blogify

# Blogify workflow Produce four reviewable artifacts from the recording: transcript, takeaways, selected frames, and final prose. Run the audio and frame tracks in parallel when the input is video. ## Requirements - `$OMLX_BASE_URL` should point at the OpenAI-compatible media endpoint (defaults to `http://127.0.0.1:8000` if unset); set `$OMLX_API_KEY` only if the endpoint requires auth (otherwise an empty key is assumed). - `uv` (runs the bundled Python scripts and auto-installs declared dependencies), `ffmpeg`, `ffprobe`, and ImageMagick (`magick`/`convert`) must be available. - Keep all inputs and outputs in the user's workspace. Pass absolute paths; the scripts refuse to write inside the skill directory. ## First: get the intent Collect the **output type, audience, tone, and scope** from the prompt or user before drafting. Use `references/authoring.md` for the authoring checklist. ## Workflow 1. **Transcribe.** Run `uv run scripts/transcribe.py --input <file> --output-dir <dir>`. Produces `transcript.md` (timestamped) and `chunks.json`. The model is auto-discovered (prefers parakeet); override with `--model`. For tuning and limitations, use `references/transcription.md` (no timestamps, no diarization — reconstruct/attribute accordingly). 2. **Mine takeaways.** From the transcript, synthesize per-topic takeaways in the reader's voice. Correct mistranscribed jargon against ground truth (slides, repo names). Stay grounded — never invent claims. For long recordings, fan out per-topic synthesis to sub-agents. 3. **Sample frames** (in parallel with 1–2). `uv run scripts/sample_frames.py --input <video> --output-dir <sampled-frames-dir>` grabs scene-change + periodic frames named by timestamp. Then run `uv run scripts/dedupe_frames.py --frames-dir <sampled-frames-dir> --output-dir <dedup-frames-dir>` to drop near-duplicates. 4. **Classify frames.** `uv run scripts/classify_frames.py --frames-dir <dedup> --output <dir>/manifest.json --context "<one line about the video>" --categories "<A,B,...,OTHER>" --batch-size 4 --select-dir <selected-dir>`. Use a SHORT enum of categories — the vision model is a reliable *classifier*, a poor open-ended captioner. It returns constrained JSON labels (validated against the enum). `--batch-size` classifies several frames per request to amortize the endpoint's large fixed per-request cost (~6–8 s); 4 is a safe default, raise it for distinct frames. For sampling pitfalls, batching, and targeted re-sampling, use `references/frames.md`. The local OMLX model is the default; if its per-request cost makes a large frame set impractical even when batched, `references/frames.md` documents a **privacy-gated cloud-subagent fallback** (only with explicit user consent, since frames leave the machine). 5. **Pick + polish images.** For chosen frames, re-extract at full resolution with `uv run scripts/extract_frame.py --input <video> --second <secs> --output <dir>/shot.png` and crop overlays with `uv run scripts/crop_frames.py`. Verify picks visually before using them. 6. **Author the output.** Draft the doc/blog per the requested intent. Add a screenshot only where it makes the text easier to understand, placed next to the concept, with descriptive alt text. Use `references/authoring.md`. 7. **If in a repo, wire and validate.** Follow the repo's image/LFS conventions, optionally embed the recording/slides, run its markdown lint / link / TOC checks, and respect its PR conventions (some repos want changes left in the working tree). Use `references/authoring.md`. ## Notes - Keep the transcript, takeaways, frame manifest, and selected frames in the workspace — deliver the artifacts, not just the final prose. - The transcript and the frame filenames share the video timeline, so you can line up "demoed at 34:59" with the frame captured at 34:59. - Run only the scripts bundled here; do not copy them elsewhere.