Skip to content

serejaris/personal-corp-os / skills/benchmark

benchmark

Verified

Use when choosing between engines or models for an agent pipeline and the answer must come from measurement on your own data, not from marketing pages: speech recognition for meeting recordings, the model that turns a transcript into notes, the critic model that checks it. Runs every variant in an isolated container, measures time, cost per hour of input (USD and RUB, with price source and date) and quality (WER/CER and course terms for ASR; code checks plus a judge of a different model for LLM steps), writes results.json/csv and one static analytics page. Triggers on "бенчмарк", "сравни движки", "сколько стоит час записи", "какую модель взять для пайплайна", "benchmark the pipeline", "compare ASR engines".

FileKindSize
SKILL.mdskill md7.3 KB
examples/adapter_example.pyscript4.9 KB
prices.example.jsonconfig3.3 KB
scripts/asr.pyscript10.8 KB
scripts/claude_step.pyscript5.8 KB
scripts/common.pyscript2.1 KB
scripts/llm.pyscript4.0 KB
scripts/report.pyscript10.4 KB
scripts/score.pyscript14.6 KB
task.example.jsonconfig1.7 KB

examples/adapter_example.py

View on GitHub
#!/usr/bin/env python3
"""Minimal LLM adapter for llm.py: transcript → notes.md → code checks → judge.

Copy it next to your task and change the three prompts and check_notes().
Input: the ASR run named in task["llm"]["input_asr"] (runs/<name>/asr/<id>/transcript.txt).

    BENCH_SKILL=<skill>/scripts python3 adapter_example.py --task task.json \
        --variant variant.json --out runs/<name>/llm/<id> --stage all
"""

from __future__ import annotations

import argparse
import json
import os
import re
import shutil
import sys
from pathlib import Path

sys.path.insert(0, os.environ["BENCH_SKILL"])
from claude_step import run_step  # noqa: E402

AUTHOR = """Read {src}. Write {dst}: a summary for a participant who missed the meeting.
Sections: `## TL;DR`, `## Decisions`, `## Tasks` (who, what, when, only if said), `## Open questions`.
Every statement must be supported by the transcript. No names of participants. Do not touch other files."""

FIX = """{dst} failed these checks:
{failed}
Fix only what the checks name. Do not touch other files."""

JUDGE = """You check a meeting summary against its transcript. Do not edit files.
Summary: {dst}. Transcript (source of truth): {src}.
Find: unsupported claims, reversed meaning, invented tasks or dates, names of participants, major topics missing.
Severity: critical (false, reversed, invented, private data), major (notable inaccuracy or missing major topic), minor.
Write only JSON to {verdict} and print it:
{{"verdict": "pass" | "fail", "issues": [{{"severity": "...", "claim": "...", "evidence": "..."}}], "summary": "..."}}
fail if any critical or more than two major."""


def check_notes(path: Path) -> list[dict]:
    text = path.read_text(encoding="utf-8") if path.exists() else ""
    checks = [
        ("file exists", bool(text)),
        ("has TL;DR", "## TL;DR" in text),
        ("has Tasks", "## Tasks" in text),
        ("no long dash", "—" not in text),
    ]
    return [{"name": n, "ok": ok} for n, ok in checks]


def main() -> int:
    ap = argparse.ArgumentParser()
    ap.add_argument("--task", required=True, type=Path)
    ap.add_argument("--variant", required=True, type=Path)
    ap.add_argument("--out", required=True, type=Path)
    ap.add_argument("--stage", default="all")
    a = ap.parse_args()
    task = json.loads(a.task.read_text(encoding="utf-8"))
    v = json.loads(a.variant.read_text(encoding="utf-8"))
    work = a.out / "run"
    src, dst, res_path = work / "transcript.txt", work / "notes.md", a.out / "result.json"

    if a.stage in ("all", "author"):
        if work.exists():
            shutil.rmtree(work)
        work.mkdir(parents=True)
        shutil.copyfile(a.out.parent.parent / "asr" / task["llm"]["input_asr"] / "transcript.txt", src)
        steps = []
        s = run_step(AUTHOR.format(src=src, dst=dst), cwd=work, log_dir=work / "logs", name="author", **v["author"])
        steps.append({**s, "role": "author"})
        first = check_notes(dst)
        final, rounds = first, 0
        if not all(c["ok"] for c in first):
            failed = "\n".join(f"- {c['name']}" for c in first if not c["ok"])
            s = run_step(FIX.format(dst=dst, failed=failed), cwd=work, log_dir=work / "logs", name="fix-1", **v["author"])
            steps.append({**s, "role": "fixer"})
            final, rounds = check_notes(dst), 1
        text = dst.read_text(encoding="utf-8") if dst.exists() else ""
        res_path.write_text(json.dumps({
            "steps": steps, "checks_first": first, "checks_final": final, "fix_rounds": rounds,
            "checks_ok": all(c["ok"] for c in final), "judges": [], "errors": [],
            "artifacts": {"notes_kb": round(len(text.encode()) / 1024, 1),
                          "sections": len(re.findall(r"(?m)^## ", text))},
        }, ensure_ascii=False, indent=2), encoding="utf-8")

    if a.stage in ("all", "judge"):
        res = json.loads(res_path.read_text(encoding="utf-8"))
        for j in v.get("judges", []):
            verdict_path = work / f"verdict-{j['provider']}-{j['model']}.json"
            s = run_step(JUDGE.format(dst=dst, src=src, verdict=verdict_path), cwd=work, log_dir=work / "logs",
                         name=f"judge-{j['provider']}-{j['model']}", **j)
            res["steps"].append({**s, "role": "judge"})
            try:
                data = json.loads(verdict_path.read_text(encoding="utf-8"))
            except (OSError, json.JSONDecodeError):
                data = {"verdict": None, "issues": [], "summary": "judge wrote no JSON"}
            sev = [i.get("severity") for i in data.get("issues", [])]
            res["judges"].append({**j, "verdict": data.get("verdict"), "critical": sev.count("critical"),
                                  "major": sev.count("major"), "minor": sev.count("minor"),
                                  "summary": (data.get("summary") or "")[:300]})
        res_path.write_text(json.dumps(res, ensure_ascii=False, indent=2), encoding="utf-8")
    return 0


if __name__ == "__main__":
    raise SystemExit(main())
writing-skillsobra/superpowersUse when creating new skills, editing existing skills, or verifying skills work before deploymentStars: 322kUpdated Claude CodeCodexVerifiedInstall
xlsxanthropics/skillsUse this skill any time a spreadsheet file is the primary input or output. This means any task where the user wants to: open, read, edit, or fix an existing .xlsx, .xlsm, .xltx, .csv, or .tsv file (e.g., adding columns, computing formulas, formatting, charting, cleaning messy data); create a new spreadsheet from scratch or from other data sources; or convert between tabular file formats. Trigger especially when the user references a spreadsheet file by name or path — even casually (like "the xlsx in my downloads") — and wants something done to it or produced from it. Also trigger for cleaning or restructuring messy tabular data files (malformed rows, misplaced headers, junk data) into proper spreadsheets. The deliverable must be a spreadsheet file. Do NOT trigger when the primary deliverable is a Word document, HTML report, standalone Python script, database pipeline, or Google Sheets API integration, even if tabular data is involved.Stars: 198kUpdated Claude CodeCodexVerifiedInstall
shipping-and-launchaddyosmani/agent-skillsPrepares production launches. Use when preparing to deploy to production, or when asking what needs to be in place before shipping. Use when you need a pre-launch checklist, when setting up monitoring, when planning a staged rollout, or when you need a rollback strategy.Stars: 119kUpdated Claude CodeCodexVerifiedInstall
childhoodthedotmack/claude-memBabysit a shipped release after merge, publication, or deployment. Verify published packages, production behavior, installed plugins, and late CI or user reports until the observation window is clean; use babysit for PR readiness before release.Stars: 118kUpdated Claude CodeCodexVerifiedInstall