User Tools — Image Generation

The user is likely to have access to at least one image-generation API. Use this file as the source of truth for which tool the user currently prefers and how an agent should call it.

This file pairs with the OG / SEO / GEO skill because banner images, plugin tiles, blog hero images, and fundraise-deck illustrations are usually generated before they’re optimized. If you (the agent) need to make a new banner from scratch, start here.

When to use this file

The user says “make a banner / hero / tile / OG image”
A new plugin page or splash route needs an OG default
Fundraise decks need slide-specific imagery
Any time you would otherwise reach for stock photos or clip art

Swap-out block — current configuration

For users: edit the YAML inside this block when you switch tools, rotate keys, or change scope. The HTML comment markers are how agents locate this block reliably — leave them in.

For agents: treat this as the source of truth. Read the block, confirm the tool is reachable, then call it. If access_method: mcp, look for the named MCP tools via ToolSearch first.

preferred_tool: ideogram
access_method: mcp                      # mcp | sdk | rest | gui-only
mcp_package: "@takeshijuan/ideogram-mcp-server"
env_var: IDEOGRAM_API_KEY
scope: user                             # mcp scope flag passed to `claude mcp add`
relaunch_after_register: true
shared_with:
  - "content-farm/plugin-modules/image-gin"  # same API key
docs: https://developer.ideogram.ai/
strengths:
  - high-quality text rendering inside images (rare among image gens)
  - strong style consistency across multi-image sets
  - reasonable cost per generation
default_dimensions:
  og_banner: "1200x630"                 # see SKILL.md rule 6
  plugin_tile: "1024x1024"
  hero: "1600x900"
default_format_request: jpeg            # match SKILL.md rule 2 for OG; webp/png fine for body
notes: |
  - The full install gotchas live in the user's memory file
    `reference_ideogram_mcp.md` — read that BEFORE attempting to register
    or troubleshoot the MCP. Do not re-derive the install steps.
  - Same Ideogram API key as the image-gin Obsidian plugin.
fallbacks:                              # if Ideogram is unreachable / rate-limited
  - tool: dall-e-3
    access_method: sdk
    notes: via OpenAI SDK if user has OPENAI_API_KEY set
  - tool: flux-pro
    access_method: rest
    notes: via Replicate or fal.ai if user has those keys set

How agents should use this

Read the swap-out block above. That’s authoritative — never use a different tool just because it’s available.
If access_method: mcp, look for MCP tools matching the package name via ToolSearch (e.g. ToolSearch query: "ideogram"). If they’re not loaded, the MCP server may not be registered or the harness needs a relaunch — point the user at reference_ideogram_mcp.md rather than guessing the install command.
Generate at the dimensions in default_dimensions unless the user asks otherwise. For OG banners specifically, 1200x630 is non-negotiable (see SKILL.md rule 6).
Request default_format_request (or the closest the tool supports). For OG banners destined for iMessage / WhatsApp, JPEG. For body images, WebP / AVIF is fine.
Save the output to the project’s image-gin pipeline or a CDN (ImageKit, per SKILL.md rule 1). Never commit raw generations to /public as the long-term home — they belong on a CDN.
Caption it sensibly. The agent picking a generation prompt should also pick a sensible og:image:alt value. Keep them in lockstep.

Why Ideogram (current preference)

Stated by the user 2026-05-05. The standout reason: text rendering inside the image is reliably correct. Most image-gen tools (DALL-E 3, Stable Diffusion lineage, Midjourney) hallucinate or smear in-image text, which makes them unsuitable for banners that include the product name or a tagline. Ideogram lands the text. For a content-farm banner that says “Content-Farm — Obsidian community plugins”, that’s the difference between usable and unusable.

Secondary reasons:

Style consistency across a generated set (useful when banners across plugins should feel like a family).
API ergonomics are straightforward; the MCP wrapper is one of several solid options.
Per-generation cost is reasonable for the splash/plugin-page volume.

Other options the user may have access to

These are not currently preferred but are worth knowing as fallbacks (already in the swap-out block):

Tool	Access pattern	Best for
DALL-E 3	OpenAI SDK	Quick generations when Ideogram is rate-limited; weak text
Flux Pro / Schnell	Replicate / fal.ai REST	Photorealistic; good prompt adherence; weak text
Stable Diffusion XL	Replicate / local ComfyUI	Tunable, free if local; needs prompt eng
Midjourney	Discord-only (no API)	Manual creative; not agent-callable

When promoting a new tool to preferred_tool, edit the YAML block above. That’s the only place the change needs to happen — agents read from there.

dididecks-ai/agent-skills/open-graph-share-seo-geo/references/user-tools-for-image-generation