# A first look at the CorvidLabs tools

These silent recordings use a tiny to-do list. Each large GUIDE caption explains the goal or result in plain words. The terminal contains real command output. You do not need to read every terminal line to follow the story.

| Chapter | Length | Watch | Read |
| --- | --- | --- | --- |
| Request & plan | 1:03 | [Video](beginner-ask.mp4) | [Transcript](beginner-ask.txt) |
| Build & test | 0:59 | [Video](beginner-mistake.mp4) | [Transcript](beginner-mistake.txt) |
| Review & resume | 0:52 | [Video](beginner-tomorrow.mp4) | [Transcript](beginner-tomorrow.txt) |

Download [the sample](beginner-corvid-task-demo.zip) to run the same commands. Its README lists prerequisites and marks the test that should fail. Run each command separately. The prepared repair restores the original code.

## What happened

- The app created, completed and reopened a task while keeping its earlier changes.
- A prepared mistake removed those earlier changes. The existing behavior tests failed: 5 passed and 3 failed.
- The prepared repair restored the original source. The same unchanged tests then passed: 8 passed and 0 failed.
- A fresh extraction reproduced this sequence and the final `chough next --environment local` command. Both test files were compared before and after.

The sample spec is a draft. Its structural check covers source-file presence and module dependency links; API/export and required-section validation are skipped. This check does not establish that the app behaves correctly. The separate tests demonstrate only their chosen assertions.

The repair is scripted. The saved checkpoint is real, with historical checks and human review still pending. These videos do not measure whether an incoming agent understands everything. SAMPLE-1 is an authored example request, not a human quotation.

## The wider development lifecycle

The guides connect the example to product intent, design, planning, development, QA, acceptance and optional release. These are explanations of possible team contributions, not a record of those people participating. Visual design and release are explained but not performed in the CLI sample. People can work through their coding agent and review results without learning the terminal commands. The page explains all seven stages and the repair loop.

## Recording details

Recorded on macOS arm64 with hi 0.8.0, specsync 6.2.0, fledge 1.8.0, chough 0.5.0, Bun 1.4.0 and VHS 0.12.1. Installed executable paths/hashes and actual command exits are retained in each chapter's `.receipts.json`. [Metadata and asset hashes](beginner-metadata.json) identify the media. Full raw receipts, VHS sources, failed attempts and encoding scripts are in the experiment archive.

The videos are 1600×1700 with the full terminal frame above authored GUIDE cards and 300px of clearance below them for native player controls. Selected test/checkpoint lines are explicitly labeled as excerpts. Transcripts contain the complete readable output; original raw terminal bytes are retained in the archive. Each video has matching English VTT captions.

An initial HC shell found a different Fledge and lacked Bun. Pinned paths fixed the recording environment. Two capture attempts timed out on long output; shorter labeled excerpts fixed the viewport issue without changing test status. The local FFmpeg build lacked drawtext, so guide cards were rendered separately and overlaid. These failed attempts are preserved.

The older recordings, GIFs and sample remain unchanged. Their [historical recording notes](README.md) and [README embedding examples](README-EMBEDS.md) describe that earlier version.