Docs
Created by Ali Alhaddad, founder of Qamar Labs.
Run the app
Needs Bun 1.1+ (or Node 20+ with npm).
bun install bun run dev # http://localhost:8080 bunx vitest run # grading + assessment tests
Run the C# tests
Needs the .NET 8 SDK. Paste your solved files over the exported ones.
bun run export:csharp cd csharp-export/tests/EventBus.Tests dotnet test
Connect Claude
- Publish the current app version. Publishing does not save local drafts or create report cards.
- Sign in to skill/bench. In Skill editor, grade a skill and click Save report card. Your account-saved snapshots appear in Report cards.
- In Claude, open Settings → Connectors → Add custom connector. Name it skill/bench and enter
https://skillbench.lovable.app/mcp. - Follow Claude’s authorization prompts, sign in to the same skill/bench account that saved the report, and approve read access.
- Enable the connector in a Claude conversation if needed, then ask: “List my saved skill/bench report cards.” Approve tool use if Claude asks.
Connector availability depends on your Claude account and organization settings. The connector reads only your saved reports; it cannot read browser-only drafts or edit skills. Revoke access in Claude’s connector settings.
Create a report-card artifact
- Open a saved card from Report cards, or use the latest grade in the Skill editor.
- Under Report card format, select Claude artifact · connection required and click Copy artifact request. This saves a new account snapshot and copies its request.
- Paste the request into a Claude conversation with the connector enabled and send it. Allow Claude to read the report if asked.
- The report stays in skill/bench’s Report cards page. Claude creates the artifact in that conversation’s artifact panel, not inside skill/bench.
A copied request is not a connection. If Claude cannot see a card, check the published version, connector authorization and account used to save it. HTML and Markdown downloads do not require Claude.
IDE
- Open IDE. Existing Markdown editor text becomes your first document without deleting the original browser draft.
- Use New project to choose Markdown, .NET, Python, Node.js, FastMCP, Gradio or a practice template. Download the current project before replacing it.
- Select a file, edit the code, and use Preview for Markdown or script-disabled HTML. The Explorer holds your files, tabs hold open documents, and the Output panel shows execution results.
- Use New file, Rename, Import, Download file or Export project. Exported
.ide.jsonfiles preserve the file paths and can be imported again. Drafts save automatically in this browser, not your Cloud account. - Run executes the selected Python, JavaScript or TypeScript file in isolation. Relative project imports work. Test all executes your .test.js and .test.ts files with synchronous test, expect and assert helpers. Network access, Node packages, host filesystem access and background servers are unavailable.
- Python uses a real interpreter with the standard library, project imports and a fresh in-memory filesystem for every run. Print output, errors and tracebacks appear in Output. Use Stop to cancel; Python runs are limited to five seconds and 65,536 characters of console text. Interactive input, third-party packages, network access and host files are unavailable. The first run downloads the bundled interpreter.
Open Agent workflow for a generator → verifier → coach starter with bounded loops and editable tests. Self improvement is available under Review. C# needs an external runtime; MCP inspection and Lambda fixtures remain simulations. The original five-problem C# assessment remains on its own page.
Blog
Read AI-efficiency articles on measuring outcomes, focused context and bounded agent reviews. Maintainers add or edit entries in src/lib/blog/articles.ts; there is no separate publishing account or article editor.
Where to edit
| Rubric scoring | src/lib/skill/rubric.ts |
| Custom binary rubrics + reports | src/lib/review/model.ts |
| HTML extraction + report formats | src/lib/review/formats.ts |
| Agent review prompts | src/routes/api.improve.ts |
| Workspace saving | src/lib/review/workspace.functions.ts |
| Description-based drafts | src/lib/skill/drafting.ts |
| OWASP security rubric + safeguards | src/lib/skill/security.ts |
| Anthropic standards, intellisense, lint | src/lib/skill/standards.ts |
| Sample skill | src/lib/skill/samples.ts |
| C# starter project | src/lib/assessment/project.ts |
| Assessment problems + checks | src/lib/assessment/problems.ts |
| Suggested fixes for weak skills | src/lib/skill/fixes.ts |
| Branding / trademark | src/lib/brand.ts |
| Server API | src/lib/api.functions.ts |
| Pages | src/routes/*.tsx |
| Colors and fonts | src/styles.css |
| IDE templates and starter files | src/lib/ide/templates.ts, templates-data.ts |
| IDE validation and simulations | src/lib/ide/project.ts, practice.ts, sandbox.ts |
| Python runtime and isolation | src/lib/ide/python.ts, python-worker.js |
| IDE workspace and panels | src/components/ide/ |
| Blog articles | src/lib/blog/articles.ts |
Notes
- In the skill editor, type
skill:or press Ctrl+Space for snippets; hover frontmatter keys for the standards. - Editor drafts and assessment progress stay in your browser. Sign in to save private skill snapshots, custom rubrics, report cards and completed improvement sessions.
- Manual reviews use equal-weight True/False indicators with evidence. Automatic grading uses the default quality rubric; security remains a separate document-coverage assessment.
- Self improvement runs a generator, verifier and devil’s-advocate coach. No commands run and proposals require approval before applying.
- HTML skills are experimental static-text reviews, not Anthropic-standard skill packages. HTML and Markdown report exports are available; Claude creates artifacts after you connect your account.
- Assessment checks in the browser read your code;
dotnet testis the real compile.
Verification and remaining issues
Integration checks passed: 77 tests, signed-in report saving and export, IDE editing and persistence, project import/export, Blog pages, and phone layouts. The preview Claude connection listed and read a temporary account’s saved report; approval and denial worked.
- The published Claude address did not serve this revision during verification. Republish, then connect Claude using the same account that saved your report. Your own saved reports and an actual Claude conversation artifact still need that final check.
- IDE projects stay in this browser. Export backups before switching projects or browsers. Python, JavaScript and TypeScript execute for real in isolation; C# needs an external runtime. MCP inspection and Lambda fixtures remain simulations. There is no remote machine, dependency installation, deployment or AWS invocation.
- Fresh self-improvement model calls and GA4 reporting were not reverified during this integration. The codebase’s docs/VERIFICATION.md contains the full checks, limitations and remaining actions for maintainers.