Skip to content

tiziano-ai/walkthrough

v0.4.1MIT

Create narrated walkthrough films from real web-product use, then revise retained footage. For Apple Silicon Macs; requires the separate Walkthrough CLI, Node 24, FFmpeg and local Voices. Setup explicitly acquires runtime dependencies and a browser; plugin installation starts no recording or inference. Removing the plugin preserves the CLI and projects.

Walkthrough

Introduce a real web product through a film people can follow.

Give your agent the product, audience and intent. It explores the application, chooses the interactions that make it understandable, records real use and delivers a narrated local film. Ask for a revision in ordinary language; the agent can change pacing, framing, narration and captions from retained footage without repeating application work.

Make a walkthrough of this web product for a colleague who has never used it. Show what it enables and how its main capabilities connect. Use fictional content and keep the film local.

The product is the subject. A concrete task connects the journey; it does not turn the walkthrough into a report about the example's outcome. A new product changes the film project, not Walkthrough's software: no adapter, selector package or site-specific program is required.

What you install

Walkthrough has two cooperating parts:

  • The native plugin gives Codex or Claude Code the authoring skill and its references. Its root contains no Node package or installation hook: installing the guidance does not acquire dependencies or start work.
  • The independent CLI owns the browser, recordings, audio and local rendering. Film projects live outside both the installed release and the source checkout.

The initial platform is macOS Apple Silicon. You need Node 24.22.1 or later on the 24 line with npm, FFmpeg/ffprobe, and a local Voices installation with the three exact model revisions reported by help audio and a saved narrator. The installer explicitly acquires the exact npm dependencies and Playwright browser; production performs no dependency or model downloads. The C toolchain is needed to build a release, not to use one.

The renderer is Remotion. Its applicable license is an operator setup decision; it is not supplied by installing Walkthrough. See licenses.

Install the plugin

Add the tiziano-ai marketplace, then install Walkthrough through your host's native plugin route. Use the CLI belonging to the host you actually use; a desktop app's bundled runtime can differ from codex on your PATH.

Codex

codex plugin marketplace add Tiziano-AI/plugins
codex plugin add walkthrough@tiziano-ai

Claude Code

claude plugin marketplace add Tiziano-AI/plugins
claude plugin install walkthrough@tiziano-ai --scope user

Reload the host where supported, or start a fresh local session. Check that its catalog contains one Walkthrough skill. Do not copy the skill into another global skills directory: the plugin owns its discovery and version.

Install the CLI and complete setup

Ask your agent to help with setup. It can inspect actual prerequisites and prepare the configuration; you retain choices about your narrator, license eligibility and any installation you have not authorized.

  1. Download walkthrough-0.4.1-darwin-arm64.tar.gz and its SHA-256 file from releases. Verify the checksum, then extract the archive.

  2. Run the installer from that extracted release:

    node scripts/install-local.mjs install \
      --release /absolute/walkthrough-0.4.1-darwin-arm64 \
      --node /absolute/stable/path/to/node
    

    Installation explicitly fetches lockfile-pinned production dependencies with lifecycle scripts disabled, applies checked patches and acquires the selected Playwright browser into Walkthrough-owned storage. It verifies prerequisites and prepared bytes before activating the release. It does not install Voices models or modify other tools.

  3. Ensure the installer's reported launcher directory is on your PATH, then read the installed identity and setup findings:

    walkthrough --version --json
    walkthrough setup inspect --json
    walkthrough help setup --json
    
  4. Use setup prepare to create a candidate from the actual tools, installed Voices, selected narrator and your Remotion assertion. Review what it selects. setup apply rechecks the candidate and selects it; creating a candidate alone changes no active configuration. setup inspect and doctor then distinguish absent prerequisites, missing configuration and changed identities.

Select Node explicitly with --node; choose the invocation path maintained by your tool installer, not a versioned directory it may remove during an update. Walkthrough retains that selection and verifies the executable it reaches. npm, FFmpeg and ffprobe can also be selected through installer flags; paths are not tied to a particular package manager. Shipped help owns the current configuration schema and command grammar. Setup does not acquire models, silently accept a changed Voices runtime or invent a license assertion. After replacing a prerequisite, inspect setup again. A changed executable hash is a changed identity even when its reported version is unchanged; review a fresh prepare/apply candidate rather than treating the previous selection as verification of new bytes.

Prepare local audio prerequisites

Read walkthrough help audio --json: audioRequirements.models supplies the exact TTS, aligner and ASR repository revisions and their Hugging Face commands. Run voices --json doctor against the Voices installation selected for setup; its model_cache is the shared cache directory to substitute in those commands. Voices has no provision or download command. If prerequisites are missing, use the upstream Hugging Face CLI with explicit acquisition authority:

hf download <repository> --revision <full-revision> --cache-dir "<model_cache>" --dry-run
hf download <repository> --revision <full-revision> --cache-dir "<model_cache>"

Use each exact repository/revision from help and acquire the whole snapshot, without filenames, include/exclude filters or --local-dir. The first command previews acquisition; it can contact the Hub and write cache metadata. The second acquires missing bytes. Hugging Face owns both snapshot storage and the trees/<revision>.json resolver metadata required for offline resolution. Copied weights alone are insufficient; do not fabricate the metadata or create a private Walkthrough model cache. After acquisition, choose an existing saved Voices narrator and use setup prepare/apply. Walkthrough's help, doctor and setup perform no model downloads.

Make a film

Tell the agent the URL, intended audience and what they should understand. Add any meaningful data/access limits or voice and language preferences. The agent handles the technical brief, exploration, capture and edit. It should not ask you to write selectors, narrate every control or prepare shot JSON.

The authoring cycle is:

Explore → choose defining capabilities → record real use → connect gesture and response → narrate observed behavior → compose → review.

Walkthrough supplies a generic observe → act → observe browser. Element references and screenshot-grounded coordinates support DOM, SVG and canvas interaction. Click, drag, wheel, keys, text, upload and downloads retain actual input and original file bytes. Authentication uses an off-camera handoff; no credentials belong in action inputs or footage.

Recording and editing are separate. The project keeps stopped takes, source clocks, marks, input events and selected artifacts in sealed source sets. The edit chooses live ranges at 1×, explicit settled reading holds, cuts, camera moves, brief emphasis and independently placed speech. Pictures set the film's length; narration does not add automatic padding or accelerate the app.

The operator's saved narrator is the default. New edits select English or Italian explicitly. Speech, alignment and independent ASR retain separate identities. ASR differences help locate words to check; agreement does not establish natural or accurate speech.

See the browser and film authoring guide for input shapes and exact coordinate systems. The installed walkthrough help --json exposes the current operations and schemas without reading implementation code.

Watch, revise and keep the result

The primary result is walkthrough.mp4, at 1920×1080/60 fps. A local review page adds native playback, fullscreen, a seekable transcript, downloads and a copyable feedback position. The bundle also contains portable VTT, canonical text, poster, edit, checksums and any actual exported files selected for delivery.

Ask for a focused revision:

Keep the opening wider. Let the explanation begin before the drag, and shorten the pause after the result. Don't reopen the application.

The author inspects retained sources, previews the connected passage and recomposes offline. --reuse-audio verifies compatible retained speech rather than synthesizing again for a crop or caption change. Matching, verifiable audio keeps its original provenance even when its former Voices runtime is unavailable. New speech or new analysis requires its applicable setup.

Technical checks can reject broken timestamps, unreadable layouts and audio defects. Watch and listen to the complete film at normal speed, then check whether someone unfamiliar with the product understands its purpose and main interactions. A successful render, ASR agreement or awaiting-human-review label does not establish that outcome.

Recovery and storage

Every action is retained before dispatch. Reusing its identity retrieves that attempt; it does not click again or promise server idempotency. After a lost response, inspect the original action and reconcile its observed effect before dependent work. Inspect and close the original session launch; inspect and resolve an incomplete audio attempt through its retained identity. An error identifies surviving state and the useful next operation; clearing caches or starting a replacement session is not routine recovery.

Keep explicit .walkthrough projects outside the release. They contain the sources and audio needed for revisions, not your authentication profile. Successful composition removes its known regenerable render scratch after verifying delivery. Preview replaces one successful current excerpt instead of accumulating versions. project compact handles exact older or settled failed compositions and failed non-current previews; project inspect lists their IDs. It does not erase projects, useful sources or superseded films. See recovery guidance.

Update or remove

The same installer supplies install, read-only check, recover and uninstall. It verifies owned outputs, preserves foreign changes and refuses to switch a release beneath active or unresolved work. Interrupted activation is recovered through its retained transaction, not a manual current symlink edit. Projects, configuration, profiles, audio and shared prerequisites survive uninstall.

Plugin updates and runtime updates are separate: inspect both the host's loaded plugin version and walkthrough --version --json. The runtime installer creates no competing standalone skill. Migrating a previous installation retires only its receipt-owned skill links after native-host discovery is verified.

Disabling or removing the plugin withdraws guidance from that host. It does not remove the independent CLI, stop its active session, delete projects or revoke authentication. Use session inspect and session close for browser custody before CLI removal; the installer refuses unresolved active work.

Development

The repository root is the native plugin. Executable code, its package metadata and build tooling live under runtime/; they are not plugin-startup dependencies.

cd runtime
npm ci --ignore-scripts
node scripts/provision-renderer.ts
node scripts/provision-session.mjs
npm run check
npm test
node scripts/package-release.mjs --out /absolute/new-release-output

Use qualified Node 24 and the Apple C toolchain for release builds. Packaging copies a committed positive inventory of owned source, skill, documentation, lockfile, patches and the compiled process reader. It does not bundle npm dependencies or their native binaries. Installation acquires those explicitly. Real browser, media and model qualification is separate from synthetic tests.

Design and operating boundaries

  • Vision: creator, viewer and value.
  • Architecture: plugin, runtime, project and media ownership.
  • Quality: acceptance gates, tests and defect ownership.
  • Security: authority, review classes, network and retention.
  • Browser authoring: public operating contract.

Scope is web products, local processing and one Remotion renderer. There is no native-desktop capture, GUI editor, permanent daemon, hosted TTS, music or automatic publisher. Ordinary browser networking is not exact-host isolation. Access to an application, agent review of its data and publication of a film are separate permissions.

Licenses

Walkthrough's original code, skill, documentation and artwork are MIT. Third-party notices separate dependency and model terms; the MIT grant does not relicense them or your footage and voice assets. The release archive distributes owned code, not modified Remotion copies or Remotion's FFmpeg binaries. Explicit installation acquires dependencies upstream.

Read the Remotion license FAQ for renderer eligibility and its FFmpeg notice for the separately licensed binaries. Voices and its models are separate installations. No personal voice assets, models, credentials or operator license assertion are included in this repository.