Imported from drPod/video-thing (
AGENTS.md). Install upstream withnpx skills add drPod/video-thing. Copyright stays with the author.
video-thing
Built on top of mem[v] (https://memv.ai). Meta DAT (Ray-Ban / AI glasses SDK) and Adobe Premiere Pro UXP (editing-side plugin surface) are separate surfaces — do not bridge.
Where to look first
- mem[v]:
docs/memv/README.md(catalog). Skills under.claude/skills/memv-*/auto-fire on memv tasks.docs/memv/llms-full.txtfor one-shot whole-corpus reads. - Meta DAT:
docs/meta-wearables/README.md(routing across 21 guides + 94 API class pages). - Premiere UXP:
docs/premiere-uxp/README.md(routing across 431 markdown pages — generic UXP platform vs Premiere-specific API split). - Delphi corpus:
docs/delphi-corpus/_indexed.md(what's been indexed in the synsci-delphi MCP).
Standing rules (always true)
- Spaces are isolation boundaries. Every memory write needs
space_id. No global writes. workspaceId(MCP) ==space_id(SDK). Same value, different name.- Use SDK types directly (
from memvai import ...). Don't invent wrappers. - No hand-rolled HTTP for mem[v]. SDK only. Escape hatches in
docs/memv/sdk/advanced.md. - Multimodal-first. Push raw text/files/video. Don't pre-process.
- Read
docs/memv/sdk/error-handling.mdBEFORE wrapping any mem[v] call in try/except. mcp__memv__*tools are dev-time only — never from app code. See skillmemv-mcp-vs-sdk.- Never guess SDK shapes. If
docs/memv/doesn't show it, flag — don't invent. - mem[v], Meta DAT, and Premiere UXP are separate surfaces. Do not bridge in code.
First time touching SDK code?
Run skill memv-bootstrap. Handles install, MEMV_API_KEY, smoke-tests.
memv MCP (4 tools, agent-side only)
whoami, list_workspaces, add_memory(workspaceId, memories[]), search_memory(query, workspaceId?, maxResults?). For anything else (file/video upload, space CRUD, graph, batch, error handling) → SDK in app code. Decision tree in skill memv-mcp-vs-sdk.
Delphi MCP (mcp__synsci-delphi__*) — external corpus only
| Need | Tool |
|---|---|
| External arxiv papers | search_papers |
| External code repos (Python/JS/TS only) | search_code / search_symbols |
| HF dataset discovery (cards only) | search_datasets |
| Library docs (React, Next.js, etc) | mcp__context7__query-docs (NOT delphi) |
| Our local mirrors / our own source | Read / rg (NOT delphi) |
Limits: never index our local mirrors or own source. Symbol search Python/JS/TS only — Kotlin/Swift/Go/Rust/C++ files index but search_symbols returns nothing for them. generate_report is markdown string-concat (title + abstract + chunks + first 20 citations + first 20 equations), NOT LLM synthesis. research (any mode: quick/deep/oracle) requires a configured provider — default Gemini, needs GEMINI_API_KEY in delphi server env. Equation/citation extraction is regex (false positives on inline $x$). HF datasets = cards only, no rows. Log every index in docs/delphi-corpus/_indexed.md.
Refresh local doc mirrors
./scripts/sync-memv-docs.sh # mem[v] docs (Tier 2: llms.txt + per-page curl)
./scripts/sync-meta-wearables-docs.sh # Meta DAT docs (Tier 3: Playwright scrape)
./scripts/sync-premiere-uxp-docs.sh # Premiere UXP docs (Tier 1: git clone source)Tier hierarchy used when integrating a new vendor surface: (1) git clone source markdown if vendor publishes it, (2) llms.txt + per-page curl if Mintlify-style, (3) Playwright + Turndown only if React-rendered with no source available.