返回资讯中心
外部精选
软件工程
#开发者实时流

Vibemetric – An open-source app that grades your Claude Code sessions

Faster builds, safer code, smaller token bills. Vibemetric shows you exactly what to change.

Hacker News 最新kisangpak2026/10/17 · 2026/10/1724 分钟阅读

以下正文同步自 Hacker News 最新,版权归原站所有,已转换为易读排版。

Faster builds, safer code, smaller token bills. Vibemetric shows you exactly what to change.

ShareDownload for MacView on GitHub

46Mon 9:41

Problem: The AI changed code 23 times last week and never ran your tests once — you caught three of the breakages yourself, a day later.

Fix: Make the agent run the tests after every change

Add this to CLAUDE.md under “Definition of done”: After every code change, run npm test and show me the output. If a test fails, fix the cause and run the tests again before you continue. For UI changes, also open the page and describe what you see. Never say a task is done until every test passes, and list the commands you ran in your final message.

71Tue 2:18

Problem: You retyped the same rule — “use pnpm, never npm” — in 9 separate sessions, and it still reached for npm twice.

Add a “Project rules” section to CLAUDE.md with this rule: Use pnpm, never npm or yarn, for installs, scripts and lockfiles. Before you run any package command, check this section. If you find a rule that I repeat in my prompts, suggest adding it here instead of asking me again. Keep each rule to one line, with the reason after a dash.

58Wed 11:05

Problem: Four multi-file tasks started with no plan and no definition of done. Two were rebuilt from scratch once you saw the result.

Fix: Start big tasks with a plan and pass/fail checks

Before you write any code, start in plan mode. Write a short plan: the files you will change, the steps in order, and the risks. Then list the pass/fail checks that prove the task is done, such as tests, commands and what I should see in the app. Wait for my OK on the plan and the checks. After the work, run every check and report each result.

83Thu 4:52

Problem: Renames, typo fixes and formatting all ran on your most expensive model — roughly a third of last week’s spend on work a cheaper one handles.

Add this to CLAUDE.md under “Model routing”: Use Sonnet for renames, typo fixes, formatting and one-line changes. Use Opus only for design decisions, debugging across several files, and security work. At the start of each task, say which model fits and why. If the task grows, stop and suggest switching before you continue.

Running the same agent on the same task varied in cost by up to 30×. Models consistently underestimate their own token spend, so the agent cannot warn you either.

Across 623 million code changes, duplication rose while refactoring fell 70% against 2022. Code that was moved — the sign of a real refactor — dropped from 21% to 3.8%.

The average security pass rate across models sits at 56%, barely moved from 55% in the first report. Log injection passes 12% of the time, cross-site scripting 15%.

In a randomized trial, 16 experienced developers took 19% longer on 246 real issues in repos they already knew. Afterwards they still believed AI had sped them up by 20%.

Today 9:12am · from 42 sessions · Score 46 / 100 ↗ +6 since Oct 6

Your agent changed code 31 times this week but ran the tests after only 8, so two “finished” tasks broke the next day.

Make the agent run the tests after every change

After every code change, run npm test and show me the result. Don’t say a task is done until the tests pass.

Small edits run on your most expensive model

Failed checks are often skipped instead of fixed

The agent often runs with every permission on

Keep doing thisYou build real tools and reuse them

Your CLAUDE.md files help, but two are out of date

Do you agree on the goal and the checks before work starts?

You plan big tasks, but small ones start with no checks

Does each project tell the AI what it must know?

Your CLAUDE.md files help, but two are out of date

Does the AI remember what it learned, correctly?

What we saw“Use pnpm, never npm” appeared in 9 prompts.

Why it mattersRepeated rules cost time, and one was forgotten.

Try thisSave repeated rules in CLAUDE.md once.

Download for Mac

How Vibemetric works

Vibemetric scans your AI coding sessions

It reads the Claude Code and Codex sessions already on your Mac. Nothing to upload, nothing to fill in.

Oct 8 · 4:51 PMFix flaky checkout testCodex

Oct 7 · 11:02 AMAdd SSO to loginClaude Code

Write two onboarding email templates: a welcome email and a trial-ending reminder.

Make them warmer, and add a P.S. with a 20% discount code.

Rewrote both in a friendlier tone and added a P.S. with WELCOME20.

Ten skills, each scored from 0 to 10 on what you actually did, with the sessions as evidence.

One fix to make first, ready to copy into your project. Then a new score every day, so you can see it work.

Route each folder to a subagent that fits it

copywriterHaiku 4.5

src/builderSonnet 5.5

infra/ migrations/architectFable 5.1

Add these files to .claude/agents/


name: copywriter description: Emails and docs in emails/ and docs/ model: haiku

Write and edit copy. Keep the tone warm and plain. Never touch code.

Copy

PRIVACY BY DEFAULT

The scanning happens on your Mac. Scoring runs through your own Claude Code or Codex, under your account. Vibemetric’s server never sees your sessions.

Planning & Specs

Do you and the AI agree on the outcome, how it’s checked, and the constraints before work starts?

Context Engineering

Does each project have a concise, well-structured CLAUDE.md or AGENTS.md that covers what AI must know, kept current?

Memory & Knowledge

Is AI memory kept accurate, so stale or mixed-up memories don’t cause mistakes?

Tool Integration

Does your recurring work run through tools and skills by default, and do you share the proven ones?

Orchestration I: Task Mapping

Is the work broken into the right pieces and wired together: order, inputs and outputs, and who does each?

Orchestration II: Feedback Loops

Do results decide the next step (fix, recheck, stop, escalate) and does the loop improve from experience?

Testing & Evals

Is checking a habit: tests that fit each project, real-flow checks of changes, and regular evals for AI features?

Human Steering

Do you examine the AI’s work and redirect it with clear reasons, then check the effect?

Security

Is the agent limited to what each task needs, and is what it builds checked for security problems?

Efficiency & Cost

Do models, effort, and usage match each task’s difficulty, adjusted from actual results?

You give AI real tools it keeps reusing, and it checks live behavior on its own. But you rarely say what “done” means up front, so design tasks end in rounds of taste corrections.

A new score every day.

Vibemetric scores your latest sessions once a day and keeps the runs side by side, so you can see whether a change you made actually helped. The score sits in your menu bar.

You give AI real tools it keeps reusing, and it checks live behavior on its own. Specs are the last gap: design tasks still end in a round of taste corrections.

Sep 30: 41 / 100Oct 1: 44 / 100Oct 2: 50 / 100Oct 3: 48 / 100Oct 4: 56 / 100Oct 5: 61 / 100Oct 6: 66 / 100Oct 7: 64 / 100Oct 8: 74 / 100Oct 9: 82 / 10082Sep 30Oct 9

PRIVACY FIRST

Your sessions are read on your Mac and scored by your own Claude Code or Codex, under your own account. Here’s exactly what happens to them.

No Vibemetric server sees them

Scoring runs through your own Claude Code or Codex, exactly as when you use it yourself. The free app has no account, and Vibemetric’s server never receives your sessions. Pro only sends your API key to check your plan.

Read-only, all the way down

Vibemetric never writes to ~/.claude or ~/.codex. The Claude run it starts can only Read, Grep and Glob. No shell, no edits, no web.

Secrets are stripped first

Before anything is read, the digest removes API keys, tokens, webhook URLs and database connection strings it recognizes.

You decide what stays

Each result keeps its redacted digest on your Mac until you delete the result. Your CLI saves the scoring run in its history, but Vibemetric leaves it out of later scores, so it can’t skew them.

The full score, on your Mac. Open source.

AI Native Score out of 100

All 10 dimensions, each with the case that earned it

Recommendations for every dimension

Score history with what changed since last time

Menu bar score, Markdown and image export

Download for Mac

Pro

$9.99per month

Everything in Free, plus tools and progress from our server.

Specific tool picks for each dimension, with setup steps

A curated tool catalog, updated every week

How you compare with other Vibemetric users

Progress over time, and whether each change worked

Team view: see your team’s scores, gaps, and trends

Start 7-day free trialThen $9.99 per month. Cancel anytime.

The app and the full score are free. Pro is $9.99 a month or $99 a year, with a 7-day free trial, and adds tool picks, comparisons, progress tracking and a team view. Scoring itself runs on your existing Claude plan through the claude command you already have installed, like any other Claude Code session.

The rubric only counts connected evidence: a rule saved in one session and used in a later one, or a failed check that was fixed and rerun. Finding those takes reading, not counting.

Scores come from an AI reviewer, so they can move a point or two between runs. Reasons always cite the sessions they come from, so you can check them.

macOS asks once about any app downloaded from the internet. Vibemetric is signed and notarized by Apple, so click Open. Updates install automatically after that.

Open the .dmg, drag Vibemetric into Applications, and open it. Needs Claude Code or Codex installed and signed in.

macOS 14+ · Apple Silicon and Intel · 5 MB download

Download for Mac

VibemetricA small Mac app that scores how you work with AI, from the sessions already on your disk.

Vibemetric 0.3.5 · macOS 14+ · Apple Silicon and Intel

正文由 探How 从来源站点 RSS 同步,内容未经改写;遇到排版缺失或需要图片、视频时请以原文为准。

打开原文
去社区讨论 →