OP here. I built skillthis.ai, a tool that takes a description of your professional expertise and generates a Claude Code skill file (a markdown prompt file that customizes Claude's behavior for specific tasks).
155 people used it over 3.5 weeks. I analyzed the results and found some patterns I didn't expect.
The headline finding: someone typed "I a bartender" (12 characters, with a typo) and scored 85/100. A 15,576-character technical specification about development process analysis scored 72/100. The bartender input was reproducible, I ran it twice.
More surprisingly, "hey bro" scored 88/100. The system generated a "Casual Communication Skill" and suggested adding "quantifiable success metrics." The grading algorithm clearly has issues (acknowledged in the post).
What actually predicted quality:
- Specific, well-understood domains (plumber, bartender, OKR expert)
- Task-oriented descriptions (what you do vs. what you are)
- Brevity with clarity (top scores averaged under 100 characters)
- Named frameworks or methodologies
What didn't: length (negatively correlated with score), vague enthusiasm, attempts to jailbreak or override Claude's behavior.
The tool uses Claude to generate the skill, then a separate Claude call to grade it. The grading inconsistency is a known problem. I built a guided question flow to address the input quality issue, which asks three follow-up questions when input is too vague.
Stack: Next.js, Supabase, Claude API. Blog post has links to every skill mentioned so you can see the actual outputs.
Author here. mpak is a package manager for MCPB bundles (the standard bundle format for MCP servers).
We built this to power our own platform internally and thought it might be useful to the broader community. Most of it is open source.
The problem: MCP servers are code, not packages. Every server has different install procedures, runtimes, and configs. Beyond the UX annoyance, this creates real issues:
1. Cold starts: pip install or npm install at runtime kills latency. MCPB vendors dependencies, so startup is just unzip and execute (some of ours were taking 60+ seconds to `npm install`)
2. Non-deterministic builds: Same repo, different day, different behavior. Transitive deps shift. Bundles are immutable artifacts, same SHA = same behavior.
3. Security: You can scan a bundle before running it. Generate SBOMs. Run grype/syft/trivy. Know what you're executing before you execute it.
The CLI:
```
mpak search weather
mpak pull @publisher/weather
mpak run @publisher/weather
```
Publishing is a GitHub Action (uses OIDC, no API keys):
- uses: NimbleBrainInc/mcpb-pack@v2
There's also a TypeScript SDK if you're building an MCP client and want to add bundle support programmatically.
Tag a release, bundle gets built and registered automatically.
MCPB is the spec (from the MCP maintainers). We run a public registry at mpak.dev, but the format is what matters. The goal is ecosystem-wide adoption, not lock-in.
"Outcomes not rules" is the whole game. On ambiguity: we don't try to solve it upfront. Get it 80% right, let them refine the rest in seconds. When something's wrong, they fix it in natural language, not by going back to eng to change logic or mess with config. The unlock is really iteration without a ticket.
Would love to swap notes on the accounting side. Similar problems, different domain constraints.
I've been writing Claude Code skills for the past year and kept making the same mistakes: wrong frontmatter format, first-person descriptions, abstract examples instead of concrete input/output pairs.
Good examples exist everywhere, but they are scattered. So I extracted what has worked best for me, combined with Claude's own best practices, and built a generator.
The tool:
1. Takes a plain English description of your expertise
2. Generates a skill following Claude's official format
3. Grades it on a 100-point rubric and gives specific suggestions
Grading criteria includes: format compliance, conciseness (Claude is smart, don't over-explain), Quick Start quality, example concreteness, and completeness.
Tech: Next.js, Claude API, Supabase, v0.
Free, no signup.
This is v1. I'd love feedback on what's working and what needs improvement.
Quick context: I've been running a company (NimbleBrain) using this pattern for ~1 year. The structure is a meta-repo with git submodules, each with a CLAUDE.md file that declares how that domain works.
This isn't intended to be a debate over monorepo vs submodules. Either one works, really. The real value is unlocked when AI can traverse strategy docs, brand voice, and code in one context. Coherence starts emerging without alignment meetings.
The slash commands (/ship-feature, /extract-knowledge) are where this gets practical.
Happy to answer questions. Curious if anyone else is running something similar.
Been using hamilton for a few months for orchestrating AI pipelines. Super lightweight and easy to use. Visualizations in dagworks are super helpful. Highly recommended!
I've had the same issue with Slack. It's incredibly difficult to find anything older than a few days. That's why I built ThreadScribe (https://www.ThreadScribe.ai). It's a AI-powered chat companion for Slack that automatically summarizes your threads. Would love to get some feedback from you on this.
155 people used it over 3.5 weeks. I analyzed the results and found some patterns I didn't expect.
The headline finding: someone typed "I a bartender" (12 characters, with a typo) and scored 85/100. A 15,576-character technical specification about development process analysis scored 72/100. The bartender input was reproducible, I ran it twice.
More surprisingly, "hey bro" scored 88/100. The system generated a "Casual Communication Skill" and suggested adding "quantifiable success metrics." The grading algorithm clearly has issues (acknowledged in the post).
What actually predicted quality: - Specific, well-understood domains (plumber, bartender, OKR expert) - Task-oriented descriptions (what you do vs. what you are) - Brevity with clarity (top scores averaged under 100 characters) - Named frameworks or methodologies
What didn't: length (negatively correlated with score), vague enthusiasm, attempts to jailbreak or override Claude's behavior.
The tool uses Claude to generate the skill, then a separate Claude call to grade it. The grading inconsistency is a known problem. I built a guided question flow to address the input quality issue, which asks three follow-up questions when input is too vague.
Stack: Next.js, Supabase, Claude API. Blog post has links to every skill mentioned so you can see the actual outputs.