carlosarias/seedproject-web
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions
seedproject-web/agents/prompts/image.md
Carlos Arias 2c969c0753 feat: content-pipeline/ → agents/ — formalize the agent system in the seed 
Adopt the agents/ architecture proven on medellin.co (reference impl):

- Move the content engine to a top-level agents/ dir: orchestrators, prompts,
  config, run.sh, admin console, shared libs. All content-pipeline literals
  repointed (config paths, scripts, admin, LLM-facing prompts/image.md string,
  configure.mjs, new-site.sh, astroagent tokenFile, .gitignore runtime block).
- Every script carries a parseable @agent-manifest header: name, title, class
  (content|operational|runtime|plumbing), trigger, model, prompts, skills (MCP),
  tools, reads/writes tables. 5 content agents + 3 plumbing scripts.
- New agents/catalog.mjs generates the catalog from the headers:
  agents/AGENTS.md (human, grouped by class) + agents/agents.json (machine
  manifest — a clone diffs it against a source to find missing tools/tables/MCP
  before running). configure.mjs regenerates the catalog on every identity
  stamp. No DB table, no watcher.
- config.json gains paths.stateDir/newsDir; publish-tick, write-daily, and
  news-radar read them instead of hardcoding.
- Full cut: content-pipeline/ deleted (the seed has no live crons, so no
  hybrid period needed). Docs updated (AGENTS.md structure + pipeline section,
  README paths).

Clones migrating from content-pipeline/: see medellin.co's
.memory/handoffs/agents-directory-migration.md for the cutover playbook
(one cron set active at a time; migrate drafts/state after repointing cron).

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01FMQeUnUrAeexcZ7P2Hxa6G
2026-07-11 17:09:15 -05:00

2.1 KiB
Raw Blame History

Cover image generation (Higgsfield MCP)

The cover is an illustrative editorial image — appetizing and on-topic, but never implying a real photograph of a specific named restaurant, dish-as-served, or identifiable person.

Flow (use the Higgsfield claude.ai MCP tools)

  1. Pick a model: call mcp__claude_ai_Higgsfield__models_explore with action:"recommend" and a goal like "editorial food/lifestyle photography cover image", to get a suitable model id and its valid aspect_ratios. If that fails, default to the preferredModel from config (soul_2).
  2. Generate: call mcp__claude_ai_Higgsfield__generate_image with params: { model, prompt, aspect_ratio, count: 1 }. Use a valid aspect ratio close to 3:2 landscape. This returns a job id.
  3. Wait for the result: call mcp__claude_ai_Higgsfield__job_status with { jobId, sync: true }; repeat (respecting poll_after_seconds) until the job is terminal. Read the resulting image URL from results.
  4. Download: use Bash curl -fsSL "<image_url>" -o "<draftDir>/cover.jpg" to save the image as cover.jpg in the draft directory. Verify the file exists and is non-empty.

If image generation fails after a reasonable retry, continue without it and record imageGenerated: false in sources.json.

Style guardrails (put these in the prompt)

  • Editorial food/lifestyle photography aesthetic, natural light, shallow depth of field.
  • Medellín / Colombian context where relevant (tropical, warm, Paisa setting) but generic.
  • No real logos, no readable signage, no recognizable real people, no text overlays.
  • High detail, web-quality, landscape orientation.

Prompt skeleton

Editorial food photography, {subject relevant to the article topic}, {Medellin/Colombian
ambience if relevant}, natural window light, shallow depth of field, warm tones, appetizing,
clean composition, no text, no logos, no people's faces. Photorealistic, high detail.

Fill {subject} from the article topic (e.g. "a vibrant brunch spread on a cafe table", "specialty coffee being poured", "a colorful arepa plate"). Keep it generic and illustrative.