Communitygithub.com

create-reels

Turn a bare topic into a finished narrated vertical explainer, or audit and re-edit an existing reel for hook clarity, visual pacing, speech gaps, phone legibility and contrast. Use for topic-only requests in a reel-making context and requests for a Reel, Short or TikTok explainer.

What is create-reels?

create-reels is a Claude Code agent skill that turn a bare topic into a finished narrated vertical explainer, or audit and re-edit an existing reel for hook clarity, visual pacing, speech gaps, phone legibility and contrast. Use for topic-only requests in a reel-making context and requests for a Reel, Short or TikTok explainer.

Works with~Claude Code~Codex CLI~Cursor
npx skills add https://github.com/dakiemdarktharr/explainer_reels_skill/tree/main/create-reels

Ask in your favorite AI

Open a new chat with this agent skill pre-loaded.

Documentation

create_reels

Own the research, script, storyboard and production when the user gives only a topic. Default to English, 1080 × 1920 and 30 fps. Aim for about 30 seconds for a simple idea. Use up to 45 seconds when the mechanism and its payoff need room. An explicit runtime takes precedence. An ordinary factual question outside a reel-making context can still receive a text answer.

Story before styling

Find the one question the viewer can understand on frame zero and the one sentence the final image should answer. Verify factual claims against primary sources. For a disputed concept, name the uncertainty. Sketch two materially different opening pictures and pick the one whose visual consequence can be demonstrated in the time available.

Write a compact timecoded storyboard with the actual hook, spoken lines, muted-screen copy, object/action, and sound cues. Budget the reading time of essential text after its entrance. Carry one concrete object or problem through a change of state. Let the outcome of that change earn the next shot. Give each shot one focal event. For this user, put only the main thought on screen. Allow picture-only sequences. Avoid lower-thirds, subheads, footers, progress counters and decorative labels. Express transitions with the object, framing, light or motion. Protect natural time to think.

Build an architecture with the viewer

For a learning architecture or mechanism, open with a concrete job for the viewer to solve. Let them try a simple first design. Put it under a test where its limitation is visible, then introduce the next component as the fix. Give each further addition a specific reason. Use direct, second-person language and vary the transitions so the script sounds spoken. Reveal the architecture’s name when the viewer has seen what earns it. Return to the original example, commit to a prediction, and play or show the outcome.

When possible, use a pair of inputs that the simple design cannot distinguish. Keep both examples consistent through the payoff. Verify that each claimed limitation follows from the pictured design and that the demonstrated fix can change the result. Present this as a design exercise unless the history of invention is documented.

For a prediction demo, decide what the model can observe and what remains hidden. Compute the forecast from those inputs when a small working example is practical. Reveal the observed outcome independently, from the source sequence or measurement. Keep example identities consistent across cuts. Identify illustrative quantities on screen or in the narration when they appear.

Picture and voice

Default to a dark field with warm-white text and one vivid action color. A second accent can mark an obstruction or change. Test contrast on the composited background, including accent text. Encode meaning with geometry, symbols or direct labels as well as color. Keep essential copy around x 65–900 and y 120–1510 in the default frame. Give the subject and its moving parts clearance throughout the camera move. Project their bounds at each frame when geometry is available. At 360 × 640, mask the rightmost 54 px and the bottom 127 px. Watch muted for labels covering clues, missing fingertips or prop edges, and answers visible before the reveal.

For a physical action, storyboard the starting state, contact, and lasting result. A key must enter the keyway and seat before it turns and releases the shackle. A page turn must uncover new artwork that is still there when the camera returns. Inspect decoded frames immediately before contact, during the action, after release, and on the next appearance. Compare the visible geometry and timing with the spoken claim.

When a physical subject benefits from detail, find a reputable free, human-authored mesh and verify its license. Embed the geometry and permitted texture maps in the HTML. Record the artist, source, license and modifications in the credits. Use first-party visual references to decide silhouette, scale, finish and light placement.

Build an offline, self-contained HTML canvas composition with a pure absolute-time draw function. Use code-drawn art, embedded fonts, and deterministic geometry. A HyperFrames timeline can drive the canvas when available. Preview controls belong outside the exported frame. Keep the live playback clock independent of frame capture.

Estimate a voice budget from word count, then measure the actual recording. A conversational 135–160 words/min including pauses is a useful starting range for English. Look at each dense phrase separately. Cut copy before raising the speed of a difficult sentence. Use word or phrase alignment to place the picture. Trim outer padding with short fades, preserve consonants, and reserve deliberate pauses for prediction or recognition.

Write initialisms normally on screen. Expand unfamiliar ones and direct the voice to speak the individual letter names where appropriate. Check the final encoded line by ear when playback is available. An independent transcription can flag “a GI” for “AGI.” Retime or regenerate the affected phrase and verify the new mix. Describe transcription-only checks accurately.

Use the requested voice service when available. For ElevenLabs, use the authenticated connector or ask for a process-scoped key through a secure entry method when needed. Keep the key out of chat, files and logs. Release temporary credential material once synthesis completes. After a provider timeout, check for an existing take before opening another paid generation. A local voice is a workable fallback when the brief permits it. Identify the actual source of the delivered narration.

Examine the isolated voice and the finished mix separately. A music bed can hide a long speech gap from a silence detector. Fix the source of unwanted hiss, clicks or bed-only passages. For a clean synthetic stem, short boundary fades and fewer sound layers often suffice. Cue Foley to an identifiable action and check loudness and true peak after AAC encoding.

Reference study and lighting

For a research-led revision, choose films that solve the actual problem. Include native vertical work and identify the aspect ratio of any craft reference. Honor a requested number of examples. Link to the creators. Review openings, the change that explains the mechanism, and the held or resolved image. Decode at the source cadence for a frame-by-frame request. Inspect adjacent frames in selected passages and record exact source times and frame numbers. State when a review used a studio excerpt. Distinguish observed visual choices from hypotheses about attention.

Test the proposed change on identical camera frames. For a miniature, start with a warm directional key, restrained fill and a cooler edge. Study the grounding at contact points, the penumbra on the receiver, moving highlights along bevels, and texture scale. Distinct roughness, a photographed studio environment and spatial ambient occlusion can clarify materials. Use depth of field only where it keeps the informative edge readable. Inspect at 1:1 for inverted normals, jagged reflections, alpha halos, transparent objects in depth passes and disconnected parts. Adjust the rig to the shot.

Render and review

For a new installation, follow Set up rendering and frame review to install the bundled renderer dependencies and Chromium.

For the included renderer, expose window.compositionReady and window.renderFrame(seconds) returning the drawn canvas. The HTML must run from disk and accept ?capture to disable live playback. Use Node with Playwright Core and Chromium, plus FFmpeg/ffprobe on PATH. Optional environment variables REELS_PLAYWRIGHT and REELS_CHROME can point to installed copies.

node scripts/render.cjs index.html out.mp4 mix.wav 45 30
python scripts/audit.py out.mp4 audit --speech voice-aligned.wav --strips --review-at 0 16.9 18.3 20.2

The audit requires NumPy. It decodes every frame and aligns audio windows on the fixed-rate timeline, producing a CSV and a summary of low-motion intervals, blank candidates and quiet passages. Each optional contact sheet contains 150 frames in reading order. A null audio-gap result means the corresponding stream was unavailable. Supply the edited, time-aligned voice stem for speech-gap results. Frame-change and level thresholds locate places to inspect. Visual intent and listening determine which ones merit changes.

Substitute the actual duration and the action times from the storyboard. For each requested time, --review-at saves the containing frame at native resolution and as a masked 360 × 640 PNG. The summary lists frame numbers and file paths. Open the images and inspect the action. For a sound-dependent reveal, check the corresponding windows of the finished AAC, including any intended rest.

Inspect decoded pictures from each beat, the opening second, and frames on both sides of cuts at full and phone size. For a frame-by-frame request, examine thumbnail strips covering each frame alongside the metrics and expand suspect intervals. Check offline playback, out-of-order frame equality, final stream metadata and the audio tail. Apply corrections and re-export. Deliver the MP4, source HTML, storyboard and concise verification note, with any untested playback conditions. Keep temporary renders and credentials outside the deliverables.

Related Skills