11 JC Lab Ideas to Film, Test and Write Today: Sandboxes, Document Parsing and a Cursor Deadline (2026)

Tokensift is the quickest high-signal start: it can turn wasted prompt space into a before-and-after test. Sesame gives you a useful privacy check with fictional credentials, while the fresh OpenAI-Cursor contract decision can anchor a timely written piece. The remaining ideas give you four hands-on videos, four app reviews, two measured crypto attention checks, and one article angle.

Video topics to film

Tokensift: find the expensive parts of a prompt

Format: Tutorial

Film today: Install the local tokensift CLI from the official repository. Run the same prompt twice: first with repeated instructions, pretty-printed JSON, and a UUID; then with the repeated text removed, compact JSON, and a placeholder. Capture the linter's findings, token counts, and estimated dollar difference. Tokensift's repository describes a local deterministic engine with 20 rules, real dollar-cost findings, and test matchers; the repository says OpenAI counts are exact while Claude figures are estimates. 1

Hook: Can a five-minute cleanup make an agent prompt cheaper without changing its answer? End with the cleaned prompt and the exact changes viewers can copy.

Boundary: Keep the Claude number labelled as an estimate. Test the two prompts on the same model, settings, and input before drawing a cost comparison.

Sesame: local password manager or unfinished beta?

Format: Scam check

Film today: Use fictional logins and a disposable vault. Import a small test set, add a password and a 2FA entry, create a backup, inspect the local storage path, and walk through the optional website account. Sesame's site says the vault stays on the device, supports imports from 15 formats, has no analytics or ads, and is AGPL-3.0-or-later. The site also identifies the app as a public beta, with sync and mobile access unavailable and a browser extension still awaiting store submission. 2

Hook: Does "local-first" survive an import, backup, and account check? Show viewers exactly what the app stores locally and what remains unfinished.

Boundary: Use invented credentials throughout. Keep a verified copy of any real vault outside the test machine and treat the independent security review as pending until the site says otherwise.

BLAST: fork a disposable VM in front of the camera

Format: How-I-tested

Film today: Follow the official setup, fork an Ubuntu VM, run a harmless command, take a snapshot, fork the VM again, and compare the files and process state. The BLAST repository describes a single-binary local sandbox tool with VM forking, snapshots, syncing, migration, resource scheduling, and a 7 MB binary under an MIT license. 3

Hook: What state survives a fork, and what state stays isolated? Make the answer visible with one file, one process, and one snapshot.

Boundary: Use a disposable VM and harmless commands. Present the result as a workflow demonstration; a short demo gives you a bounded observation about this setup, not a security certification for hostile workloads or VM escapes.

Conduct: block the agent before the command runs

Format: Comparison

Film today: Configure one harmless allow rule, one block rule, and one warning rule. Run the same small task through the supported Claude Code, Cursor, Copilot, or Codex integration, then inspect the audit trail. Conduct's repository describes signed policies, blocks and warnings, hash-chained audit logs, a fail-closed default, a router, compliance packs, and a self-hosted Docker Compose path. 4

Hook: What changes when a guardrail stops a tool call before execution instead of recording it afterward? Put the three outcomes on screen: allowed, warned, and blocked.

Boundary: Run harmless local commands in a throwaway project. Treat the result as a comparison of the tested policy path; the run does not establish protection against every agent, integration, or policy bypass.

Tools and apps to review

Hy4 Preview: can the long-context claim survive a small task?

Format: Review

Film today: Start from Tencent Hy's official access path and record whether Hy4 Preview is available in your region or account. Give Hy4/Hy3 and one comparison model the same short coding or document task. Show context length, setup time, test execution, recovery after an error, and the final output. Product Hunt's launch entry describes Hy4 Preview as a 770B mixture-of-experts model with 49B active parameters and a 1M-token context for agentic coding, game development, and document analysis. Tencent's official site currently exposes Hy3, AI Studio access, and a Hy4 preview video. 5 6

Hook: The review question is practical: can the model hold a longer task together and recover when the first attempt fails?

Boundary: Attribute the 770B, 49B, and 1M figures to Product Hunt's launch copy. Treat access, speed, and quality as results from your recorded test rather than settled specifications.

Cohere Parse 5: invoice, table, diagram

Format: How-I-tested

Film today: Submit one synthetic invoice, one scanned table, and one diagram. Compare Markdown output with block output, then inspect whether the returned bounding boxes preserve the information a search or agent would need. Cohere's quickstart shows model="parse-v5.0", Markdown output, optional blocks, tables, and bounding boxes. Cohere's August 27 announcement lists $1.50 per 1,000 pages and vendor benchmark results, so the price and performance claims need a retest. 7 8

Hook: Does the parser preserve structure, or does it merely turn a page into readable text? End with one field that downstream search can recover and one field it loses.

Boundary: Use synthetic documents and a disposable API key. Label the price and benchmark figures as Cohere's claims, and record the model version and document type beside every result.

seendiff: review a large agent diff in chunks

Format: Review

Film today: Create a deliberately large disposable code diff. Open it in seendiff, review the proposed chunks, mark sections seen and unseen, and ask the coding agent to explain one change. Product Hunt's launch entry describes seendiff as a local diff viewer for large coding-agent changes with review chunks, progress tracking, and agent explanations. 9

Hook: Does chunked review reduce the changes you miss when an agent edits a whole project?

Boundary: Work in a throwaway repository and verify every changed line yourself. Frame the episode around review workflow; a smoother diff view is evidence about navigation, while code correctness still needs tests and human review.

Any Command: phone remote versus second screen

Format: Comparison

Film today: Pair an Android phone with a Windows PC and compare three jobs: mouse and keyboard control, one-window streaming, and gamepad input. Test the local-network path and Bluetooth path separately, then compare setup time, latency, reconnects, and the free-tier limit. Product Hunt's current launch entry describes Any Command as an Android-to-Windows remote that can act as a second screen, trackpad, keyboard, and gamepad, with one-window streaming, local-network or Bluetooth connections, and a free tier without an account. 10

Hook: Which job does the phone handle well, and where does a cheap dedicated peripheral still win?

Boundary: Use a spare Windows profile and a trusted local network. Check the app's permissions and pairing screen before connecting, and record the exact app version and connection method.

Fogo (FOGO): a search spike worth paper-trading, not chasing

Format: Trend react

Film today: Capture the current CoinGecko trending list, then refresh the Fogo page immediately before recording. The returned 24-hour list placed Fogo sixth among 15 coins, with a market-cap rank of 651; a separate price call returned a snapshot of $0.00761689. CoinGecko defines this trending list as coins ordered by popular user searches over the last 24 hours. 11 12

Hook: Can attention survive a repeatable check? Show the search position, liquidity and spread on your chosen exchange, then compare the same token again after a fixed interval.

Paper rule: Record a hypothetical entry at the capture price, a fixed exit time, and a fixed maximum loss. Use zero real funds and publish the result as an attention experiment.

Boundary: The search position is a popularity signal. It supplies no thesis, price target, or buy/sell instruction. Refresh every price and rank immediately before filming because both can move between the morning scan and the recording.

Velvet (VELVET): separate the trend list from the tape

Format: Trend react

Film today: Repeat the same card used for Fogo. CoinGecko's returned list placed Velvet tenth among 15 coins, with a market-cap rank of 457; the separate price call returned a snapshot of $0.104239. The CoinGecko Velvet page provides the coin-specific detail route for the check. 11 13

Hook: Does a higher market-cap rank make the attention signal easier to trade, or does the same short-lived noise remain?

Paper rule: Log the snapshot, spread, volume, hypothetical entry, fixed holding period, and exit before looking at the result. Compare the card with Fogo using the same fields.

Boundary: Popular searches and a live price are observations, not investment advice. Use paper trading only, avoid leverage, and show the capture time and refreshed price on screen.

Article idea

OpenAI and Cursor: what a model-provider exit means for developers

Format: Article

Write today: Build a short explainer around OpenAI's August 28, 2026 notice that it intends to wind down its contract supplying OpenAI models to Cursor after Cursor's acquisition by SpaceX. OpenAI proposed November 12, 2026 as the shutoff date, said the date used the maximum notice allowed by its contract, and said future models would not be provided to Cursor. 14

Reader promise: Give developers a checklist for separating a contract notice from an immediate product shutdown: check which model is active, whether a workspace has a provider fallback, how API billing is configured, and which project settings need export before the proposed date.

Reporting boundary: Keep OpenAI's reasons attributed to OpenAI. Verify Cursor's current model list and access during reporting, and label every prediction about migration, pricing, or user impact as an open question until Cursor publishes it.

Film in this order

Capture the Fogo and Velvet screens first, then refresh both prices immediately before recording.
Shoot the Tokensift before-and-after prompt test and the Sesame disposable-vault check.
Record the Any Command comparison if the phone and Windows machine are already available.
Reserve the longer setup blocks for BLAST, Conduct, Hy4 Preview, and Cohere Parse.
Use the seendiff diff review as a compact screen-recording episode.
Write the OpenAI-Cursor explainer after checking the current Cursor model list.

Market prices, search positions, access, launch-board copy, and product features can change during filming. Refresh the time-sensitive screens and label vendor or launch-board claims before publishing.