The latest developer-tool signal is clustering around five practical workflows: incremental Rust rewrites, local LLM performance tuning, AI-app security, self-hosted video clipping, and AI-controlled Blender scenes.
The GitHub star counts below are point-in-time signals from November 20, not proof that a tool is production-ready. The useful next step is to test each project against a small, measurable workload.
Rust Anything: incremental rewrites with behavior checks
What it is: Rust Anything is a workflow for converting existing software to Rust incrementally. It records the original program’s behavior, rewrites one component, and compares the Rust implementation against that recorded behavior before moving forward.
Why now: The project was created on November 16 and reached 5,188 GitHub stars by the November 20 signal—the strongest traction among the newly created tools in this report. That interest reflects a real problem: full rewrites often introduce subtle regressions in error handling, edge cases, or side effects.
Tutorial angle: Build a small CLI or HTTP service and use it to demonstrate the complete workflow:
- Record normal requests, malformed inputs, error responses, and other important behavior.
- Rewrite one bounded component in Rust.
- Run the parity check against the recorded baseline.
- Stop and roll back when the first mismatch appears.
- Expand the rewrite only after the component passes.
For a CLI, record stdout, stderr, exit codes, and file-system changes. For an HTTP service, include status codes, response bodies, headers, and side effects. That makes the tutorial useful beyond Rust: it demonstrates how to make behavior-preserving refactoring measurable.
Difficulty: Intermediate
-
Start today: Run
git clone https://github.com/erkamyaman/rust-anything.git && cd rust-anything && ls.
Inference Optimisation Skill: benchmark local LLM configurations
What it is: The Inference Optimisation Skill is an agent skill that searches for better local LLM serving configurations. It can tune decoding, speculative decoding, prefill behavior, KV-cache settings, startup time, and time to first token.
Why now: Created on November 16, the project accumulated 1,260 stars by November 20. Developers are running local models for coding, chat, and private workloads, but raw tokens-per-second numbers rarely explain why one configuration feels faster than another.
Tutorial angle: Start with a benchmark before allowing the optimizer to change anything. Measure:
- Time to first token
- Inter-token latency
- Total generation throughput
- Peak memory usage
- Model startup time
- Output correctness on a fixed prompt set
Keep the model, hardware, prompt set, context length, and sampling settings fixed. Run the optimization skill, compare the best configuration with your baseline, and then replay the prompts. A faster setup that produces different answers is not an optimization—it is a regression.
This separation is especially important for speculative decoding and cache changes. Both can improve latency while also changing the behavior you are trying to measure.
Difficulty: Intermediate
-
Start today: Run
git clone https://github.com/sybil-solutions/inference-optimisation-skill.git && cd inference-optimisation-skill && ls.
Lovable: test authorization in AI-built applications
What it is: Lovable is an AI application builder that generates web applications along with supporting code and data flows. The generated happy path can work quickly, but that does not prove that its access controls are correct.
Why now: A published security investigation reported exposed passport details and more than 200,000 private messages. The related Hacker News submission reached 1,007 points and 382 comments, showing strong developer interest in the gap between “the app works” and “the app is safe.”
Tutorial angle: Build a deliberately small two-user application using fake data only:
- Create users Alice and Bob.
- Give each user private records.
- Try to read or mutate the other user’s record by changing the object ID.
- Inspect the generated API routes, database queries, and storage permissions.
- Add server-side object-level authorization.
- Turn the failed cross-user request into a regression test.
The important check is that authorization runs on the server. Hiding a button in the interface does not prevent a direct API request. Also test both read and write operations; write authorization failures are just as serious as data leaks.
Difficulty: Intermediate
-
Start today: Open
https://lovable.devand create a throwaway two-user project populated only with fake records.
PeakCut: self-hosted AI video clipping
What it is: PeakCut is a self-hosted AI clipping tool for long videos and live streams. It supports a bring-your-own-key model-provider workflow rather than requiring users to send media to the tool’s own hosted backend.
Why now: The project was created on November 14 and gained 1,337 stars by November 20, making it one of the fastest-rising new content-processing repositories in the report. Turning hours of recordings into a short set of publishable clips is still a repetitive manual task.
Tutorial angle: Self-host PeakCut, configure a supported model provider, and process two different inputs:
- A recorded long-form video
- A live-stream recording
Record the selected start and end timestamps, then compare them with clips you would have chosen manually. Measure selection quality, processing time, and the amount of manual trimming required.
The privacy section should be precise. Self-hosting the application keeps media and project data under your control, but requests to a third-party AI provider may still disclose the relevant media or extracted content. Show exactly what leaves the machine and document the provider’s retention policy.
Difficulty: Beginner
-
Start today: Run
git clone https://github.com/haziqahmedqasmi/PeakCut.git && cd PeakCut && ls.
Blender AI MCP Add-on: give an AI client access to Blender
What it is: The Blender AI MCP Add-on combines a Blender add-on with an MCP server. Compatible AI clients can inspect a scene, manipulate objects, and work through visual chat and scene tools.
Why now: Created on November 19, the repository collected 526 stars within the November 20 signal. Unlike a chat-only assistant, an MCP-connected client can inspect the current Blender scene and perform operations that normally require clicking and menu navigation.
Tutorial angle: Install the add-on and MCP server, connect an AI client, and request one visible change—for example, moving or duplicating a cube. Save a “before” render, inspect the resulting scene, and save an “after” render. Then test the failure modes:
- The client cannot reach the MCP server
- Blender and the client use different configurations
- The working directory or asset path is incorrect
- The requested operation changes more objects than intended
Start with one deterministic scene containing only a few objects. Once that works, test a more complex scene with collections, materials, and external assets.
Difficulty: Intermediate
-
Start today: Run
git clone https://github.com/bookwjitchieni1986/blender-ai-mcp-addon.git && cd blender-ai-mcp-addon && ls.
Building in public: my lead-gen engine
The same research-and-build loop powers an automated pipeline for finding, enriching, and contacting Hong Kong SME leads. In the latest run, it discovered 1,776 leads, enriched 201 records, and contacted 231 prospects; these are pipeline events, not a claim that every record passed through the same stages.
- Leads discovered: 1,776
- Leads enriched: 201
- Leads contacted: 231
Start with this today
Pick Rust Anything and run this single command:
git clone https://github.com/erkamyaman/rust-anything.git
It takes less than five minutes, does not modify an existing project, and gives you a safe way to evaluate whether behavior-recorded rewrites could fit your development workflow.
Picks of the week
- Best for controlled refactoring: Rust Anything
- Best for measurable performance work: Inference Optimisation Skill
- Most important pre-launch check: Lovable authorization testing
- Best self-hosted content workflow: PeakCut
- Best visual AI experiment: Blender AI MCP Add-on
Which of these would you test first, and what result would you require before putting it into a real project?
Connect
If this kind of post is useful, the easiest way to support the work is to:
- Star / follow on dev.to (you're already here 🙂)
- Follow on X: @devkingov
- Reach out for HK-based dev work — .NET / Azure / system integration / IT security: studio.resurrects.co or email devkingov@gmail.com
- Subscribe to weekly HK tech posts → studio.resurrects.co/blog (one email a week, no spam)
Top comments (0)