Jack-Tar — Deckhand
"A Jack-Tar never lets the deck go untended."
AI-powered skills and agents for building conference-quality presentation decks with Claude Code and Claude Desktop.
Vision
The existing pptx skill produces solid slide decks, but conference-quality presentations demand more: bespoke hero imagery, data-driven infographics, speaker-ready layouts, and a visual identity that holds up on a 40-foot projector screen in front of 2,000 people. This project closes that gap.
Jack-Tar is a coordinated suite of Claude skills and orchestration agents that combine image generation models, layout intelligence, and content-authoring tools into a single end-to-end pipeline. A speaker describes their talk, and the system delivers a polished, stage-ready .pptx — complete with generated visuals, consistent branding, typographic hierarchy, and speaker notes.
Problem Statement
Building a conference deck today — even with AI assistance — still involves a fragmented workflow:
- Draft an outline in one tool.
- Generate images in another (Midjourney, DALL·E, Flux, etc.).
- Manually place and resize assets in PowerPoint or Google Slides.
- Iterate on layout, colour, and typography by hand.
- Write speaker notes separately.
- Run a visual QA pass yourself.
Each handoff loses context. The image generator doesn't know your slide dimensions. The layout tool doesn't know your narrative arc. Nobody enforces brand consistency across 30 slides. Jack-Tar eliminates these seams by keeping the entire pipeline inside Claude's skill and agent framework, where every component shares context.
Architecture Overview
The system is organised into three layers:
┌─────────────────────────────────────────────────────────┐
│ Orchestration Layer │
│ deck-conductor agent (top-level) │
│ Receives talk brief → coordinates all skills → .pptx │
└────────────┬──────────────┬──────────────┬──────────────┘
│ │ │
┌───────▼──────┐ ┌────▼─────┐ ┌──────▼───────┐
│ Content │ │ Visual │ │ Assembly & │
│ Skills │ │ Skills │ │ QA Skills │
│ │ │ │ │ │
│ • narrative │ │ • imagegen│ │ • layout │
│ • speaker- │ │ • iconset │ │ • brand-qa │
│ notes │ │ • palette │ │ • slide-qa │
│ • outline │ │ • chart │ │ • pptx-build │
└──────────────┘ └───────────┘ └──────────────┘
Orchestration Layer — A top-level agent (deck-conductor) that accepts a talk brief (topic, audience, duration, tone) and breaks it into a sequenced plan. It calls the content, visual, and assembly skills in dependency order, passing shared context (palette, narrative arc, brand tokens) between them.
Content Skills — Responsible for the intellectual structure of the deck: outline generation, slide-by-slide narrative, headline copywriting, and speaker note drafting. These skills understand conference communication patterns (the "rule of three," progressive disclosure, audience callbacks).
Visual Skills — Handle all image and graphic asset creation: hero images via generation models, icon set curation, colour palette derivation, and data visualisation/chart generation. Every visual skill is resolution- and aspect-ratio-aware for standard slide dimensions (16:9 at 1920×1080 or 2560×1440).
Assembly & QA Skills — Take the content and visual outputs, compose them into .pptx using the existing pptxgenjs pipeline, enforce layout rules, and run automated visual QA (overlap detection, contrast checking, margin enforcement, text overflow).
Project Components
1. deck-conductor — Orchestration Agent
The central agent that owns the end-to-end workflow.
Responsibilities:
- Parse and validate the talk brief (topic, audience profile, session length, tone, branding requirements).
- Generate a deck plan: slide count, slide types (title, section divider, content, data, quote, closing/CTA), and narrative flow.
- Coordinate skill invocations in dependency order.
- Maintain a shared context object (
DeckContext) that carries palette, fonts, brand tokens, image manifest, and outline across all skill calls.
- Handle iteration: when QA finds issues, re-invoke the appropriate skill with corrective instructions.
Key design decisions to make:
- How
DeckContext is serialised and passed between skills (JSON blob vs. file on disk).
- Retry and fallback strategy when an image generation call fails or returns low-quality output.
- Whether the conductor runs as a single long session or checkpoints to disk for resumability.
2. narrative-architect — Content Outline Skill
Transforms a talk brief into a structured slide outline.
SteveGJones
