differently. Here's how to get the most out of each major model.
The best match for Promgrammer's tag-style prompts. SDXL and its fine-tunes (Pony, Juggernaut, RealVisXL) excel at dense comma-separated prompts. Front-load quality tags, keep the total tag count in the 15–25 range, and lean heavily on the Negative category. If you're using an anime fine-tune like AnythingV5 or Pony, include score_9, score_8_up style quality tokens where the model documentation says to.
NovelAI is built on Danbooru-style tags, so Promgrammer's vocabulary maps almost one-to-one. Use underscores between words (the curated tags already do this). NAI v4's prompt weight system {tag} and [tag] for emphasis works well after you copy the prompt — add the braces manually in the NAI UI.
Midjourney prefers natural-language prompts over tag lists. Copy the Promgrammer prompt and rewrite the subject portion into a sentence (e.g. "a young woman with long black hair wearing a school uniform, standing in a sunlit classroom"). Style, lighting, and quality tags can stay comma-separated at the end, and the --ar, --stylize, --chaos flags go after that.
These models want full sentences, not tags. Use Promgrammer as a checklist to ensure you haven't forgotten any dimension (lighting? background? pose?), then write the prompt as you'd describe the image to a friend. Negative prompts aren't supported in most of these — just leave them off.
Fuzzy-match search across English tags, Korean labels, category names, and sub-category names. Results group by category so you see context, not a flat list.
Star the tags you reach for constantly. Toggle the favorites filter to hide everything else and rebuild prompts in seconds from your most-used set.
Add your own tags and sub-categories. Useful for LoRA trigger words, personal style preferences, or niche tags not in the default library.
Save a finished prompt with a name, the model you used (SDXL, NovelAI, etc.), LoRA notes, and free-form memo. Useful for "good settings I want to come back to."
Attach generated images to saved prompts. Over time you build a visual index of which prompts produced which results — far faster than scrolling through raw SD WebUI output folders.
Drag selected tags to rearrange prompt order. Drag the category sections themselves to change the top-to-bottom order that drives your default assembly pattern.
Hide tags you never use from browse and search. The default library has 1,700+ tags, but you probably care about 200. Hiding keeps the browse view focused.
Every prompt you've copied is auto-saved. Accidentally cleared your prompt? Recover it from history. Want to see how your prompt style evolved over a month? Scroll history.
Promgrammer's tag library is curated, not exhaustive. We deliberately excluded tags that are either too vague to produce consistent results (e.g. "beautiful", "amazing") or too narrow to justify the browse weight (e.g. hyper-specific character-name tags from a single anime). The goal is a library where every tag, if you click it, meaningfully changes the image in a predictable way.
The 18-category split is the second curation decision. Earlier drafts had 8 categories, 30 categories, and finally 18. The current split reflects the decisions a prompt writer actually makes independently: a scene's lighting choice is independent of its subject's clothing choice, so they're different categories. Angle is independent of composition quality, so they're different categories. This independence matters because it makes the UI combinatorial — you can pick freely from each category without worrying about conflicts between them.
The default within-category ordering (quality before style, common tags before obscure ones) is also curated. It's what an experienced prompt writer would reach for first. New users can follow top-to-bottom and get reasonable prompts; experienced users can jump around via search.
Yes. All core features — browsing tags, clicking to build prompts, copying the result — are free and require no account. Optional features like favorites, saved prompts, custom tags, and the image gallery are unlocked with a free Google sign-in.
No. Every tag has a Korean label shown beside the English prompt text. You can read the Korean meaning while the copied prompt stays in English, which is what AI image models expect.
The tags are curated for booru-style prompt engines, so they work best with Stable Diffusion (SD 1.5, SDXL, SD 3), NovelAI (v3, v4), and similar comma-separated-tag models. Midjourney uses natural language prompts, but selected tags still translate reasonably well when pasted.
Promgrammer itself imposes no restrictions on the prompts you build. Commercial use of generated images depends on the AI model's own terms (check SDXL, NovelAI, Midjourney licenses separately).
AI image models were trained predominantly on English-labeled datasets (Danbooru, LAION, COCO). Korean tags produce much weaker results. We show Korean labels for learning and English tags for copying — the best of both.
Yes — sign in and use the Custom Tags feature to create your own categories and tags. They appear alongside the built-in 1,700+ tags and are searchable.
No. Promgrammer is strictly a prompt builder. You copy the finished prompt and paste it into your own AI image generator (local SD WebUI, ComfyUI, NovelAI, Midjourney, DALL·E, etc.).
Yes. Only you can see your saved prompts, favorites, custom tags, and uploaded images when signed in. See our Privacy page for details.
Most AI image models weight earlier tokens more heavily. A tag at position 1 has more influence than the same tag at position 20. Promgrammer's default category order (Quality → Composition → Subject → Details → Background → Style) reflects best practice, and drag-and-drop lets you override it when needed.
Copy the prompt text with the Copy button — that's the prompt. The tag selections themselves aren't sharable as a link yet, but since all 1,700+ tags are searchable, a friend can reconstruct a prompt in seconds.
Selection is which tags are in the prompt (the vocabulary). Order is where each tag sits in the comma-separated list (the emphasis). Two prompts with identical tags but different orders produce noticeably different images — earlier tokens influence composition more, later tokens fine-tune details. If your output is close but not quite right, try reordering before changing selection.
Three common reasons. First, model checkpoints differ — SDXL base and SDXL Juggernaut share an architecture but respond to tags differently. Second, sampler and CFG scale affect how strictly the model follows the prompt. Third, seed matters more than most people realize: the same prompt on seed 42 vs seed 10000 can look like two different images. If a friend's prompt underperforms for you, match their model + sampler + CFG first before changing tags.
Use tags like 2girls, 1boy_1girl, or multiple_characters. Most booru-trained models handle up to 2 or 3 characters reliably; beyond that, expect anatomy issues. For complex multi-character scenes, generate them separately with a consistent style seed and composite in an image editor — that's usually faster than fighting the model.
Promgrammer itself is model-agnostic and doesn't load LoRAs — your image generator (SD WebUI, ComfyUI, Forge) does that. To use a LoRA with a Promgrammer prompt, sign in and create a Custom Tag like <lora:myStyle:0.7> along with its trigger words. Save those as a single preset in My Prompts. The custom tag then appears in your library next to the curated tags and includes automatically when you click it.
Yes. The layout is responsive — browsing, search, clicking tags, drag-and-drop reorder, and copying all work on phones. That said, assembling a complex 25-tag prompt is physically easier on a desktop or tablet with a keyboard for search. Most users browse on mobile and build prompts on larger screens.
All core features are available without an account — click through to the prompt builder and start assembling. For favorites, saved prompts, custom tags, and the image gallery, sign in with Google. No email marketing, no up-sells, no paid tier — the product is free and supported by display ads.
If this is your first time, the Prompt Writing Basics guide walks through a complete example from zero to a polished prompt in ~5 minutes. From there, the tag index and prompt examples are good jumping-off points for exploration.
Better prompts make better images.
Feedback, bug reports, tag suggestions, or "please add X"? Email siwooeo@gmail.com. Korean or English both fine.