iiinova — where your digital self keeps evolving

Samuel Berthe @samber.bsky.social · 28/05/2026

Europe is as bad at fighting heat-related deaths as the United States is at preventing gun deaths. hannahritchie.substack.com/p/heat-guns-...

000

Samuel Berthe @samber.bsky.social · 20/04/2026

Dans 6 mois, la France déploie la facture électronique. Just saying...

000

Samuel Berthe @samber.bsky.social · 04/04/2026

I wonder why OpenAI isn’t mentioned more in the race for proactive LLM agents. They acquired the core developer of Openclaw but never outlined a clear, dedicated strategy, and never announced a dedicated subscription. We’ve been used to seeing these labs move faster.

000

Samuel Berthe @samber.bsky.social · 04/04/2026

The Claude Code fork in Python has just been renamed to "claw-code". This is the ultimate troll. 😈 github.com/ultraworkers... During that time, there was a huge retaliation from Anthropic: they banned customers using their Claude sub in Openclaw.

media.tenor.com

a man in a striped shirt is sitting on a couch looking at a laptop .

ALT: a man in a striped shirt is sitting on a couch looking at a laptop .

010

Samuel Berthe @samber.bsky.social · 31/03/2026

Feature request for AI agents: Add support for skills groups, we could enable/disable on demand. ☠️ Adding too many skills breaks the context. ☠️ Enabling a bunch of skills for a 5-minute task is painful.

100

Samuel Berthe @samber.bsky.social · 24/03/2026

In 2027, Brazil and California will require age verification + KYC at the OS level. Including Linux. 30 years ago, when hackers (GNU, FSF...) built the internet and fought for digital freedom, this would have sparked a global rebellion. What the f*** is going on ??

020

Samuel Berthe @samber.bsky.social · 24/03/2026

Some skills are very factual and straightforward (code style, naming...). But others need more high-level thinking about methodology (troubleshooting, safety...).

110

Samuel Berthe @samber.bsky.social · 24/03/2026

Yep, some skills are very close and might be loaded at the same time (eg: observability+benchmark+troubleshooting+performance). I had to update 100% of skills during review. I'm not able to tell you how much % remains at the end.

110

Samuel Berthe @samber.bsky.social · 23/03/2026

BTW, some markdown files are just long list of interfaces and API descriptions: - github.com/samber/cc-sk... - github.com/samber/cc-sk... This is not where your attention goes.

000

Samuel Berthe @samber.bsky.social · 23/03/2026

Yep. I would estimate it took around 8–10 days to pull all the skills together into a production-grade collection I'm comfortable sharing publicly.

100

Samuel Berthe @samber.bsky.social · 23/03/2026

Your SKILL.md should describe the methodology, mention tools, list a few common mistakes, and give references to secondary markdown files that will be loaded lazily Bonus: write some external references If the description is too short, you can add a complementary "When to use" section at the top.

000

Samuel Berthe @samber.bsky.social · 23/03/2026

My Go skill set: github.com/samber/cc-skills-golang npx skills add samber/cc-skills-golang Open to contributions. If you're building skills for other languages/frameworks, fork the structure.

github.com

GitHub - samber/cc-skills-golang: 🧑‍🎨 A collection of Golang agentic skills that works

🧑‍🎨 A collection of Golang agentic skills that works - samber/cc-skills-golang

051

Samuel Berthe @samber.bsky.social · 23/03/2026

300k tokens of skills later, the hard part isn't writing instructions. It's writing instructions that are small enough to load, specific enough to trigger, and clear enough that an AI agent does the right thing without you watching.

110

Samuel Berthe @samber.bsky.social · 23/03/2026

And finally: Test, test, test. Iterate. Your skills won't work from the first shot. And review everything MANUALLY!

100

Samuel Berthe @samber.bsky.social · 23/03/2026

Add extras only when needed: Bash(benchstat:*) for perf, WebFetch for docs, Bash(govulncheck:*) for security. Minimal tool surface = smaller attack vector = cleaner Snyk scans.

100

Samuel Berthe @samber.bsky.social · 23/03/2026

Declare your allowed-tools explicitly. Default set for Go skills: Read Edit Write Glob Grep Bash(go:*) Bash(golangci-lint:*) Bash(git:*) Agent

100

Samuel Berthe @samber.bsky.social · 23/03/2026

Teach reasoning, not just rules. Bad: "NEVER use math/rand for crypto" Good: "math/rand output is predictable; an attacker can reproduce the sequence. Use crypto/rand."

110

Samuel Berthe @samber.bsky.social · 23/03/2026

Worst format: walls of prose. The agent will miss the signal.

100

Samuel Berthe @samber.bsky.social · 23/03/2026

Format for scanning, not reading. Agents don't read your skill like a human. They scan for structure. Best formats: - Good/Bad code examples with inline "why" - Diagnose: lines pointing to CLI tools - Imperative prose - Numbered RFC-style (MUST/MAY/SHOULD) for strict rules

100

Samuel Berthe @samber.bsky.social · 23/03/2026

Snyk found that 36% of ClawHub skills contained prompt injection patterns. Their scanner is now integrated into skills.sh as a gate on install. Don't be the skill author who gets flagged. Scan locally first.

skills.sh

The Agent Skills Directory

Discover and install skills for AI agents.

100

Samuel Berthe @samber.bsky.social · 23/03/2026

skills.sh

The Agent Skills Directory

Discover and install skills for AI agents.

100

Samuel Berthe @samber.bsky.social · 23/03/2026

Run Snyk Agent Scanner before publishing: snyk-agent-scan --skills ./skills/ It detects prompt injection, tool poisoning, toxic flows, credential leaks, and malware payloads.

100

Samuel Berthe @samber.bsky.social · 23/03/2026

Target: ~10 assertions per 1,000 tokens of skill content. Minimum 50. Then review all assertions manually 🥵 Store results in the repo. Track with/without score delta over time.

100

Samuel Berthe @samber.bsky.social · 23/03/2026

Eval methodology that works: First run: human-as-judge (you review outputs manually) Follow-up runs: LLM-as-judge (automated, faster iteration) Clear your agent session before eval. Load ONLY skill-creator. Ask Claude to be adversarial and NOT load other skills

200

Samuel Berthe @samber.bsky.social · 23/03/2026

Evaluations must be adversarial. A good eval has a "trap" the model falls into WITHOUT the skill but avoids WITH it. Test the skill's UNIQUE value. The subtle tradeoffs, non-obvious choices, the specific gotchas the model would miss by default.

100

Samuel Berthe @samber.bsky.social · 23/03/2026

If your skill uses a CLI, add a cli-reference.md Agents discover tool capabilities through examples, not --help output. Create references/cli-reference.md with every useful subcommand, flag combo, and realistic example. Generous examples > terse documentation.

100

Samuel Berthe @samber.bsky.social · 23/03/2026

WARNING: don't write explicit MCP instructions in the SKILL.md body (e.g. "call resolve-library-id"). Snyk Agent Scanner flags this as prompt injection. Instead: "This skill is not exhaustive. Refer to library documentation. Context7 can help as a discoverability platform."

100

Samuel Berthe @samber.bsky.social · 23/03/2026

Point the agent to Context7 for up-to-date library docs. For library-specific skills, add Context7 tools to "allowed-tools" and recommend fetching docs.

100

Samuel Berthe @samber.bsky.social · 23/03/2026

When two skills claim the same trigger keywords, the agent loads the wrong one. Add boundary disclaimers: "Not for measurement methodology → See samber/cc-skills-golang@golang-benchmark"

100

Samuel Berthe @samber.bsky.social · 23/03/2026

Write atomic skills. Cross-reference, don't duplicate. Each concept lives in exactly ONE skill. Everything else uses → See owner/repo@skill-name. Example: performance optimization vs benchmarking vs profiling = 3 separate skills with clear boundaries.

100

Samuel Berthe @samber.bsky.social · 23/03/2026

Never parallelize the write in a single file. Background agents: non-blocking checks while main agent keeps coding Up to 5 parallel sub-agents. Explicitly describe how to split in your SKILL.md.

110

Samuel Berthe @samber.bsky.social · 23/03/2026

Teach your skill to use sub-agents. For audit/review modes covering a full codebase, instruct the agent to split work: Split by concern: Agent 1 handles injection, Agent 2 handles auth, Agent 3 handles crypto Split by scope: Agent 1 covers pkg/, Agent 2 covers internal/

100

Samuel Berthe @samber.bsky.social · 23/03/2026

If deep analysis benefits from extended thinking, say **Thinking mode:** Use ultrathink for... If a company-specific skill should override yours, declare it: "This skill may be superseded by your internal X skill"

100

Samuel Berthe @samber.bsky.social · 23/03/2026

The top of your SKILL.md body is prime real estate. Use it for directives the agent needs immediately: If the skill has multiple modes (coding vs reviewing vs auditing), explain the difference up front

110

Samuel Berthe @samber.bsky.social · 23/03/2026

Set "user_invocable: false" when the skill is contextual. Most skills should auto-trigger based on context, not require a /slash-command. Setting it to true pollutes the command namespace. Only use it for skills the user explicitly invokes (like /commit or /security-audit).

100

Samuel Berthe @samber.bsky.social · 23/03/2026

Be specific. Be slightly "pushy." Under-triggering is worse than over-triggering because you never notice it.

100

Samuel Berthe @samber.bsky.social · 23/03/2026

Your description IS your trigger. Get it wrong, skill never fires. Bad: "Implements X in Golang using library/foo" Good: "Implements X in Golang using library/foo — feature A, B, C. Apply when using or adopting library/foo, or when the codebase imports github.com/library/foo."

100

Samuel Berthe @samber.bsky.social · 23/03/2026

Progressive disclosure is how you stay lean without losing depth. Put CLI docs, code examples, config templates in references/ and assets/. The agent reads them lazily. Think of it like a function: small signature, deep implementation behind a pointer.

100

Samuel Berthe @samber.bsky.social · 23/03/2026

The #1 mistake: fat skills. Your token budget: Description: ~100 tokens (loaded at startup for ALL skills) SKILL.md body: < 2,500 tokens (loaded on activation) Full directory with refs: < 10,000 tokens Blow the budget and you degrade EVERY response. Measure with: tiktoken-cli

330

Samuel Berthe @samber.bsky.social · 23/03/2026

Use the skill-creator skill to bootstrap. Don't start from a blank SKILL.md. The skill-creator walks you through intent capture, draft, eval design, and iteration. It's the meta-skill: a skill for building skills. npx skills add vercel-labs/agent-skills --skill skill-creator

220

Samuel Berthe @samber.bsky.social · 23/03/2026

Before you write anything: check if it already exists. Visit skills.sh or skillsmp.com. There are 500k+ indexed skills now. If nothing fits, THEN create your own. Don't reinvent what's already built and battle-tested.

skills.sh

The Agent Skills Directory

Discover and install skills for AI agents.

100

Samuel Berthe @samber.bsky.social · 23/03/2026

I wrote 300k+ tokens of agent skills for Claude Code with 100% manual reviews. Here's everything I learned about writing skills that actually trigger, stay lean, and don't get flagged by security scanners. A thread on best practices for building agent skills in 2026. 🧵

441

Samuel Berthe @samber.bsky.social · 19/03/2026

Days since a CTO proudly said their team codes with ChatGPT + the $10 Copilot plan: 0 Please don't.

000

Samuel Berthe @samber.bsky.social · 17/03/2026

- Go, Python, Ruby - SystemD, process-exporter, IPMI, Proxmox VE, Keycloak - AWS, Azure, DigitalOcean, GCP, OpenStack - eBPF, Cilium, Envoy, SNMP, WireGuard, certmanager - GitLabCI, Spinnaker - Jaeger, Grafana Tempo, Grafana Mimir - Apache Flink, Apache Spark, Memcached, OracleDB

000

Samuel Berthe @samber.bsky.social · 17/03/2026

🚨📣 Massive update on samber/awesome-prometheus-alerts project with 30 new exporters and 343 new rules distilled from the community. 🎉 samber.github.io/awesome-prom... #buildInPublic

132

Samuel Berthe @samber.bsky.social · 15/03/2026

Hey @datadoghq.com Feature request: Give me a datadog-cli that: - exports any indicator to json/csv - draws graphs in ascii art - and publish a Claude Code skill with strong how-to Congrats, you just built an autopilot debugger. 🤪

011

Samuel Berthe @samber.bsky.social · 13/03/2026

Claude Code Skills tip: If your skill needs a broad understanding of the codebase (migrations, refactoring, architecture review), have it spawn parallel sub-agents via the Agent tool, each targeting a distinct scope (packages, file patterns, concerns). Cuts research time dramatically on large repo.

000

Samuel Berthe @samber.bsky.social · 11/03/2026

🤑🤑 Business idea for 2026: -> Sell keyboards with 4 big buttons: 1. Yes 2. Yes, allow all edits during this session (shift+tab) 3. No + 1 button for mic.

000