Blog Image Gen

Blog Image Generation

Generate editorial hero images for posts on bogdanbaciu.com via the blog repo's image-generation scripts. Images land at priv/static/images/posts/<slug>/hero.png and are auto-rendered on the post page by Bogdan.Content.Post.detect_hero/1 (no frontmatter changes needed).

When to use

• "Generate the image for <slug>."
• "Generate images for every post that's missing one."
• "I dumped some PNGs into art/ — wire them up."
• "Show me what's done and what's still pending."

Required setup (verify once per session)

1. Work from the blog repo root. On this machine the expected repo is: /Users/danb/Desktop/bogdanbaciu-dot-com.
2. Confirm the project scripts exist before invoking anything:

test -f scripts/gen_blog_image.py && test -f scripts/ingest_art.py

If either script is missing, stop and locate it with:

rg --files /Users/danb/Desktop | rg '(^|/)gen_blog_image\.py$|(^|/)ingest_art\.py$'

3. OPENAI_API_KEY is set in the shell that will run the script.
4. pip show openai returns a version (install with pip install --upgrade openai if not).
5. Verify current OpenAI image-generation model names, tool syntax, pricing, size limits, and quality options from official OpenAI docs before quoting or relying on them. Do not treat model names or per-image costs in this skill as evergreen.
6. Confirm estimated batch cost with the user before any --all run. ChatGPT subscriptions do not imply API credits.

Workflow

Step 1 — Check what's pending

python3 scripts/gen_blog_image.py --list

This prints one line per post showing whether it's already done (and where) or PENDING. Posts are considered done if EITHER of these exists:

• priv/static/images/posts/<slug>/hero.png (script output, served)
• art/<slug-no-hyphens>.png (manual download — needs ingest)

Step 2 — If the user has manual downloads in art/, ingest first

python3 scripts/ingest_art.py --dry-run    # preview
python3 scripts/ingest_art.py              # do it

This copies art/<compact>.png → priv/static/images/posts/<slug>/hero.png by matching the compact (no-hyphen) filename to a real post slug. Orphans (filenames that don't map to a slug) are reported but left alone.

Step 3 — Smoke-test on one post (dry run, then real)

python3 scripts/gen_blog_image.py <slug> --dry-run    # prints final prompt
python3 scripts/gen_blog_image.py <slug>              # one real call (~$0.30)

Open the generated priv/static/images/posts/<slug>/hero.png and confirm quality before any batch run.

Step 4 — Batch the rest

python3 scripts/gen_blog_image.py --all

Sequential, with SLEEP_BETWEEN_CALLS_SEC = 3 between calls. Already-done slugs are skipped automatically. Failures are logged but don't halt the loop — re-run to retry.

Step 5 — Verify rendering

Boot the server (mix phx.server) and visit /thoughts/<slug>. The hero should appear between the title and the lede. If it doesn't:

• Check Bogdan.Content.Post.detect_hero/1 — it reads via :code.priv_dir(:bogdan), which requires a recompile after adding files (touch lib/bogdan/content.ex or restart the server).
• Confirm the file is at the exact path priv/static/images/posts/<slug>/hero.png.

Step 6 — Prompt-quality and thumbnail review

Before accepting a generated image:

• The image matches the post's concrete subject, not just its mood.
• It is editorial and specific, not generic stock imagery.
• It has no fake screenshots, fake UI, fake charts, unreadable text, distorted hands/faces, or accidental brand/logo claims.
• It works as a crop at article-card size and as a full hero.
• It is visually distinct from nearby posts in the image inventory.
• The prompt mentions the one or two actual concepts the article is about and avoids stuffing in every paragraph of the post.

Output a compact review table:

If a thumbnail needs revision, keep the prior image until the replacement is generated and reviewed.

For batch review, build a local contact sheet from the blog repo root:

python3 /Users/danb/src/skills/blog-image-gen/scripts/thumbnail_contact_sheet.py --repo .

Open priv/static/images/posts/contact-sheet.html and scan for duplicate visual ideas, awkward crops, distorted text, and subject drift before accepting a batch.

Step 7 — Commit

git add priv/static/images/posts/
git commit -m "blog: add hero images for <N> posts"

The PNGs are committed to the repo (Fly.io builds the release from this tree; priv/static/ ships in the artifact).

Skill maintenance evals

When editing prompt parsing or review workflow text in this skill, run:

PYTHONDONTWRITEBYTECODE=1 python3 /Users/danb/src/skills/blog-image-gen/evals/prompt_fixture_eval.py

The fixture locks the contract that --dry-run output must expose the final prompt as a fenced FINAL PROMPT block before any real API call is made.

Prompt sourcing

scripts/gen_blog_image.py resolves the prompt for each slug in this order:

1. Pre-written: if blog-image-prompts.md has a section ## N. <slug> — ... with a fenced code block, that exact prompt is used.
2. Auto-generated: otherwise the script prepends the house-style preamble to a one-liner derived from the post title + description (or first ~300 words of body if no description). The auto path is fine for batch but pre-written prompts produce noticeably better results — encourage the user to draft one in blog-image-prompts.md for any post they care strongly about before running.

Tunable knobs (top of scripts/gen_blog_image.py)

Verify current allowed values in official OpenAI docs before changing these.

On "thinking" / reasoning mode

The script may use the Responses API with an image-generation tool and a separate planning model. Verify the current OpenAI API surface before editing this path; model names, tool payloads, and reasoning options are time-sensitive.

client.responses.create(
    model="<current-planning-model>",
    input=prompt,
    tools=[{"type": "image_generation", "model": "<current-image-model>",
            "size": "<verified-size>", "quality": "<verified-quality>",
            "output_format": "webp", "moderation": "auto"}],
    reasoning={"effort": "<verified-effort>"},
)

Image bytes come back as a base64 string on the image-generation output item. Do not downgrade to another API path without checking the project script and the current official docs.

What this skill does NOT do

• Does not edit blog post frontmatter — hero detection is automatic via filename.
• Does not push images to a CDN — they ship in the Phoenix release artifact.
• Does not retry failed API calls inline — re-run the command to retry skipped slugs.
• Does not write prompts for you when a pre-written one is desired; draft those in blog-image-prompts.md first.