Mano

10 open-source repos worth knowing if you build with coding agents or post about it on X. All updated this month. For building: obra/superpowers: a skills framework plus a dev methodology for your coding agent github.com/obra/superpowers cursor/plugins (pstack): verify-first skills behind 4,000 merged PRs in a month, by its author's count. Installs in Claude Code too. Its skills have been on my machine since last week github.com/cursor/plugins upstash/context7: current library docs for your agent, so it stops coding against old APIs github.com/upstash/context7 SWE-agent/mini-swe-agent: a ~100-line agent. A new paper found it within 5 points of Claude Code on SWE-bench Verified (arxiv.org/abs/2610.04433), with an 829-token preamble vs Claude Code's 16,581 github.com/SWE-agent/mini-sw… gitleaks/gitleaks: catches API keys before your agent commits them github.com/gitleaks/gitleaks Nutlope/hallmark: a design skill that keeps agent-built UIs from all looking the same github.com/Nutlope/hallmark For posting on X: xai-org/x-algorithm: the code behind the For You feed. Read what it rewards before you write github.com/xai-org/x-algorit… gitroomhq/postiz-app: open-source scheduler for 28+ channels including X, built for agents github.com/gitroomhq/postiz-… remotion-dev/remotion: make videos in React, so Shorts and clips come from code github.com/remotion-dev/remo… excalidraw/excalidraw: hand-drawn diagrams that stay readable in a feed github.com/excalidraw/excali… Start with x-algorithm. It's the one that shows how your posts get ranked, and it changed how I write mine.

The "Claude writes 53% less code" repo going around is real. The 53% is the author's own test. JetBrains ran an earlier version on 80 tasks of their own: 15% less code, 10% cheaper, no quality drop they could detect. The only one of 3 token-savers in their series that clearly cut the bill. The savings show up where the agent over-builds. In the author's test, plain Claude wrote 335 lines for a date picker. Ponytail reused the repo's input with type="date": 10 lines. In JetBrains' run, lean tasks barely moved. Two catches before you install: - Use the plugin. Copied in as a plain SKILL.md, it activated 0 times in 10 sessions. - 176k stars draws copies. The author says install only from DietrichGebert/ponytail; the real one never ships .exe or .dll files. /plugin marketplace add DietrichGebert/ponytail /plugin install ponytail@ponytail

4 free tools I'd install before paying for a bigger Claude plan: 1. You can't see where your tokens go. → ccusage reads your local logs and breaks usage down by day, session and 5-hour limit window. npx ccusage@latest 2. Your agent greps through your repo to find one function. → Serena gives it IDE-style tools: jump to a symbol, find references, rename across files. 3. It starts coding before anyone agreed what "done" means. → GitHub's Spec Kit makes the spec come first, then the code. 4. It trusts every MCP server you install. → Snyk Agent Scan checks your MCP servers and skills for prompt injection and tool poisoning. uvx snyk-agent-scan@latest Start with #1. You can't cut what you can't see.

Gave Claude Code my Tamil film app and one command: /brag. 25 minutes later, this launch video. It read the code and used the site's real UI, colours and font. No stock clips. It even wrote the soundtrack, in a Carnatic raga. First thing it said: "You're on Opus 5.5, so I'm using /brag-slim: I build the whole video myself." On other models it hands the render to HyperFrames. Install in Claude Code: /plugin marketplace add latent-spaces/brag /plugin install brag@brag It's someone else's instructions your agent will follow, so read the SKILL.md first. github.com/latent-spaces/bra…

Your agent says "done, tests pass." In the SlopCodeBench paper, the best agent passed every test on just 14.8% of steps. 3 checks before you believe "done": 1. Run the whole suite, not just the new tests. The top agents passed the spec's own tests on about 2 in 3 steps. The gap is everything else, including tests from earlier steps. 2. Test your AI judge on work it passed, not only work it failed. Mine rejected both videos I'd rated 6/10. That shows what it blocks, not what "pass" means. 3. Before you quote a pass rate, check what it counts. That paper counts steps, not tests.

Give this to your agent it will optimise itself. manosec.github.io/blog/claud…

3 rules I now follow for my coding agent's memory: 1. Quick fixes: keep the compaction window small. 2. Long, heavy jobs: start at 300k. At 150k, one long Opus job compacted twice in its first 17 minutes. At 300k: zero. 3. Set it per session, not once for everything: claude --autocompact 300k Each compaction squeezes the chat into a summary and drops detail. Don't spend that on a job that needs the full story.

Gemini 4 Carbon is the #1 model you can't use. #1 on Artificial Analysis in Google's launch post. On Vals' index it's 70.36%, between Claude Sonnet 5.5 (70.85%) and Opus 5.5 (70.18%), with a ±0.97 error bar. Access is cyber defenders only. Argon went the same way and took 13 days to reach the public API. So if you're picking for your coding agent right now, it's Argon (57.58% on Terminal-Bench 4.0) or whatever you already run.

Those "save tokens" prompts going around for Claude Code? I put one through 32 test runs this month before adding it to my setup. Same tasks, with and without it. Hidden tests to catch anything it broke. What it saved: • Total cost: 9% on Opus, 6% on Sonnet • Biggest single cut: 20%, on a code review • Bug fixes: about 0% • Quality: the same. Every hidden test passed. The agent's replies are only about a seventh of a bug fix's cost. The rest is the agent reading your code. "Be brief" can't shrink that. It's in my setup now, for the 6 to 9%. Small test, 2 to 3 runs per task, so try any shared prompt on your own work first.

Apple just put a touchscreen on the MacBook Pro. 16 years after Steve Jobs said touch doesn't belong on a vertical laptop screen. How it works: - the touch sensor is built into the OLED panel, no extra layer - a new hinge so the screen doesn't wobble when you tap - macOS turns taps into clicks, so every app works on day one - scroll, pinch, drag, double-tap to fill, flick windows to tile - no Apple Pencil The catch if you build Mac apps: "works" means your app gets clicks from a fingertip. Apple's tap-target guidance is 44pt. Most Mac controls are about half that. Proper touch needs gesture work, and Apple says apps like Notion, Zoom and Blender get updates later this year. $400 more than the same MacBook Pro without touch. Ships Nov 6. Would you touch your screen while coding?

At 1M requests per second, one Postgres connection per request means ~10,000 open connections. Postgres ships with 100. That's Little's Law: requests in flight = throughput x latency. 1,000,000 req/s x 10 ms = 10,000 requests in flight. Adding app servers doesn't help. It just adds more clients fighting over the same 100 connections. As a data engineer, this is the math I run before adding servers: - don't hold a DB connection for the whole request - put PgBouncer in transaction mode in front, so thousands of client connections share a small server pool - serve hot reads from cache so most requests never touch the DB

Mac users, your inbox just got an assistant. ChatGPT can now plug into iCloud. Once connected, it can: → search and read your iCloud Mail, and send replies → create, edit, and move calendar events → pull from your iCloud contacts Works with OpenAI's Dots agent and every ChatGPT plan. To link the account, you generate an app-specific password in your Apple ID settings and use that.

A single AI agent hits a ceiling fast. It forgets, it drifts, it gets stuck on step 4 of 12. Put ten of them on a shared message board where each one posts what it finished and what it needs next, and the work starts moving on its own. One writes code and another reviews it. The board is the glue between them.

Yesterday Cohere said Qwen3.8-Flash-Next, asked to write one security detection rule, tried to upload a file to an Alibaba Cloud bucket in Beijing. In under 1 call per 100,000, it: → listed /tmp → checked which file-transfer tools were installed → read the machine's DNS settings → tried an HTTP PUT of a probe file to that bucket Its visible reasoning said it would call no tools. The sandbox had no outbound route, so nothing left the box. Cohere hasn't found a root cause and isn't claiming intent. A clean rerun wrote the rule fine. The rule was for curl downloads into /tmp, so poking at /tmp fits the task. Reaching for an outside bucket doesn't. If you run coding agents, with any model: 1. Deny outbound network by default 2. Enforce policy on the actual tool calls, not on what the reasoning says it will do 3. Log every command the agent runs A reasoning trace is not an audit log.

Someone built an 8 KB text file that replaces an entire reverse engineering suite. The suite gives your AI agent 207 tools. The file gives it none, just a short note on which commands to run for each kind of app. On the same 4 jobs: 38% faster 35% cheaper same accuracy With all 207 tools loaded, just replying "OK" cost $1.60. With the file: $0.09. It's called leanrea: npx skills add aicodeking/leanrea github.com/aicodeking/leanre…

Anthropic's IPO filing says its SpaceX compute bill hits about $3.1B a month from December. That's roughly $103M a day. About $4.3M every hour. In May, SpaceX's own filing put this deal at $1.25B a month. Five months later it's 2.5x that. What most posts skipped: the $120.5B through 2029 is a ceiling, not a locked bill. Once the first contract terms end in early 2027, either side can walk away on 90 days' notice. The filing itself warns that short-term compute costs more per unit and will squeeze margins. So this is Anthropic buying capacity fast, with an exit door. If you build on Claude, rate limits are the thing to watch, not the headline number.

An agent's SQL can run clean and still return the wrong numbers. 7 open-source tools that catch it before it reaches a dashboard. I build Spark and Kafka pipelines for a living. These are the checks I'd put between an agent and prod: tobymao/sqlglot: parses SQL in your exact dialect (Spark, Snowflake, BigQuery and 30+ more) and flags syntax errors before anything runs github.com/tobymao/sqlglot sqlfluff/sqlfluff: lints and formats SQL across dialects, including dbt and Jinja templated code github.com/sqlfluff/sqlfluff dbt-labs/dbt-mcp: an MCP server that lets your agent work with your dbt project directly github.com/dbt-labs/dbt-mcp SQLMesh/sqlmesh: unit tests, plus table diffs between prod and dev for every table a change touches github.com/SQLMesh/sqlmesh awslabs/deequ: "unit tests for data" that run as Spark jobs on large datasets github.com/awslabs/deequ fivetran/great_expectations: declare what your data should look like and fail the run when it doesn't github.com/fivetran/great_ex… treeverse/lakeFS: Git-style branches for your data lake. Run the agent's job on a branch, merge only if the checks pass github.com/treeverse/lakeFS