11 JC Lab Ideas to Film, Test and Write Today: Creator AI Editors, Local Agent Grids and a 64% Reliability Retest (2026)

Start with ChatCut Desktop, TeXbrain, or typeclip. Each one gives you a visible creator result in a short session: an agent-edited timeline, a PDF compiled in the browser, or a human-looking typed demo. The rest of the pack adds four tool reviews, two crypto attention checks, and one article that retests a famous "64%" LLM claim.

Video topics to film

ChatCut Desktop: can an agent cut a real creator clip on an editable timeline?

Format: How-I-tested

Shoot: Install the free desktop build, import one short JC Lab clip you already own, and give the same edit brief three ways: ChatCut's built-in agent, a connected ChatGPT/Codex session, and a Claude Code session if you have one. Watch every cut land on the timeline, then reverse one caption, one B-roll insert, and one music cue by hand. ChatCut's Product Hunt launch page says the app is built for humans and agents to edit together, that changes stay fully editable, that export can go out as XML to Premiere Pro, DaVinci Resolve, or CapCut, and that everything runs locally. 1

Hook / verdict frame: "Did the agent leave me a usable timeline, or a pretty first draft I still have to rebuild?" End on a three-column card: brief, agent result, minutes of manual cleanup.

Check before publishing: Use footage and music you own. Treat "runs locally" and "no editing experience needed" as claims to verify in the network panel and export settings, not as proof from a launch page.

TeXbrain: can a browser tab replace a paid Overleaf git sync for one real document?

Format: Tutorial

Shoot: Open TeXbrain in Chrome or Edge, load the article template, compile a short document with one math formula and one figure, then open a local project folder and compile again. Time first compile versus second compile, turn the network off after packages cache, and try one push or commit only if you use a disposable repo. TeXbrain's repository and live site say pdfTeX runs in WebAssembly with no backend, packages load on demand and cache locally, git works through isomorphic-git, Chromium browsers get File System Access, and Firefox/Safari fall back to a virtual filesystem. The author also lists hard limits: pdfTeX only, no XeTeX/LuaTeX, no bibtex/biber in the engine yet, and packages pinned to a TeX Live 2020-era set. 2 3 4

Hook / verdict frame: "Does the free browser path finish a real paper workflow, or only the first draft before bibliography breaks it?" Show the PDF, the offline recompile, and the first missing package or bibtex failure.

Check before publishing: Do not put thesis drafts with private data through a CORS git proxy you have not reviewed. Prefer a local folder and a throwaway remote.

MulmoTerminal: can one grid show which of five coding agents needs you first?

Format: How-I-tested

Shoot: Run npx mulmoterminal@latest on a machine with Node 22.9+ and a logged-in claude CLI. Open three sessions on the same toy repo in separate worktrees: one easy file rename, one failing test fix, and one permission-heavy task. Film the grid colours, the cockpit roster line for each cell, and the moment an amber "needs you" state appears. MulmoTerminal's repository says every cell is a real PTY, sessions can persist through tmux, work isolates in git worktrees, and Web Push can reach a phone when a turn finishes. 5

Hook / verdict frame: "When five agents run, is the bottleneck the model or me noticing the blocked one?" Cut between the amber cell, the roster summary, and the time it took you to answer.

Check before publishing: Use a disposable repository and dummy secrets. Do not point five agents at a production app, a paid cloud account with uncapped spend, or a client repo.

typeclip: does human-looking typing make a demo look real, or just slow?

Format: How-I-tested

Shoot: Install with Homebrew or build from source, grant Accessibility to the host terminal, copy a short demo prompt, and record two takes into the same field: a plain paste, then typeclip --seed 42 --wpm 85. Replay the same seed, abort mid-run with Cmd+Tab, and show the printed seed in the terminal. typeclip's README says it posts real keyboard events with jittered rhythm, optional typos, a focus guard that halts on app switch, and reproducible takes via --seed. 6

Hook / verdict frame: "Does the audience believe the typed prompt more than a paste, and is the extra take time worth it?" Put the paste cut and the typed cut side by side with timestamps.

Check before publishing: typeclip can type whatever is on the clipboard into the focused field. Clear secrets from the clipboard first, and keep the countdown preview on screen so viewers see you chose the text.

Tools and apps to review

Tellie Prompter 1.5: a teleprompter that follows what you actually said

Format: Review

Shoot: Load a 60-second script, record one take where you pause and skip a line, then one take where you improvise a hook. Watch whether the prompter stays with the spoken words, whether the "points you have not said" warnings appear, and whether the app is invisible in a Zoom or screen-recording frame. Tellie's Product Hunt page describes a 3 MB Mac app that runs on-device with no account and no telemetry, tracks important points and overruns in version 1.5, offers a 10-day free trial, and sells a one-time license with a Product Hunt code. 7

Hook / verdict frame: "Can I go off script without losing the next line I still need?" Show the skipped line, the recovery, and the debrief if one appears.

Check before publishing: Confirm the invisible-to-recorder claim with a real screen capture. Treat pricing and the Product Hunt discount as launch-page claims and re-check the checkout screen before stating a final price.

LatticeDB: one local file for graph links, vectors, and full-text

Format: How-I-tested

Shoot: Install the Python or CLI path, create a tiny knowledge graph with three documents, three chunks, and two authors, store placeholder embeddings with the built-in hash helper, then run one query that mixes vector distance, full-text match, and a graph hop. LatticeDB's repository describes a single-file embedded property-graph database with HNSW vectors, BM25 full-text, WAL durability, and bindings for Python, TypeScript/Node, and Go. It also warns that hash_embed is a deterministic placeholder, not a semantic model. 8

Hook / verdict frame: "Can one local file replace three tools for a small agent-memory demo?" Show the file size, the query, and the point where a real embedding model becomes mandatory.

Check before publishing: Keep the test dataset synthetic. Vendor latency tables on the README are author benchmarks on specific hardware; re-run your own tiny workload instead of repeating those numbers as universal facts.

Screenify Studio: can an AI agent film a product demo you would still publish?

Format: Review

Shoot: On a Mac with Apple Silicon, record one plain screen capture of a simple web flow, then try the plain-English agent/CLI path if the free tier exposes it. Compare free-plan limits against what you need for a YouTube short: watermark, export length, resolution, and whether AI captions are locked. Screenify's site says the app is macOS 13+ Apple Silicon only, free forever for unlimited recording with a watermark and 1080p exports up to five minutes, and paid one-time Pro tiers for 4K and on-device AI features. Product Hunt also frames it as demos recorded by an AI agent that can be driven from a CLI. 9 10

Hook / verdict frame: "Is the free path enough for a publishable short, or only a watermarked draft?" End with a checklist: watermark, max length, AI captions, and minutes to a final export.

Check before publishing: Re-read the free versus Pro feature matrix on the live site before stating what AI features ship without payment. Demo only apps and sites you are allowed to film.

Apache Maka: a local-first agent workspace that keeps the run record

Format: Scam check

Shoot: Clone the repository on an Apple Silicon Mac, follow the source-build path (npm ci, npm run dev), add one disposable model connection, and run one bounded task such as "summarise this toy repo." Inspect whether messages, tool calls, approvals, and stop reasons are written to a local record you can reopen after a restart. Apache Maka's README describes a local-first agent workspace under Apache incubation, with Desktop plus CLI surfaces, sandbox approvals, durable execution records, and an explicit note that prebuilt downloads are not yet approved Apache releases. Desktop currently targets Apple Silicon; Windows is an unsigned preview; Linux is not supported yet. 11

Hook / verdict frame: "If the agent crashes mid-task, can I prove what it already did?" Show the session timeline, the local record, and the first permission prompt outside the sandbox.

Check before publishing: Build from source and say so. Do not present incubation status, early public builds, or author speed claims as a finished product endorsement. Keep API keys in a disposable account.

Bitlayer / BTR: can a search spike survive a liquidity check?

Format: Trend react

Shoot: CoinGecko's 24-hour trending return placed Bitlayer second, with a market-cap rank of 829. A morning simple-price snapshot recorded about $0.083. The public coin page title also carried a roughly $82 million 24-hour volume figure at fetch time. 12 Refresh the page while filming, then record the live price, volume, spread, order-book depth, and the timestamp from one liquid venue.

Hook / verdict frame: "Did attention arrive with enough depth for a paper trade you can measure, or only a thin book?" Apply one written paper rule, include estimated slippage, and compare the result with a fixed holding period.

Check before publishing: The trend list measures search attention. It does not establish a trade, a catalyst, or a future price. Use paper trades only and label every market number with the capture time.

peaq / PEAQ: same paper rule, different book

Format: Comparison

Shoot: CoinGecko's 24-hour trending return included peaq and gave it a market-cap rank of 416. A morning simple-price snapshot recorded about $0.022. The public coin page title also carried a roughly $5.4 million 24-hour volume figure at fetch time. 13 Refresh both cards while filming, then put peaq beside Bitlayer and apply the same paper-trading rule, window, and exit condition.

Hook / verdict frame: "Does the higher-ranked name still lose the clean test once spread and depth enter the sheet?" Keep the rule identical so the only variable is the book.

Check before publishing: CoinGecko ranks and prices move during the day. Re-capture both screens immediately before filming, state the local capture time, and avoid a buy, sell, or price target.

Article idea

Julia Bench: retest the "64% reliable" LLM claim on a fixed question set

Format: Article

Write: Open with the public claim that large language models are "relevant" or "reliable" about 64% of the time, then show where the number comes from: Bang et al. (2023) reported 63.41% answer accuracy for a December 2022 ChatGPT snapshot on 634 hand-picked reasoning items, not a true/false scan of millions of facts. Walk readers through the Julia Bench reconstruction: a balanced 200-question subset, released gold answers, merge and scoring scripts, and the author's own re-runs that scored modern models far above the original figure on that same task. 14 15

Reader promise: Separate three things viewers usually mash together: the original paper's narrow accuracy number, the later media claim about general reliability, and what a 2026 model scores on the reconstructed set. Include one sample question, the grading rule, and a small table of reported scores. State clearly that a single fixed set is not "LLM reliability" in general.

Check before publishing: Quote the paper and the reconstruction repository, not second-hand interview summaries alone. If you run models yourself, log model name, date, temperature, and the exact CSV split. Do not turn the article into investment or product advice.

Film in this order

Run TeXbrain first while Chrome is open: first compile, offline recompile, and the bibtex limit are quick to capture.
Record the typeclip paste-versus-type comparison next; it needs only Accessibility permission and a short clip.
Install ChatCut Desktop and complete one agent edit while the machine is free for a longer timeline test.
Review Tellie and Screenify on the same Mac session, because both are creator-facing and macOS-bound.
Spin up MulmoTerminal with three disposable agent tasks, then build LatticeDB and Apache Maka from their repositories.
Refresh CoinGecko, capture Bitlayer and peaq, and paper-trade the same written rule.
Write the Julia Bench article after you have opened the CSV, the scoring script, and at least one sample question on camera or in screenshots.