PaperBanana Skills

Academic figures and editable slide decks, from your agent.

English · 中文

Quick start · Examples · Documentation

Turn a methods description into a figure, plan a consistent slide deck, or build PowerPoint objects you can edit. Use the skills in Claude Code or Codex; choose an image backend when your task needs one.

Archived figure; the previous README records GPT Image 2. Click for the full-resolution image. More outputs and provenance.

Quick start

1. Install the skills

Codex — use the skills installer:

npx skills add PlutoLei/paperbanana-skill -a codex

Claude Code — use the plugin marketplace:

claude plugin marketplace add PlutoLei/paperbanana-skill
claude plugin install paperbanana@paperbanana-skills

For full deck orchestration, also install paperbanana-slide-deck@paperbanana-skills with --scope project. Installation details and manual setup →

2. Choose your output

I want to…	Start here
Generate a figure or an image-based slide deck	Set up the Python backend with your chosen image/VLM providers
Build editable PowerPoint slides	Run the local PPTX example; rendering needs Bun, with no image API call
Keep using Gemini / Vertex	Keep the existing backend and authentication setup; see provider roles

3. Describe the task

Ask your agent:

Use paperbanana with Gemini to draw a four-layer CNN with batch normalization. Label the input, convolution blocks and classifier; return the image and its review status.

In Claude Code you can also invoke /paperbanana. Name your preferred provider when you have one. Starter prompts →

Preview feature: Codex native image routing and explicit Image 2.5 controls are on codex/image25-adapter, ahead of the current v4.5 release. Setup and live smoke results describe that branch separately.

Selected examples

Scientific figure Gemini · Starter prompt	Research presentation Image slide · Starter prompt
Consistent slide deck View the set · Starter prompt	Editable PowerPoint Download PPTX · Source spec

The first three are archived outputs; the PPTX is a reproducible local rendering example with illustrative data. Starter prompts are new recipes, not recovered historical inputs. Full gallery and provenance →

How it works

• paperbanana guides figure, plot and slide requests through the configured Python pipeline: retrieval, planning, styling, generation and Critic review.
• paperbanana-slide-deck plans a deck and coordinates its content and visual style. Choose image slides or native editable PPTX objects explicitly.
• Python backend supplies model integrations and the generation/review pipeline. The local editable renderer lives in this repository.

The host, image provider and review provider have separate roles. A VLM integration does not necessarily generate images. Providers, runtime support and routing →

Validation and limits

CI checks the skills, manifests and PPTX renderer. The editable example includes a source spec and real native objects. Review generated labels, arrows, numbers and scientific claims before publication; a Critic score is an automated opinion, not scientific acceptance.

Image outputs can contain text or diagram errors. If review fails, describe the result as unreviewed. Historical comparisons and recent smoke tests do not establish a universal model ranking, fixed generation time or cost. Evidence and review guide →

Documentation

Guide	Use it for
Getting started	Installation, Python setup and first output
Gallery · Starter prompts	Outputs, full-resolution files and task recipes
Providers	Claude Code / Codex, Gemini / Vertex, OpenAI and preview features
Styles	Bundled styles, optional libraries and backend presets
Editable PPTX · Spec reference	Build and edit native PowerPoint objects
Commands and troubleshooting	Common commands, recovery and review status
Evaluation	Evidence, quality checks and limitations
Changelog · Releases	Version history

Current release components: marketplace 4.5.0, paperbanana 4.4.0, paperbanana-slide-deck 1.3.0. These are separate component versions.

Contributing and attribution

See the contributing guide and code of conduct. Describe the runtime, provider/model and reproducible steps when reporting an issue.

Built on the PaperBanana project, with additional integrations in the maintained Python fork. See style sources for the preset libraries.

MIT · Maintained by Lei Yuxuan.