project-workflow

name	project-workflow
description	Guides Dailies demo creation and self-review. Use when creating, compiling, rendering, evaluating, or improving Copilot-directed demo scenarios, especially Relay CLI demos with editor, terminal, audio, and video artifacts.

Dailies Project Workflow

Use this skill for Dailies work that benefits from focused guidance beyond AGENTS.md: creating demo scenarios, compiling timelines, generating safe audio fixtures, rendering previews, evaluating artifacts, or deciding whether a candidate is ready for human feedback.

Workflow

1. Read AGENTS.md and any relevant local docs.
2. Check git status --short.
3. Identify the smallest useful demo slice and the artifact that will prove it moved forward.
4. Use the anchored workflow: deterministic input collection, bounded AI decisions, deterministic validation, then rendering or handoff.
5. Keep live side effects off by default. Scenario commands are fixtures unless a task explicitly approves live execution.
6. Make the change with nearby documentation updates.
7. Run the closest available validation, usually npm run check for the initial scenario slice.
8. Improve the harness if the work reveals a durable command, constraint, validation step, or agent gotcha.
9. Ask for human feedback only after a complete candidate passes current self-review gates, or report the specific blocker.

Anchored Demo Loop

1. Plan: write or update a scenario with editor text, terminal commands, expected output, audio cues, and self-review criteria.
2. Compile: run npm run compile:demo -- <scenario> to produce a timeline artifact.
3. Preview: run npm run render:preview -- <scenario> to produce a self-contained local HTML preview.
4. Evaluate: run npm run evaluate:demo -- <scenario> and read the JSON report.
5. Iterate: fix the scenario, compiler, renderer, or evaluator based on concrete findings.
6. Candidate: run npm run render:candidate -- <scenario> --provider say for the complete local loop.
7. Audio: use npm run generate:audio -- <scenario> --provider speechify only when real provider output is intended.
8. Candidate gate: inspect npm run evaluate:candidate -- <scenario> output and sampled frames before asking the user to watch.
9. Review: inspect source, timeline, preview, evaluation report, visual captures, audio manifest, and video before asking the user to watch.

For scenarios with dailies:media, use Chrome capture. Confirm the media file stays under assets/, ffprobe accepts the requested source window, ffmpeg extraction maps timeline time to the intended source frame, the render manifest binds the source hash, and sampled frames visibly contain the expected media acts.

Quality Loop

1. Prefer quality over speed for consequential changes.
2. Use critical review before adding new third-party services, persistence, deployment, credential handling, permission changes, or broad architecture.
3. If an agent miss happens, update AGENTS.md, this skill, docs, scripts, tests, or guardrails with the smallest durable improvement.
4. Keep the evaluator honest: deterministic checks are a gate, not a claim that the demo feels good.

Relay and Audio Safety

• Do not touch the user's real TSRS queue by default.
• Do not execute live relay commands unless the task explicitly asks for that mode and the command path is allowlisted.
• Do not speak live audio from Dailies. Real voice generation, when requested, should write files through the TSRS non-speaking Speechify wrapper.
• Keep Speechify/network generation opt-in; offline stub fixtures are the default.
• Never commit secrets, API keys, raw private queue content, generated provider metadata, or shareable videos containing private paths.

ZShot Visual Checks

Use ZShot when browser rendering, visual state, or captured page artifacts would improve confidence. Default command path on jonmagic's Mac: ~/Library/Application Support/ZShot/zshot.

1. Start with zshot --agent-help when unsure.
2. Prefer HTML/MHTML smoke captures when license support for screenshots or PDFs is unavailable.
3. Put temporary outputs under zshot-artifacts/ or another ignored path.
4. Do not capture secrets or sensitive private pages unless the user explicitly approves the local-only artifact.

Do Not Use For

• Generic questions that AGENTS.md already answers.
• One-off notes that should live in README or docs instead of a skill.
• Live Relay queue operations outside an explicit Dailies demo workflow.