Sebastian's Benchmarked Skills

Reusable AI agent skills, each with a regression eval track record. That track record is the main artifact, so start there before cloning or running anything.

The 2026 World Cup has wrapped, but the eval track record lives on and is still worth a look.

Or browse the same history locally after cloning:

runme eval view

Install world-cup-picks-report globally from this repo:

npx skills add -g https://github.com/sourishkrout/skills --skill world-cup-picks-report

Or add it through the Claude Code marketplace:

/plugin marketplace add sourishkrout/skills
/plugin install world-cup-picks-report@sourishkrout-skills

Everything below is for running, promoting, and deploying the evals locally.

Runme's eval workflow is documented at https://docs.runme.dev/eval/.

Current skill:

• world-cup-picks-report in skills/world-cup-picks-report

The regression verifier uses Anthropic-backed LLM judges. Export an Anthropic API key before running the evals:

export ANTHROPIC_API_KEY=...

Run the full regression eval from the repo root:

runme eval skills/world-cup-picks-report/evals/regression \
    --agent codex \
    --ak reasoning_effort=xhigh

To run only the oracle:

runme eval skills/world-cup-picks-report/evals/regression

You can substitute another supported agent, such as cursor-cli, claude-code, or openclaw, for codex.

After running an eval, compare the latest local job against the latest Git-tracked baseline:

runme eval compare skills/world-cup-picks-report/evals/regression

Preview the promotion before committing evidence:

runme eval promote skills/world-cup-picks-report/evals/regression --latest --dry-run

If the result should become the new baseline, stage the related source changes and promote the eval evidence:

git add <changed-files>
runme eval promote skills/world-cup-picks-report/evals/regression --latest

Promotion records compact eval evidence by default. Use --artifacts only when you need full logs and trial outputs; artifacts can contain sensitive information. Use --evidence-only when promoting eval evidence without source changes.

Review artifacts before promoting, because the Harbor showcase is public. Push the promotion commit to main and GitHub Actions publishes the fresh results.