Instagram Reel preview

Create one 1080x1920 RGB JPEG per preview, using the local render.py beside this file. This is an image workflow, not a video edit or an official Instagram embed. Do not change source videos, approved covers, or another task's scripts. No account, online service, API key, transcription model download or upload is needed by the renderer.

Default publisher: fightthestroke, always with the actual approved FTS logo. Only an explicit request for another publisher permits an override. Subject names are subjects, never account owners. Read BRANDING.md for the approved source URL, pinned hash and safe local cache/download. Supply --publisher-logo; missing or mismatched FTS artwork fails rather than drawing initials or a substitute. Keep the logo whole and proportional: no redraw, recoloring, stretching or clipping.

Evidence before copy

1. Establish the actual local source, intended audience, display name, output directory and locale. Do not infer claims, identity or project branding from filenames. Publisher defaults to FTS; project labels remain optional and supplied.
2. Extract a local contact sheet at useful timestamps and inspect it visually. Choose an expressive real frame: eyes open, face in focus, sane headroom, no accidental gesture or obscured features. Inspect the selected frame at full size. If the default crop cuts the face, adjust focus_x/focus_y or use photo_fit: contain (full frame above the text).
3. Read the actual transcript before writing copy. Use an existing transcript or an already-installed local transcription engine/model. Inspect the engine's help and run it locally; verify names and uncertain phrases against the audio. No engine available: request a supplied transcript or approved copy rather than guessing. Never upload private media without explicit destination-specific authorization. Never automatically download a speech model.
4. Write 1-3 short editorial headline lines plus a supporting detail grounded in that transcript. Make a genuine curiosity hook, not a literal quotation or invented outcome. No fake followers, engagement, endorsements, verification badge, pity, medical sensationalism, or disability-as-inspiration framing. Preserve the person's agency. Keep timecoded support for each factual claim in the task's local evidence, outside git.
5. Choose/infer the language from the current task, not a prior session. English is the renderer's default only. Translate all authored text consistently: headline, support, CTA, Reels label and any descriptive project label. Preserve supplied proper names. For non-English video covers, pass cta and reels_label explicitly; the renderer does not translate or fact-check copy. For an existing cover, check its language first; the wrapper cannot translate the text baked into it.

Local contact-sheet example (choose a new destination; -n refuses overwrite):

ffmpeg -hide_banner -loglevel error -nostdin -n -protocol_whitelist file,pipe \
  -i "/path/to/source.mp4" \
  -vf "fps=1/5,scale=216:-1,tile=5x4" -frames:v 1 \
  "/path/to/output/contact-sheet.png"

This shows the first 20 sampled frames, not necessarily the entire video. For longer footage, inspect further intervals with -ss before -i; record the absolute chosen timestamp. The renderer is not a substitute for this inspection.

Render locally

Resolve SKILL_DIR to the directory containing this SKILL.md (not the working directory). Use Python 3.10+; check it with python3 -c "import PIL", and ffmpeg -version. Only if Pillow is missing, install the adjacent requirements.txt into an explicitly chosen environment. No blind installs. Supply a readable TTF/OTF font with the selected language's glyphs; optionally supply a separate body font. No proprietary fonts or media ship with the skill.

python3 "$SKILL_DIR/render.py" \
  --video "/path/to/source.mp4" --timestamp 12.5 \
  --title-line "A SMALL CHANGE." --title-line "WHAT HAPPENS NEXT?" \
  --subtitle "A transcript-supported detail." --display-name "Supplied name" \
  --publisher-logo "/path/to/local/cache/approved-fts-logo.png" \
  --font "/path/to/headline.ttf" --sans-font "/path/to/body.ttf" \
  --output-dir "/path/to/user/output" --output-name preview-new.jpg

The example copy is illustrative, not a factual claim to reuse. For repeatable work, copy example.json to a local, non-versioned location, replace its placeholders and run python3 "$SKILL_DIR/render.py" --config /path/to/job.json. CLI options override config values; repeated --title-line replaces the full list. Paths in JSON are relative to the JSON file, CLI paths to the current directory. --revise is CLI-only and requires the user's explicit revision request. Do not use it just to clear a filename collision.

Two styles

Cover margins are 48 px left/right, 85 px top, 86 px bottom at 1080x1920. The bordered card is 984x1749, so two full-size covers placed side by side have 96 px of white separation. Keep this spacing in a contact sheet; do not crop it away to make a collage. JPEG edge compression can slightly tint the pixels immediately beside the border.

For instagram-post, use the video options above with --style instagram-post, or wrap an approved cover directly:

python3 "$SKILL_DIR/render.py" --style instagram-post \
  --cover-input "/path/to/approved-cover.jpg" --display-name "Supplied name" \
  --publisher-logo "/path/to/local/cache/approved-fts-logo.png" \
  --font "/path/to/body.ttf" --output-dir "/path/to/user/output" \
  --output-name preview-post.jpg

Optional --project-label "Supplied label" adds supplied text next to the subject, not a replacement account. The header publisher remains fightthestroke. Only an explicitly requested different --publisher permits other branding or an optional --avatar; a neutral silhouette is never allowed in the default FTS preset. Do not supply headline/subtitle/timestamp with an existing cover. Every pixel of the input cover is represented through scaling into a 984x1480 maximum area, not cropped. Smaller text will therefore become smaller; inspect at phone size. This decorative framing is Instagram-inspired, not a captured post or official embed. Do not describe the controls as functional.

Configuration

--help lists all flags; JSON uses underscore names. Required video keys: video, timestamp (seconds), title_lines, subtitle, display_name, font, output_dir, output_name, plus publisher_logo for the FTS default. Wrapper input replaces video/copy keys with cover_input. Common optional keys: style, sans_font (same font when absent), locale (en), cta (WATCH THE VIDEO), reels_label (REELS), project_label, publisher (defaults to fightthestroke; explicit override only), publisher_logo (approved local original), avatar (non-FTS post only), photo_fit (cover or contain), focus_x, focus_y (0..1 crop centering; defaults 0.5).

Text shrinks only within readable limits, then fails with Text overflow. Shorten the copy or choose a fitting font; do not remove bounds checks, truncate silently, or claim multilingual glyph coverage without inspecting the output.

Delivery and safety

Keep outputs in the explicit user-selected folder (normally beside their media, never in this skill's source). Filenames are safe .jpg basenames. Each output has a .jpg.json provenance sidecar containing input/logo SHA-256 hashes, publisher, subject, approved FTS source URL when applicable, timestamp, drawn text/bounds and scaling geometry. Inputs are read-only and rehashed before publication. Hashes prove identity, not editorial accuracy.

Existing files are refused by default. Explicit --revise accepts only an unchanged generated JPEG with its matching generator/hash sidecar. Originals, symlinks and hard links to inputs are refused. Never remove a user's original. Keep the sidecar for revision; it contains local paths/copy and is private task data. A process interruption can leave a lock or mismatched JPEG/sidecar pair: report it and use a new output name, do not bypass the provenance check. Pair publication is not a filesystem transaction.

Open both full-size and phone-size outputs. Inspect faces, all glyphs/copy, whitespace, gradient border, safe text bounds and the wrapper's uncropped photo and CTA. Verify the JPEG is 1080x1920 RGB and the original hashes are unchanged. Report exact output paths and whether visual inspection happened. Do not claim publishing, Instagram UI accuracy, or human approval from a local render.