{"id":20523,"plugin_id":"plugins_6aa518f7d9608191aa8b89c668c2fbe1","kind":"skill","collection_source":null,"comparison_source":null,"observed_at":"2026-09-30T23:16:19.041Z","digest":"fa281e9e662a2984a867b0be2e1baf29e099d6aed89bfa43f49716109f24eb65","against":null,"payload":{"name":"prompt-writing","description":"Use when writing, improving, or debugging image generation prompts or choosing prompt parameters.","included_files":[],"skill_md_contents":"---\nname: prompt-writing\ndescription: Use when writing, improving, or debugging image generation prompts or choosing prompt parameters.\nallowed-tools: Bash(astria:*)\n---\n\n# Prompt Writing\n\nAlways write generation prompt text in English, even when the user communicates\nin another language. The surrounding conversation may remain in the user's\nlanguage, but every prompt shown to the user or sent to Astria must be English.\n\nBefore writing a prompt, get to know the user with the `astria` CLI (see the\n`astria-api` skill):\n\n- `astria prompts list --limit 20` — their recent prompts, to learn their style and parameters.\n- `astria packs list` — their packs. If the request is about a specific subject (e.g. a shirt), find a pack whose `main_class_name` matches and inspect its template prompts with `astria prompts list --pack-id `.\n- `astria tunes list` — their references/tunes. If any match what they want to generate, reference it in the prompt via ` name` syntax (see the `astria-api` skill).\n\nDo not send the user off to browse packs or tunes themselves — query with the\nCLI and bring concrete suggestions back to them. When offering choices, your\nask-user question tool (`AskUserQuestion` in Claude Code, `ask_user` in the\nAstria chat agent) with image thumbnails helps them pick.\n\nIf no tune matches the request, follow intent:\n- For headshots / models / avatars, do NOT ask for a reference first. Propose ready-to-generate prompt options and trait options (look, age range, styling, framing, lighting), then confirm generation settings.\n- For product or person-specific identity requests where likeness matters, ask the user to provide a reference image (upload it as a tune with `astria tunes create`, or drag it into the prompt box in the web app).\n- Do not suggest web search for this flow.\n\nAfter writing a prompt, show the user the prompt text and let them review or\nedit it before generating. Confirm how many images per prompt (via an\nask-user question) before calling `astria generate`.\n\n# Types of request\n\n1. **E-commerce / product shots** — reference tunes to create a new image.\n2. **Image editing** — pass `astria generate --input-image ` to modify an existing image (change background, change style, add/remove objects).\n3. **Upscaling** — `astria generate --model gemini --text \"Recreate this image in 4K\" --input-image `. If the image has prominent text labels, mention them in the prompt so they survive the upscale.\n\n# Prompt Writing Guide\n\nBy default generate with the **gemini** model (`astria generate --model gemini`).\nNo need to ask the user about model type unless they explicitly mention another\none or ask for a recommendation.\n\n## Default intent mapping\n\nIf the user asks to create \"models\", \"avatars\", or \"headshots\", interpret this\nas a request for realistic, unique face headshots on a clean white studio\nbackground:\n- Do not use any reference tune for this request type.\n- Keep framing as a face headshot (not full body), white `#fff` studio backdrop, studio lighting.\n- Proactively offer prompt/trait choices with an ask-user question instead of asking for references.\n- Treat nationality words (\"israeli\", \"french\", \"japanese\") as styling/trait guidance for facial features and casting diversity, not as a request for references.\n- Do NOT suggest fashion, lookbook, lifestyle, full-body, or outfit-led prompts for this request type unless the user explicitly asks for those.\n\nSee the `unique-headshot` skill for the detailed face-trait template.\n\n## Interaction rule (strict)\n\nNever ask a question and suggest prompt text in the same response.\n- If you use the ask-user question tool, return only questions/options in that turn.\n- After the user answers, return the prompt suggestion(s) in the next turn.\n\n## References\n\nFor a reference tune always use ` tune.name` to reference\nthe trained subject:\n- `portrait of woman in a garden wearing dress` — CORRECT\n- `portrait of John in a garden` — WRONG (the model doesn't know \"John\")\n\nThe trailing `:1` is a fixed part of the token syntax — it is NOT a weight or\nstrength. Always write `:1`; never vary it, and never tell a user to change it.\n\n## Reviewing bad results\n\nIf the user says results are bad, figure out what went wrong:\n1. Inspect each reference's `orig_images` (`astria tunes get `) and check the `name` matches the image content. If `name=woman` but the image is a full-body shot including clothes or a hat, tell the user to re-crop the training images on that tune's page (`/tunes/` → \"training images\" → crop tool).\n2. The `name` must represent the main subject. \"jewelry\" is a poor name — it should be \"ring\" or \"necklace\" depending on the subject. If a tune named \"jewelry\" holds a ring, suggest renaming it (`astria tunes update --name ring`) and retraining.\n3. Distorted or competing faces (e.g. a LoRA and a FaceID of the same person in one prompt): turn on \"Inpaint faces\" in the composer settings, or remove the extra face reference (✕ on its chip). Do NOT suggest adjusting the numbers inside reference tokens — there is no reference-weight control.\n\n## Key parameters\n\n| Parameter | CLI flag | Common values |\n|-----------|----------|---------------|\n| Prompt text | `--text` | Required |\n| Number of images | `--num-images` | 1–4 |\n| Aspect ratio | `--aspect-ratio` | `1:1 16:9 9:16 21:9 9:21 3:2 2:3 5:4 4:5 4:3` |\n| Resolution (gemini only) | `--resolution` | `1K`, `2K` (default), `4K` |\n\n## Tips for better results\n\n- Include background descriptions: \"clean white studio\", \"blurred urban street\", \"autumn forest\".\n- For product shots, describe the surface and arrangement.\n\n## Keeping the background identical across generations\n\nGemini / Nano Banana does not hold a backdrop exactly from one generation to\nthe next — the shade and brightness of a studio background drift even when a\nbackground reference is attached. Wording cannot fix this; the reliable fix is\nthe post-processing flag, appended to the prompt text:\n\n```\nastria generate --text \" woman in a studio --background_color #f2f0ed\"\n```\n\n`--background_color #RRGGBB` recolors the detected background to that exact hex\nafter generation, so a whole catalog lands on the same backdrop. It runs on\nGemini / Nano Banana, partner models and Kontext — not on Flux. In the web app\nit is the \"Background color\" row in the composer's options cog.\n\nTwo related mistakes to check when a user reports a drifting background:\n\n- Text like \"match the provided background reference\" with no background\n reference actually attached — the model then invents a backdrop every run.\n Attach it as a reference (` background`) or describe the exact\n colour instead.\n- Five or six references stacked in one prompt (subject + outfit + shoes + hat\n + bag + background). They compete, and the background gives way first — keep\n the essentials and add accessories in a second pass.\n\n# Fashion and garments\n\n1. Always work with a consistent face reference. If the user has none, suggest a faceid tune from the public gallery (`astria tunes list --gallery --model-type faceid --limit 200`) or generate a face first (no reference) and convert one of those outputs into a tune with `astria tunes create`.\n2. Figure out the intent — a lookbook (e.g. prompt `look book plain white background #fff`) or a campaign shot. Campaign example: `A direct flash paparazzi style shot of woman moving through a crowded bar. She looks straight into the lens with an intense expression. She wears the dress and the bag on her shoulder. The background is dark and out of focus. High contrast, sharp flash.`\n"},"changes":[],"summary":"First saved snapshot. No earlier version is available for comparison.","summary_kind":"deterministic","summary_metadata":{}}