Know Today
👀 Keep an eye on
- Reported screen-and-audio skill recording could turn demonstrated work into reusable automations, making explicit decision criteria more valuable than recorded clicks.
- Terminal coding agents are converging on a common workflow, so task-level cost, latency, diff quality, and intervention rate matter more than a single benchmark.
- Tool-using agents need default-deny network access, scoped credentials, audit logs, and approval gates because sandbox mistakes can become external actions.
- Highly directed video generation is becoming practical through many reference inputs and timeline controls, but coherence failures and cost still limit volume use.
- Generative features tied to real places need misuse testing and a rapid rollback path before broad release.
🛠️ Practical workflows
| Pattern | Useful implementation detail |
|---|---|
| Record a process | Narrate why a choice is correct, its exceptions, and rejected alternatives; preserve the policy, not merely the UI path. |
| Compose automation | Break work into small skills with explicit inputs, outputs, owners, and failure behavior before adding an orchestration layer. |
| Generate and check | Pair every generator with a rubric using good and bad examples, then require evidence for each pass or fail decision. |
| Choose a model | Separate frontier cloud reasoning, self-hosted cloud deployment, and lightweight local tasks; open weights do not make a multi-trillion-parameter model locally practical. |
💭 Opinions worth testing
- Opinion: Teaching an agent to evaluate quality may produce more durable value than teaching it to generate output, because acceptance criteria travel across tasks better than a long prompt.
- Opinion: Video demonstrations can transfer tacit, visual judgment better than written instructions alone, while written specifications remain necessary for permissions and governance.
- Prediction: Teams that capture how reviewers think—not just how operators click—will get more useful automation from recorded workflows.
⚠️ Caveats
- Verify before adopting: reported product availability, pricing, integrations, benchmarks, and security behavior may be incomplete, stale, or inaccurately described.
- A workflow can appear correct while breaking semantics, such as replacing spreadsheet formulas with static values; test fresh data, permissions, side effects, and idempotency.
- Claims about incidents, rollbacks, and consumer-agent capabilities need independent confirmation before they inform policy or architecture.
✨ Try this today
Build a reviewer before a generator
Choose one small artifact such as a pull-request summary or spreadsheet update, give an LLM two good and two bad examples, and ask it to produce an evidence-linked checklist. The outcome should be a review rubric you can run manually before trusting generation.
Run a two-agent coding comparison
Give two coding agents the same bounded repository task in separate working copies, with tests and a fixed budget. Compare passing tests, diff readability, elapsed time, and how often you had to intervene; review all changes before merging.
Turn a stable routine into atomic steps
Document a repeatable task as three to five units with inputs, outputs, and an explicit stop condition, then ask an LLM to draft an API-first implementation. The outcome should reveal which steps are safely automatable and which still need human judgment.
Test a media brief, not a prompt
Create a short video brief with reference images, a timing plan, and a defined acceptance checklist, then generate only a few candidates. You will learn whether controllability and usable-clip cost fit the work better than headline demo quality.