| name | transcreveai-video-intelligence |
| description | Use to turn video URLs or media files into evidence-backed knowledge dossiers with TranscreveAI. Trigger on Reels, YouTube, TikTok, Loom, Vimeo, X/Twitter video links, local media files, video summaries, dossier requests, RAG over video runs, or requests to use TranscreveAI. |
TranscreveAI Video Intelligence
Use this skill when the user sends a video URL/file or asks Codex to extract, summarize, analyze, index, or ask questions about video content with TranscreveAI.
Tooling Preference
- Prefer the TranscreveAI MCP tools when available:
sources_probe for source pre-checks.
agent_run for the full probe/analyze/index/ask workflow.
agent_batch for saved lists of sources.
index, ask, runs_list, runs_show, and shared_catalog for retrieval.
- If the MCP tools are not available in the current Codex thread, use the CLI.
- CLI command preference:
- First try
transcreveai.
- If it is not on PATH, use the plugin wrapper at
./scripts/transcreveai from the installed plugin root.
- MCP server command for local registration:
bash ./scripts/transcreveai-mcp --transport stdio from the installed plugin root.
- The plugin wrappers add
/opt/homebrew/bin and /usr/local/bin to PATH so Homebrew FFmpeg/Tesseract installs are visible to Codex-launched processes.
- If TranscreveAI is not installed globally, the wrappers create a venv under
~/.cache/transcreveai-codex-plugin and install transcreve-ai[mcp,rag] from pinned commit 7490586e86b57eefae377dc6839e9476f89ffd8d of https://github.com/DeHor-Labs/transcreve-ai.git.
Required Nested Handoff
Whenever TranscreveAI is used as a nested capability for another agent or workflow:
- Keep
run_id, out, and index-db identifiable for the caller.
- Use temporary retention by default when the caller only needs extraction, summary, or an answer from YouTube, Reels, TikTok, local media, and similar sources.
- Preserve artifacts or index in the user's real knowledge base only when the caller/user asks to save, index, audit later, or reuse the dossier.
- When durable reuse is requested, run
transcreveai share RUN_ID --json or call MCP share_run after analysis. If the run used an isolated index, pass the same --index-db or use transcreveai share --run-dir "$RUN_DIR" --json. It writes handoff.md, manifest.json, knowledge.md, and analysis.json, and updates share-root catalog.json/index.md. To rediscover durable packets later, use transcreveai share --catalog --json or MCP shared_catalog.
- If the dossier is preserved or indexed, explicitly say:
O dossie que voce criou foi salvo para voce como conhecimento.
- Include the path to
knowledge.md, the run_id, and whether the knowledge was saved in the user's real index or in an isolated agent index.
- If the run was temporary and cleaned up, do not claim it was saved; report the
run_id, that the answer was based on generated artifacts, and that temporary files were removed.
Safe Defaults
- For smoke tests, demos, and automated validation, isolate state:
--index-db /tmp/transcreveai-agent.db
--out /tmp/transcreveai-agent
--ai off
--provider local
--force
- For temporary production-like agent runs, create a dedicated temp directory:
TMP=$(mktemp -d "${TMPDIR:-/tmp}/transcreveai-agent.XXXXXX"),
use --index-db "$TMP/index.db" and --out "$TMP/runs", read the generated artifacts, then rm -rf "$TMP".
- If a temporary run used the real index, remove it with
transcreveai runs rm RUN_ID --force before deleting files.
- Do not expose API keys, cookie contents, or complete sensitive URLs in logs or final answers.
- Use
--cookies-browser chrome only for user-owned browser state and only when needed for sources such as Instagram.
- Base final answers on generated artifacts, especially
knowledge.md, analysis.json, and template files. Do not create a parallel manual dossier and pretend it came from TranscreveAI.
- If analysis fails before artifacts are written, check whether
ffmpeg, ffprobe, and tesseract are visible in PATH before treating it as a TranscreveAI bug.
Recommended Agent Flow
- Probe the source:
transcreveai sources probe "SOURCE" --json
Read kind, adapter, requires_cookies, and notes. If cookies are required, prefer --cookies-browser chrome for user-owned sources.
- Run the agent workflow:
TMP=$(mktemp -d "${TMPDIR:-/tmp}/transcreveai-agent.XXXXXX")
transcreveai --index-db "$TMP/index.db" agent run "SOURCE" --out "$TMP/runs" --json
For an isolated no-cost smoke:
transcreveai --index-db /tmp/transcreveai-agent.db agent run "SOURCE" \
--out /tmp/transcreveai-agent \
--ai off \
--provider local \
--force \
--json
- Add templates when useful:
- Use
--template content for creator, marketing, product, sales, distribution, or content workflow videos.
- Use
--template skill for videos about agents, prompts, skills, Claude, Codex, automations, or reusable workflows.
- Read generated
content.md/content.json/content.csv and skill.md/skill.json before answering about those artifacts.
- Read the evidence:
- Always inspect
knowledge.md.
- Inspect
analysis.json for structured metadata, source, paths, transcript quality, and run details.
- If the user asks a question over the run, index and query:
transcreveai index RUN_ID
transcreveai ask "QUESTION" --run-id RUN_ID --top-k 8
- Report compactly:
- Summarize what the video actually supports.
- Separate evidence from inference when making product, business, or technical recommendations.
- Cite artifact paths only when the dossier is preserved. For temporary runs that are cleaned up, cite the
run_id and cleanup status instead.
- State limitations when transcript, OCR, visual context, or source access was weak.
- Unless the user asked to preserve/index the dossier, remove the temp directory after reading the artifacts and mention that temporary files were removed.
Batch Flow
For multiple URLs or files:
transcreveai agent batch ./sources.txt \
--template content \
--template skill \
--strict \
--json
Use --strict when any failed item should block the caller. Read success, ok_count, failed_count, batch.md, batch.json, and per-run template_paths.
Expected Artifacts
knowledge.md: human-readable dossier.
analysis.json: structured run metadata and analysis.
- Optional
content.md, content.json, content.csv.
- Optional
skill.md, skill.json.
- Optional
batch.md, batch.json for batch runs.
Example Starter Commands
transcreveai sources probe "https://www.instagram.com/reel/..." --json
transcreveai agent run "https://www.instagram.com/reel/..." --template content --template skill --json
transcreveai agent batch ./sources.txt --template content --template skill --json
transcreveai ask "What decisions does this video support?" --run-id RUN_ID --top-k 8
Palestra tecnica, aula ou tutorial longo
Use --frame-strategy slides: os frames saem na troca de tela em vez de a cada
N segundos, entao nenhum slide se perde nem aparece repetido. O padrao auto ja
liga isso sozinho quando o video e longo demais para o intervalo cobrir.
transcreveai analyze "<url>" --ai auto --frame-strategy slides
Codigo mostrado na tela e reconhecido e sai em bloco cercado no knowledge.md,
com a indentacao reconstruida a partir da posicao do texto na imagem. Em
slides os frames sao gravados em PNG: o JPEG borra texto fino e arruina o OCR
de codigo.