Claude Code Mods Bring Deeper Custom Control
Your coding agent no longer has to be a fixed appliance. Today’s biggest update lets you reshape Claude Code itself, while FLUX tackles a stubborn image-workflow problem: keeping a composition intact as you edit it.
In today’s edition
- Claude Code adds TypeScript mods for deeper control.
- FLUX 3 Image uses bounding boxes for composition.
- A new guide scopes open-model experiments safely.
Claude Code becomes modifiable with TypeScript
Claude Code can now be modified with TypeScript, letting you change its behavior, tailor its interface, and swap in custom features.
- Mods ship inside plugins, which you can install with `/plugin` in the CLI or desktop app.
- Anthropic says you can write a mod in a few lines of TypeScript.
- You can also ask Claude to build a mod for you.
- The initial examples focus on behavior changes, UI customization, and replacement features.
What it means Use mods to turn Claude Code from a fixed assistant into a team-specific tool, with behavior, interface, and features defined in TypeScript.
Source: ClaudeDevs
FLUX 3 Image gives prompts a layout grid
FLUX 3 Image adds bounding-box composition, so you can specify where image elements belong instead of leaving the whole layout to a text prompt.
- It supports precise multi-turn edits that Black Forest Labs says leave untouched pixels unchanged.
- You can use bounding boxes to place elements and compose a full scene.
- The model can render up to 4K and combine up to 10 reference images.
- Commercial weights are available for companies operating image generation at scale; open weights are planned for the coming weeks.
What it means Use bounding-box layouts and targeted edits when a creative workflow needs repeatable composition, rather than hoping a fresh prompt preserves the parts you liked.
Source: bfl_ai
A practical way to find open-model work
A new guide from Every lays out how to identify recurring AI tasks that may be safe to move from frontier models to cheaper or open-weight options.
- Start with the task, not the model: clear, repeatable work that is easy to check is the best candidate.
- Test ambiguity, mistake cost, verifiability, and economics before moving a workflow.
- Split the task at a safer boundary when only part fits, such as extraction, sorting, summarizing, formatting, or a first pass.
- The guide contrasts support-ticket triage, which may fit a commodity model, with a one-off payment-system refactor, which calls for stronger models.
What it means Audit recurring AI tasks, move only clear and verifiable substeps to an open or cheaper model, and keep ambiguous, high-stakes judgment on frontier models.
Source: every.to