Slides Grab Image-Native Skill
In Claude Code or Codex, generate whole-slide raster images one page at a time against reference template pages, for maximum fidelity to corporate templates, brand layouts, and visual density.
The skill is pure workflow instruction with no destructive commands; asset rules forbid remote http(s) image URLs, scope work to the slides-dir, and data flow (template PNG → reference → generated PNG/wrapper) is disclosed, with three provider options documented. Deductions: default provider reuses ~/.codex/auth. to send prompts/images to a (per README possibly unsupported) private backend without per-run user confirmation; reliance on the unaudited god-tibo-imagen package; no rollback/cleanup mechanism described.
SKILL.md is self-consistent, explicitly bans the batch command known to yield poor output, and gates on validate and the design gate; the repo contains codex-imagen, image-contract and design-gate tests covering image-generation contract and export gating, with tested error messages (codex login hints). Deductions: static review only; the real success path (OpenAI gpt-image-2 / private backend) cannot be verified statically and may fail on backend unavailability.
Scenario and boundaries are well declared: for 'visual fidelity to a template over editable text', with explicit trade-offs vs HTML mode and experimental PPTX/Figma exports. Deductions: core image generation depends entirely on overseas APIs (OpenAI/Codex private backend/Google) with no mainland-China reachability statement; README warns some paths rely on unsupported private backends; no Chinese-language support noted.
Documentation is well layered: SKILL.md focuses on one pipeline and references plan/design/export sibling skills; repo has MIT license, npm versioning, CI (pages/publish workflows), bilingual README and contribution flow. Deductions: publisher identity (vkehfdl1 / Jeffrey) unknown to the registry; the skill itself has no changelog or known-limitations list, and the pinned god-tibo-imagen dependency has unclear provenance.
The workflow (template pages rendered as references, per-slide generation, design-gate gating) logically targets the 'template fidelity' task and a live showcase exists. Deductions: static review cannot verify the actual usability and text-rendering correctness of generated PNGs (high spelling/layout risk for image models); README implies non-trivial review effort; comparative-benefit evidence is limited.
The repo provides a real test suite (codex-imagen.test.js verifying provider fallback and error messages, design-gate.test.js verifying export gating and template fidelity) plus npm publish/pages CI — auditable primary evidence. Deductions: static cap of 5 applies; no independent third-party execution evidence; image-generation tests use mock providers and do not cover real render quality.
- The default provider reuses local ~/.codex/auth. and sends prompts to a private backend; use --provider openai or nano-banana with your own API key if unacceptable.
- Image-generation quality (text spelling, layout fidelity) is independently unverified; manually review every PNG before use — the README itself admits some backends may be unsupported.
- Core function depends entirely on overseas image-generation APIs; may be unreachable from mainland-China networks without a proxy and offers no offline fallback.
- Dependencies include the pinned god-tibo-imagen package of unclear provenance and audit status; review before installing.
- PPTX/Figma exports are experimental and may require manual cleanup.
What does this skill do, and when should you use it?
slides-grab-image is one of 7 skills bundled in the slides-grab monorepo, located at skills/slides-grab-image/SKILL.md, and implements the image-native presentation pipeline. Instead of generating editable HTML slides, the generated PNG itself is the slide: each page is produced with slides-grab image, passing a matching template page via --reference so the model copies the template's layout, color bands, typography hierarchy, and density, swapping only text and photos. The workflow has three stages: a slides-grab-plan outline matched to the template's per-slide density, per-slide generation gated by a design review, and export via slides-grab-export to PDF or 2160p PNG. It is aimed at cases where visual fidelity matters more than downstream editability.
1) Plan stage: reads the user's reference template (PPTX/PDF/HTML), imports it with slides-grab import-template, and renders template pages to PNG (e.g. pdftoppm) under .slides-grab/template-previews/; it identifies each page's layout type and information density, writes a dense outline with one slide per template page tagged template-page: NN, and records mode: image-native and style: template-pack in slide-outline.md. 2) Design stage: for each slide it runs slides-grab image --image-native --name slide-XX with the matching template page and real photos as --reference inputs; the prompt describes only the content to inject while requiring exact copying of the reference's visual design. The command writes the PNG, a slide-XX.html wrapper, and regeneration metadata under .slides-grab/image-native/. It then runs slides-grab validate and the design gate (PNG evidence, Pass A + Pass B review), regenerating until all Critical findings are resolved and a Proceed receipt is recorded. 3) Export stage: with a fresh Proceed receipt, runs slides-grab pdf or slides-grab png --resolution 2160p (PPTX/Figma conversion is experimental/unstable). Revisions regenerate a specific slide with the same --reference and adjusted --prompt, or open the image-native editor with slides-grab edit-image.
- Marketing or brand teams who hold a corporate PPT template or a filled example deck and need new AI-generated slides to match its layout, colors, and typography hierarchy exactly
- Content creators embedding real product shots or member photos into a fixed brand layout and wanting the model to place them automatically
- High-information-density slide needs (many bullets, stats, table data), since a single raster has no HTML semantic structure limits
- Image-led decks or rapid visual exploration where composition matters more than later text editing
- Team workflows that require per-slide design-gate review (PNG evidence + Pass A/B) before PDF/PNG export
What are this skill's strengths and limitations?
- Reference template pages define the visual design, giving fidelity to existing corporate templates and brand layouts that HTML pipelines struggle to match
- A single raster carries more text and data per slide than semantic HTML, with no structural limits
- Default codex-imagen provider reuses the local Codex login — no separate API key required
- --image-native writes the PNG, HTML wrapper, and regeneration metadata together, staying in contract with slides-grab edit-image for revisions
- Enforced validate + design gate (A/B review, Proceed receipt) before export adds a quality process
- Image-native slides are far less editable than HTML: no direct semantic text edits; revisions require slides-grab edit-image or full slide regeneration
- The PNG is the slide — text is not searchable, selectable, or accessible; the README explicitly recommends HTML mode when editability, accessibility, or searchability matters
- codex-imagen calls an unsupported private Codex backend that may break without notice and requires a ChatGPT account entitled to image generation
- Batch generation (generate-images) is forbidden; one-at-a-time generation is slow for large decks
- PPTX and Figma exports are officially experimental/unstable and need manual cleanup
- Export is blocked by the design-gate receipt — skipping review means no pdf/convert/figma
How do you install this skill?
Install the whole slides-grab collection (this skill ships with it): npm install slides-grab && npx playwright install chromium && npx slides-grab install-skills --target all --scope user. Requires Node.js >= 20. The default image provider, codex-imagen, reuses your local Codex login (run codex login once; no separate API key). Optional providers: --provider openai (OPENAI_API_KEY) or --provider nano-banana (GOOGLE_API_KEY/GEMINI_API_KEY). The README does not document installing skills/slides-grab-image in isolation, excluding the sibling skills.
How do you use this skill?
After installing via docs/installation/claude.md or docs/installation/codex.md in Claude Code or Codex, trigger the skill with a prompt that includes the reference template, e.g.: "Make a 10-page quarterly product review in the exact visual style of references/acme-qbr.pptx." The skill renders template pages to PNG and writes an outline (mode: image-native); after you approve it, each slide is generated like: slides-grab image --image-native --name slide-01 --prompt "Follow the reference image layout exactly... Replace the text with: Title..." --slides-dir decks/my-deck --reference <template-page.png> --reference <real-photo.png>. Then run slides-grab validate, clear the design gate, and export with slides-grab pdf or slides-grab png --resolution 2160p. Slides are fixed at 720pt × 405pt (16:9).
How does this skill compare with similar options?
The source positions slides-grab overall as a "Claude Design Open Source Alternative." Within the repo, this skill's counterpart is the semantic HTML slide pipeline (the HTML mode of sibling skills slides-grab-plan/design/export): HTML mode is editable, searchable, and easy to convert, while image-native prioritizes visual composition over editability. No other specific competitors are named in the source.