gpt-image-2-techniques

117-technique catalog wrapped around kie.ai's GPT-Image-2 endpoint, with a discovery step that proposes the top 3 best-matching techniques per request and lets the user pick. Brand-locked. Production-ready.

Shared Infrastructure

• API client: kie_client.py (included in this skill folder — reads KIE_AI_API_KEY from environment)
• Brand kits: brand_kits/ (local folder — acme-consulting template; copy it for your own)

Architecture

gpt-image-2-techniques/
├── SKILL.md                         (this file — operating instructions)
├── kie_client.py                    (kie.ai API wrapper: submit_text_to_image, submit_image_to_image, poll, generate)
├── techniques/
│   ├── _index.md                    (one-liner directory of all 117 techniques — read at discovery time)
│   └── library/
│       ├── 01-brand-identity.md           (13 techniques, full prompt templates)
│       ├── 02-editorial-info.md           (10)
│       ├── 03-web-app-ui.md               (13)
│       ├── 04-marketing-ads.md            (15)
│       ├── 05-multi-tool-stacking.md      (7)
│       ├── 06-reverse-style-transfer.md   (14)
│       ├── 07-personalization-portraits.md (13)
│       ├── 08a-music-brand.md            (10)
│       ├── 08b-client-b2b.md              (10)
│       ├── 08c-content-research.md        (11)
│       └── 09-video-keyframe-bible.md     (1 — keyframe asset for video pipelines)
└── brand_kits/
    ├── _index.json                  (registered kits)
    └── acme-consulting.json         (generic starter template — copy and customize for your brand)

The discovery workflow

Every request runs through these 4 steps. Do not skip discovery.

Step 1 — Parse the request

Identify these dimensions from the user's natural-language request:

1. Output type: image-only (always for this skill — img2img is fine, but never video)
2. Specific deliverable: share card, landing hero, dashboard, PDF cover, carousel slide, ad, infographic, album cover, magazine, mockup, etc.
3. Audience:
   • B2C consumer (broad)
   • B2B enterprise (HR/L&D directors, formal register for HU)
   • Personal/portfolio (your own brand)
4. Brand fit: acme-consulting / your-own-kit / none — generic
5. Language: en / your-language / bilingual
6. Aspect/format hint: square, vertical, wide, print-ready, etc.
7. Mode: text-to-image or image-to-image (img2img if user mentions "based on this", "similar to", "use this as a reference", or attaches/references existing assets)
8. Stylistic register: premium-restrained / cinematic-mythic / hand-drawn-friendly / editorial-illustration / sepia-academic / etc.
9. Hidden constraints: any avoid list, must-include text, etc.

Step 2 — Scan the index, propose top 3

1. Read techniques/_index.md once (always do this; it's the discovery directory).
2. Match candidates against the request dimensions. Prioritize:
   • Output deliverable match (highest weight — share card request → share-card techniques)
   • Audience match (B2B request → B2B-flavored techniques)
   • Score (★★★★ and ★★★★★ outrank ★★ and ★)
   • Foundational tags — when the request is about brand consistency, multi-asset coherence, or "do this for everything," elevate [foundational] techniques (a1-t3, a3-t2, a3-t3, a3-t10, a6-t2, a8c-t2)
   • Synth tags — when the request is about a multi-step deliverable or workflow, prefer [synth] entries (these are recipes, not single prompts)
3. Pick the 3 best matches spanning at least 2 categories where possible (give the user variety in approach, not 3 takes on the same idea).
4. For each pick, prepare a concise rationale: why it fits, what makes it different from the other two picks, what variables the user will need to provide.

Step 3 — Present and wait for user selection

Output a structured proposal in this exact format (markdown table or list — pick whichever fits the response register):

Based on your request "[paraphrase]", here are 3 techniques that fit:

1. **[ID] [name]** ★★★★ — [rationale: why this fits, what variables needed]
   _Strength:_ [the one thing this does best]
   _Trade-off:_ [what you give up vs the other options]

2. **[ID] [name]** ★★★★★ — [rationale]
   _Strength:_
   _Trade-off:_

3. **[ID] [name]** ★★★ — [rationale, often this is the wildcard or premium reinterpretation]
   _Strength:_
   _Trade-off:_

Which one do you want me to run? Or describe what's missing and I'll re-propose.

If the request is unambiguous and one technique is overwhelmingly the right answer (e.g. "I want a Pokémon-style trading card for X" → a7-t5 is the only real match), still show 3 but make it clear option 1 is the canonical pick — don't overwhelm with weak alternatives.

Step 4 — Execute

Once the user picks (or specifies):

1. Read the full technique from the appropriate library/{category}.md file.
2. Load the brand kit from brand_kits/{name}.json (default: acme-consulting; copy it as a starter for your own kit).
3. For Hungarian content: Apply correct register (Ön for B2B enterprise, te for SME/consumer), native term substitutions (megtérülés not ROI; versenyelőny not "kompetitív előny"), compound-word rules, correct punctuation, and date format (2026. április 18.).
4. Assemble the final prompt:
   • Start with the technique's prompt template
   • Substitute all {placeholders} with user-provided variables (ask for missing required vars)
   • Inject brand kit fields (palette hex codes, typography, logo description, voice, hard negatives)
   • Append the brand kit's negatives at the end
   • For HU register: explicitly state "use Ön form" or "use te form (consumer register)" — never let the model default
5. Call kie.ai via kie_client.generate():

   from kie_client import generate
   result = generate(
       prompt=final_prompt,
       aspect_ratio="16:9",     # or whatever the technique specifies
       resolution="2K",          # 1:1 cannot be 4K — use 2K
       input_urls=[...] if mode == "image-to-image" else None,
       output_path="./out/asset.png",
       log_fn=print,
   )

6. Confirm save + show the result to the user. Cost estimate: ~€0.10–0.20.

Hungarian content guidelines

When generating Hungarian text inside images, apply these rules:

• Register: Ön for B2B enterprise (formal), te for SME/consumer (informal)
• Native terms: megtérülés (ROI), versenyelőny (competitive advantage), adatvezérelt (data-driven), bevételnövekedés (revenue increase), costs: költségcsökkentés
• Compound words: single word, no hyphen, no space: termelékenységnövekedés, hatékonyságnövelés; hyphen before suffix on abbreviation: AI-képzés, MI-korszak
• Quotes: „lower-upper" or »guillemet« — never ASCII straight quotes
• Comma before hogy: ALWAYS — non-negotiable
• Date format: 2026. április 18. (note: month in lowercase)
• Name order: family name first (Hungarian convention)
• Decimal separator: comma (4,44 not 4.44); thousand separator: space (1 200 000)

Avoid these known kie.ai failure modes

Brand kit reference (quick)

When the user mentions:

• Your own brand → copy brand_kits/acme-consulting.json, rename it, fill in your palette/typography/voice

If the user doesn't specify a brand and the project context isn't clear, ask before assuming. Alternatively, use acme-consulting as a neutral dark-ink palette for any dark-background digital asset.

Cost + latency

• Cost per image: ~€0.10–0.20 on kie.ai (depends on aspect/resolution)
• Latency: ~80–120s per image, ~7-min hard timeout server-side
• Image-to-image: same cost, slightly slower (~100–140s)

Iteration

If the first generation isn't right:

• Compositional issue (layout wrong, missing elements) → re-prompt with more explicit positioning. Same technique.
• Brand drift (wrong palette, wrong logo) → switch to img2img mode and pass the brand-kit reference images as input_urls (foundational technique a3-t10 / a1-t3).
• Style register wrong → propose a different technique. Don't over-iterate on a wrong base.
• Failed safety filter → soften the prompt (see avoid table above), don't bypass.

When this skill is the wrong choice

• The user wants a video → this skill only does single images. Video generation is a separate workflow (Kling/Runway/Veo via kie.ai's other endpoints) — explain and offer to set up the handoff.
• The user wants vector / SVG editable output → GPT-Image-2 produces raster only. Suggest exporting to Canva/Figma after generation, or tracing in Illustrator.
• The user wants a logo from a description with no constraints → this skill works best for branded assets in established kits. For pure logo design, suggest manual designer brief instead.
• The user wants to edit a real photo of a person with face-swap or identity-spoofing intent → refuse politely, point to safer alternatives.