for
Claude Code and
Codex
Empower every student.
Assess every AI skill.
A coach that helps each student build apps and agents with AI and walk into technical interviews prepared, and a clear read on ten AI skills across your computer science class.
Feed
Updated just nowFull logs added about 12,000 input tokens per debugging session.
Read only the failing test and the relevant error lines.
Keep the useful context. Skip the repeated output.
Define “done” before the first edit.
Add acceptance criteria so your AI can check its own work.Every student in your class uses AI.As a class, you’re still guessing.
Each student prompts their own way, at their own level. The tokens and the misses stay in their own chat.
Maya (Software Engineering)
the checkout test in my final project is flaky. use opus and keep rerunning it until it passes
Reran the suite 14 times.
Wrapped the test in a retry, so it passes now.
$ npm test checkout
✕ checkout total · run 1–13
+ retry(3, () => expect(total).toBe(42))
attempt 3 of 3
checkout total · run 14
This used 1.24M tokens.
Over a million tokens on the most expensive model, and the bug is still there, just hidden behind a retry.
From disconnected AI users to
a connected AI-native class.
As each student’s AI fluency goes up, the links between them get stronger: shared rules, shared skills, one way of working.
Today each student uses AI at a different level, from 3 to 6 out of 10, and the links between them are weak. In an AI-native class, everyone is at 9 or 10 and the links are strong: one shared AI workflow, and the class’s productivity rises with it.
Introducing Vibemetric
One simple app. Every student better at AI, together.
A coach for every student.
For students
Walk into the technical interview ready.
When an interview asks you to build with AI, the interviewer looks at how you work, not just whether the code runs. Vibemetric gives each student an in-depth read on how they build apps and agents with Claude Code and Codex, so they practise exactly that before it counts.
Plan before you code
Planning & Specs · Task Mapping
Settle what done looks like, then split the work into pieces you can check. It’s the first thing an interviewer asks about.
Steer the agent out loud
Human Steering · Context Engineering
Say exactly what’s wrong, ask for evidence, catch wrong assumptions, and give the agent the context it needs.
Prove it works, and keep it safe
Testing & Evals · Security
Run it, test it, and keep keys and personal data out of the chat, instead of taking the AI’s word for it.
Real examples to talk about
Every reason cites a session
“Tell me about a time…” Every point comes from the student’s own projects, with dates, so they have real stories to tell.
Coaching for real habits, not a script to memorize.
The score
Ten AI skills. A score out of 100. Every reason checkable.
The same rubric for every student, applied to the work they actually did with AI.
01
It reads the sessions already on the Mac
Vibemetric reads the student’s Claude Code and Codex history and builds a short digest: how they plan, give context, check results and fix problems. API keys and tokens it recognizes are removed first.
02
The student’s own AI scores it
Their own Claude Code or Codex scores the digest against a fixed rubric: ten skills, 0 to 10 each, added up to 100. It scores habits, not the best session; one great example moves a skill by a point at most.
03
Each skill shows its checks
Every skill has two to four checks, each marked Yes, Partly, Not yet or Not enough evidence, with counts such as 5 of 13 changes checked. Not enough evidence never lowers a score.
04
Every reason cites its sessions
Each reason names the project and date it came from, so the student can open that session and check it. The score says which level they reached and what the next level needs.
Review prompts
Prompts that show how a student worked with AI.
Write a prompt and share it with the class. Each student runs it in their own Claude Code or Codex, on their own laptop, and sends you the answer: how carefully they planned, how much research they did, and whether they checked the AI’s work, with dates and quotes.
1You write the prompt
2The student runs it
Joe (Algorithms)
~/algorithms-hw4
Reading 23 of their own sessions
Read-only: no files changed
3The answer comes back
The sessions never leave the student’s Mac; the student runs the prompt and decides what to send back. Vibemetric does the same across all ten skills and turns it into a score, with the sessions behind every reason.
See your whole class in one place.
Students choose to share their score. You see who’s strong at what, where the class gets stuck, and who can show the others how.
CS project studio
5 students shared a score
Maya
Software Eng.
StudentOverall score 52/100
Plans each feature before asking for code
Accepts fixes without running them
Joe
Algorithms
StudentOverall score 44/100
Keeps a short CLAUDE.md for the project
Lets the AI guess when a test fails
Priya
Web Dev
StudentOverall score 61/100
Asks for evidence before a fix
API keys sit in the code
Sam
Robotics
StudentOverall score 38/100
Tries each robot move in the sim first
Tunes gains without checking the arm’s limits
Alex
Capstone
StudentOverall score 47/100
Writes clear, specific corrections
Starts each session without project notes
- Each student chooses to share. Their scores and one line per area are all that leaves their Mac.
Compared with the usual answer
A fraction of the cost of AI consulting.
| Compared on | AI consulting | Vibemetric |
|---|---|---|
| Cost | $25,000 | Freethe app is free; class pricing on request |
| Time | 60 days | 1 dayhelps from the first score |
| Approach | Top-down | Every studenteach student gets better, every day |
| Over time | Goes stale | Keeps upstudents keep up when the tools change |
AI readiness assessments typically run $10,000–$30,000 in 2026 pricing guides; 60 days is a round figure for an engagement.
Questions
Questions CS professors ask first.
How does Vibemetric work out where each student stands?
It reads the Claude Code and Codex history already on the student’s Mac and builds a short digest of their sessions: how they plan, give context, check results and fix problems. The student’s own Claude Code or Codex then scores the digest against ten AI skills, and every reason points back to the sessions it came from.
Do students’ prompts, code or transcripts go to Vibemetric’s servers?
No. Scoring runs through each student’s own Claude Code or Codex, under their own account, the same as any session they run themselves. Vibemetric’s server never receives sessions, transcripts, prompts or code. It gets only each score’s numbers, to keep the score history.
What do I see, and who decides?
Each student decides. Nothing reaches the class until they choose a score and share it, and then only that score and one line per area leave their Mac. Never their prompts, transcripts or code. They can take a shared score back at any time, and leaving the class team removes everything they shared.
Can I use the score as a grade?
We advise against it. It’s coaching for each student. An AI reviewer writes it, so a skill can move a point or two between runs. Use it to start conversations and to choose what to teach next; every reason cites the sessions it came from, so students can check it.
What do students need?
A Mac (macOS 14+ · Apple Silicon and Intel) and Claude Code or Codex signed in with their own account. Scoring runs on that account, like any other session they run.
What does it cost?
The app is free and open source, and every student can score their own work for free. The class view is part of Pro. For class or school pricing, write to us below.
Can Vibemetric change students’ code or settings?
No. It never writes to ~/.claude or ~/.codex and never changes a transcript. The scoring run can only read files: no shell, no edits, no web.
Every detail is on the privacy page.
Contact
Bringing it to your class? Talk to us.
Questions about setup, privacy or class pricing. We reply by email.
Stop guessing. Go AI native.
Score how each student works, fix one thing at a time, share it with the class, and spread what works.
Free and open source at vibemetric.app · macOS 14+ · Apple Silicon and Intel · 12 MB download