RedCube AI

When a task moves from "make a few images" to "produce a visual deliverable I can actually use," the hard part is usually the full workflow, not one page:

Source material, notes, screenshots, reference images, and old drafts are scattered. How do they become one coherent deliverable?
After many generated versions, which review comments were addressed, and which version should be rerun?
Slides, Xiaohongshu notes, and posters need different routes. Can the system choose the right creation path for the deliverable type?
During longer generation, review, and export runs, can the user still understand what is happening?
At handoff, can exported files, review records, and source material still match each other?

RedCube AI is built around those questions. For knowledge-heavy visual work, it keeps source organization, page generation, review, revision, progress reporting, and export evidence on the same delivery line so a draft can move toward deliverable files.

It does not reduce visual delivery to "generate one image." A usable deliverable often needs several visual directions, layout comparison, missing-material handling, review and repair, and final export checks. RedCube AI keeps those creative decisions and delivery evidence on one line so every revision can explain why it changed and where it landed.

Core Highlights

Continuous Creation Around The Deliverable It does not stop at a single generated image. It keeps working around a concrete slide deck, note series, or poster, organizing material, generating pages, absorbing review feedback, and preparing final exports.

Source To Deliverable In One Workspace Lecture notes, project summaries, references, screenshots, old drafts, and review comments stay on one delivery line for inspection and reuse; runtime artifacts live in the task workspace, not in the source checkout.

Traceable Review And Revision Each review round, rerun, revision target, and export result stays connected, so operators can see why the current version changed.

Routes Match The Output Type Slides, Xiaohongshu notes, and knowledge posters use different default routes; editable PPTX and HTML routes are explicit selectable routes.

Template Intelligence Before Rendering When a template or reference deck matters, RedCube profiles layout zones and placeholder capacity before visual direction or native PPTX authoring consumes those constraints.

Progress Stays Visible During Long Jobs During generation, checking, reruns, and export, RCA progress and review surfaces expose the current step, remaining issues, and the next review focus.

Room For Visual Exploration And Comparison Formal visual delivery often needs multiple directions, repeated-failure diagnosis, variants, and export checks. RedCube AI does not lock creation into one path; candidates, review, repair, and handoff can continue together.

One-Sentence Quick Start

You can start with prompts like:

"Turn these lecture notes and references into a polished teaching deck, keep the progress visible, and export the final PPTX/PDF. If I ask for editable slides, use the native PPTX route."
"Use this source package to draft a Xiaohongshu note series, tell me what is still missing, and keep each review round traceable."
"Make a poster from this project summary, track the review comments, and export the final delivery files when the content is ready."

What It Helps With

Turning notes, outlines, references, screenshots, and draft fragments into formal slide decks, note series, and poster-style deliverables.
Keeping multi-round review, reruns, and export checks tied to the same workspace.
Showing human-readable progress while longer-running jobs continue in the background.
Delivering exported files that stay connected to their source material and review history; editable PPTX is an explicit route when requested.
Comparing visual directions, finding repeated failure patterns, generating variants, absorbing review, and completing export checks inside the same delivery stage.

Current Delivery Focus

Slides for teaching decks, academic talks, internal briefings, and formal reports. The current default PPT route is image-first full-slide authoring; HTML and editable native PPTX are explicit selectable routes.
Xiaohongshu notes for knowledge posts, science communication, and serialized publishing. The default route is GPT-Image-2 full-page 3:4 PNG authoring through author_image_pages; HTML remains an explicit maintenance route.
Knowledge posters for single-page visual delivery.
Academic paper and conference poster lanes continue to be evaluated case by case.

How It Works

Experts provide the source material, audience expectations, and final judgment.
The AI operator handles direction exploration, generation, revision, reruns, export, and progress reporting.
The workspace keeps sessions, review state, rerun history, artifact refs, and export outputs together for inspection.

Personal and team knowledge stays in its existing home. RCA can use authorized Obsidian notes alongside project Markdown, or use the working directory alone when there is no personal library. Reusable product methods, owner preferences, topic experience and project decisions have distinct scopes; each task consumes a versioned snapshot, and accepted updates return to the original source. See the visual memory policy.

Current Boundary

RedCube AI is an independent visual-deliverable Foundry Agent. Its first public identity is visual delivery: source intake, staged visual authorship, review, repair, export, and file handoff.
In the OPL family, RCA is OPL Package(kind=agent): RCA owns its executor-neutral identity, capabilities, work-item and typed-view semantics, and all visual-deliverable authority.
The first public surface is the single redcube-ai app skill. It is the current Codex carrier projection, not the Package identity or the complete installed truth; Codex, OPL, and other general agents can reach the same stable RCA capabilities through supported routes.
It can be used as the Presentation Foundry inside One Person Lab, and it can also be called directly by Codex or another agent through stable capability entries.
RedCube owns material intake, visual generation, review loops, export, and file handoff.
Content framing, audience fit, and final acceptance stay with experts.
External publishing and upload steps stay under human supervision.

Installation

Install through the standard OPL Package entry:

opl packages install rca --json
opl packages status --package-id rca --json

The publication channel is ghcr.io/gaofeng21cn/one-person-lab-packages/rca. Immutable versions identify exact releases; latest-stable selects the current version. OPL and the native plugin manager handle installation and updates. Separate GitHub Release pages and attachments are not used for distribution.

Start a new task after installation to load the professional skills. Package installation, runtime callability, and domain acceptance remain separate; see Current Status.

Usage And Documentation

Installing the Codex entry skill still requires the complete RCA Package and callable OPL-hosted runtime to execute a deliverable. Installation, action selection, and human review steps have one owner: the Usage Guide.

This repository supplies the declarative visual pack, professional methods, machine contracts, and Python native helpers. Its current evidence contract still records a live StageRun acceptance gap and does not claim domain or production readiness. See Current Status for source facts and evidence limits.

Docs Guide: navigation by reader need.
Architecture: Package, carrier, executor, and domain authority ownership.
Contracts: executable machine inputs.
Document Lifecycle: maintenance and retirement rules.

Start developer verification with scripts/verify.sh; lanes are defined in scripts/test-registry.ts. Retired implementations and plans remain traceable in Git history and are not execution entrypoints.