New
ralphy unit — curated deliverables. Projects now have an append-only units/ layer: ralphy unit create <project> --slug <s> --format <f> --from "<glob>" copies selected assets into a versioned deliverable with a unit.json provenance manifest (template + style + recipes + assets). See ralphy unit.ralphy compose — timeline-aware composition. Build the HyperFrames timeline from the asset manifest with explicit scene ordering and timing. See ralphy compose.ralphy image — local image ops. chromakey, cutout, fit, and svg-rasterize for sticker cutouts, aspect fitting, and overlay prep — no model call needed. See ralphy image.ralphy guideline — prompt-library rules. list and show <slug> surface the field-tested prompt guidelines (photoreal humans, anti-AI-slop, broadcast realism) so they can be folded into any generation prompt. See ralphy guideline.ralphy research run — deep-research engine. Planner + fan-out retrievers + a deterministic 5-level citation verifier, including native video analysis of TikTok/Shorts via gemini-3.1-pro-preview. See eval & research.ralphy hyperframes namespace. HF utility verbs with gen-log wiring, including save-version snapshots of index.html before structural edits. See ralphy hyperframes.More new verbs. ref locate (vision-based bounding boxes), ref pull-site (Playwright brand-DNA crawl), ref pull-bulk + assets unpack (bulk pulls), template extract (promote a project to a template), video frame / extend (i2v anchoring), generate batch / variants / image-batch with cost --dry-run, voice clone, editor preflight / trim-analyze, project assets / verify.Updates
Provider reliability. Transient errors retry with exponential backoff, per-endpoint concurrency self-throttles, oversized PNG anchors auto-resize for i2v, and ElevenLabs ToS rejections return structured errors with auto-resubmit. Voice/scribe/vision/size preflights run before paid calls.Render flags.ralphy render gains --grade / --quality, plus a --from-clip loudnorm mode for single-clip finishing. New audio/video recipes: grade, VHS, mix-music, compress. See ralphy render.Templates: format → styles taxonomy. Every template now declares a media format (video, image, carousel, fb-creative, motion-design, poster, sticker-pack); ralphy template suggest takes --format. See Picking a template.Fixes
ralphy new unified with project create. No more orphan project dirs invisible to generate / render (#031).Gen-log canonical schema. Every generation logs a consistent record with a slot field; analyze-video calls are logged too (#032).Captions per-slot output. ralphy generate captions writes per-slot files, returns a silent empty array for no-speech clips, and wraps to the caption safe zone (#010).May 29, 2026
New
Ralphy Desktop (preview). A new desktop app that drives a Ralphy project through an embedded Claude Code chat — onboarding, live project panel, and a confirm-with-cost permission modal. Runs against your localclaude install, so it uses your existing subscription credit instead of a separate API key. Ships as an early spike; expect rough edges.--provider flag and ralphy provider list. Generation commands (image, video, voiceover, music, sfx) now accept --provider <id> to pin a specific backend, and ralphy provider list prints the capability matrix for every connector you have keys for. See ralphy provider and ralphy generate.Skills marketplace. The public skills page is now a real marketplace — grid, filters, search, per-skill icons, and a detail modal with the full file tree for each skill.New skills.ugc-toon-action— subject-agnostic Spider-Verse/Arcane action-clip skill with style locks and multi-clip continuity.poster— fixed poster-as-landing layout (banner → wordmark → hero → sticker collage → copy → footer) with parameterized slots.- Four niche UGC skills:
unboxing,ad,model-swap, androckstar.
ralphy skill list to install.analog-horror-pick-a-door template. A full reproduction kit for the “pick a door” brand-comparison format — composition skeleton, monster cutouts, VHS/SFX bed, beep system, font, and a prompt cookbook. See ralphy template.y2k-streetwear-poster guideline. A new image-prompt guideline with three validated square examples, surfaced in the library and referenceable as @guideline:y2k-streetwear-poster.Updates
Templates are remix-only. Generic “make a video about X” prompts now route to a generalized skill, not a one-off template. Templates are only applied when you explicitly opt in (e.g.@template:<slug>, swap X for Y). See Picking a template.Fixes
gpt-5.4-image-2 honors aspect ratios. The --size flag (9:16, 1:1, etc.) is now passed through correctly for image generation across all flows; previously the model ignored the in-prompt hint and returned 1024² regardless. See ralphy generate.May 26, 2026
New
HyperFrames is the default render engine. A new HTML + GSAP renderer ships as the default forralphy render. It’s faster to iterate on, supports per-render --fps, --quality, --format, and --resolution flags. See ralphy render and the rendering guide.Skills gallery. Browse the full catalog of Ralphy skills on the new skills page — filterable by namespace, with motion-driven previews.Reference-consent gate. Ralphy now classifies prompts that imply a real person or branded look and asks for a reference image before generating. Pass --no-ref-consent to opt out per command. See brand, persona, refs.Updates
Default image model is nowgemini-3-pro-image-preview. New projects pick up the upgraded default automatically. Existing projects keep whatever they’ve pinned. See Models.Skill namespaces rebranded to ralphy: and ralphy-dev:. Skill IDs now follow a clean kebab namespace. Reinstall via ralphy skill install to pick up the new names. See Skill format.Render flag surface. ralphy render gained --fps, --quality, --format, and --resolution. Run with --dry-run to preview the resolved plan without burning a render. See ralphy render.Fixes
Pretty CLI output no longer prints[object Object] for arrays of objects in generate, render, doctor, and the editor flows. See Output modes.ralphy skill now errors cleanly when run from a directory with no source bundle, instead of failing mid-resolution.