METR · July 2025
19% slower
Experienced developers took 19% longer with AI on their own code. They believed it made them 20% faster.
16 developers, early-2025 tools.
Read the study ↗Let your work with AI develop you, too.
The model gets an upgrade every quarter. When did you?
Agents do more of your work every month. How do you know you're still good?
The way we work with AI keeps changing. Is your judgment keeping up?
How often do you ask the AI whether you're right?
Sense reads your own AI chats, on your Mac or PC, and tells you how you're really doing: where your judgment is sharp, where it slips, and what to practise next. It makes you better, not the AI.
Get Sense
Works in Claude Code, Codex, Antigravity and Claude desktop.
$ claude
Sense · 74 for your last stretch of work, up from 69.
New about you: your first message in a chat is your best. By the fourth, you're typing “continue”.
/korture for the full reading · /improve to practise
>
Film by cottonbro studio / Pexels
Film by ProImageContent / Mixkit
Film by ProImageContent / Mixkit
Film by FrameStock / Mixkit
Film by DC_Studio / Mixkit
The best advice on working with AI keeps changing.
Four people who use AI as hard as anyone, a few months apart. Watch the advice move.
Keep consequential architecture decisions for yourself. Hand over the mundane.
Hand the agents the details. Stay in charge of the spec, the judgment and the taste, and stay in the loop.
On tasks with one right answer, where AI already beats most professionals, check it, then trust a correct result.
Hand a long-running optimisation goal to the agent and let it work for weeks.
Each was right for its moment. Your sense of when to trust, check and stop was formed somewhere else. Is it still right?
- The code you know
- The length of the run
- Agents in parallel
- The tool
- The model
- The hour
- What just happened
- The setup
Why your impression can miss it. Two studies
And from the inside, it's hard to tell which way you're going.
Anthropic · January 2026
17% lower
Learning a new library with AI cost 17% in understanding. Asking it why kept the understanding; handing over the code didn't.
A small first study.
Read the study ↗Paraphrased from the originals, linked. No endorsement implied.
Measure
Know how you're really doing.
One number out of 100 for every stretch of work, read from what you actually typed. It counts what you decided, what you kept for yourself, and how the work felt. Next to it sits your usual, so you can tell a good day from a lucky one.
A reading of the stretch, not a verdict on you.
Your last stretch of work · 16 October, 21:52Example
71
- What you did 60%74What you decided, handed over, checked and corrected
- What you built 20%62Notes, plans and checklists you wrote for yourself, then used
- How you sounded 20%70Calm, stuck or fed up. Anger at the AI's own mistakes never counts against you.
What went well
Asked to see the rollback before running the payments migrationshow me the rollback first
What cost you
Said “try again” four times before deciding it wasn't workingtry again (four times, 22:40)
Your history
20 stretches agonow
Improving. Your last 10 stretches typically read 70. The 10 before that, 62.
Every move, against your usual
- Brief74, usual 70
- Plan69, usual 71
- Hand over80, usual 72
- Steer66, usual 68
- Check78, usual 65
- Correct70, usual 69
- Not sure64, usual 66
- Reuse72, usual 70
- Stop48, usual 67
- Check You asked to see the rollback before anything ran.
- Stop After 9pm, you let wrong turns run three times longer before stopping them.
this stretch your usual
Understand
Who are you when you work with AI?
Not a list of tips. A profile of you, built from your own chats: what you're best at, what you do on autopilot, how you sound when it breaks. Then the patterns no single chat could show you.
Your profile
- Your best move
You say what done looks like.
You reach 100 on Steer more often than on any other move. You name the finished thing, not just the next step.
“make the import idempotent; done when the replay test passes”your best brief this month
- Room to grow
You check the summary, not the work.
On long runs you approve what the agent says it did. Your best checks name one thing to look at, then look.
“great, ship it”47-minute run, 31 files
- Your habit
You start from your own notes.
Half your chats open by pointing the AI at a plan you wrote. Those chats score 9 points higher than the ones that don't.
“read my handover note, then do step 3 only”Monday, 09:12
- How you write
Short, fast, and often asking permission.
Your typical message is 14 words. One in four ends with “ok?” or “right?”.
“use the queue for this, right?”Codex, Tuesday
- How you sound
Calm, until the AI deletes a test.
Calm in 6 messages out of 10. When you're upset, 8 times in 10 it's at the AI's mistake, and that never counts against you.
“you just deleted the test you were meant to fix”Thursday, 16:40
- What you learned
On Monday you asked what a CTE was. By Thursday you were briefing with one.
Sense notices what you ask about and later use yourself. That's you getting better, not the AI.
“write it as a CTE: the dates first, then the joins”Thursday, your own brief
Your patterns
Things you'd never spot from inside one chat. Some say check more. Some say trust it more.
An example, read from the words below.
Over the months
In May you made 7 in 10 of the calls. Now the AI makes most of them.
Sometimes that's the point. The question is whether these are the calls you meant to give away.
Follow this finding into a practice- Calls you made, May70
- Calls you made, October42
“whatever you think is best”October, choosing the database
Try the example: does this read you right?
See 8 more patterns
Your first message is your best. By the fourth, you're typing “continue”.
- Your first message in a chat81
- Every message after it69
“continue”four times in a row, same chat
Try the example: does this read you right?
You check hardest in code you know. In code you don't, you say “go”.
- Check in your own services82
- Check in code new to you52
“looks reasonable, go”payments repo, new to you
Try the example: does this read you right?
On short runs you read the diff. On long ones, you read the summary.
- Diffs you read on 6-minute runs90
- on 40-minute runs20
“great, ship it”47-minute run, 31 files
Try the example: does this read you right?
Your fourth agent gets your least attention, and the most rework.
- Steer in your fourth session41
- in your first74
“continue”five times, session four
Try the example: does this read you right?
When the AI breaks something, you stop checking for the next half hour.
Poker players call it tilt.
- Check, 30 minutes after an AI mistake44
- the rest of the day68
“just fix it”three times, after the failed migration
Try the example: does this read you right?
After 9pm, you let it go round three times longer before you stop it.
- Stop after 9pm48
- before 9pm69
“try again”four times, 22:40
Try the example: does this read you right?
After a bad deploy you grabbed the wheel, and kept it for a week.
- Hand over, the week after38
- the week before72
Trust it more here.
“I'll do this one myself”a migration that was already passing
Try the example: does this read you right?
Since Opus 5.5 you hand over bigger jobs, and say when they're done.
- Hand over since Opus 5.580
- before it64
“move the nightly import to the queue; done when it finishes under ten minutes”Opus 5.5, last week
Try the example: does this read you right?
Improve
Practise the part only you can do.
Your AI can take more and more of the work. It can't decide what good looks like, or know when to stop. Sense picks practices from what your own chats show. You choose many, one or none, and your own AI coaches you on them.
- 01 · Direction
Make your call before you ask
One in four of your messages ends with “ok?” or “right?”. The AI agreed with almost every one.
The AI agrees too easily, and its first guess bends yours. Decide first, then let it argue.
Your AI, as coachBefore it answers a decision, it asks “What's your call?” and waits for yours.
Lesson: Sam Stephenson with Peter Yang ↗ - 02 · Taste
Say why you said yes
You gave a reason for 1 in 5 of your yeses, and 2 in 5 of your noes. Your taste lives in the reasons you never wrote down.
Saying why is how the direction of the work stays yours.
Your AI, as coachWhen you accept its work without a reason, it asks once what made it right.
Lesson: Anne-Laure Le Cunff ↗ - 03 · Understanding
Ask why, then say it back
You asked “why” 300 times last month. You explained the answer back in your own words twice.
Learning happens in your head, not while you read the AI's answer.
Your AI, as coachAfter it builds something that matters, it asks you to explain it back, then tells you what you missed.
Lesson: Justin Sung ↗ - 04 · Knowing when to stop
When it goes round, stop and make the call
Your longest loop: “try again” seven times in 22 minutes. Your best stops come after one.
Every choice you hand away costs you a little of your future judgment.
Your AI, as coachWhen the same fix fails twice, it says so and hands the call back to you.
Lesson: Alex Lieberman on How I AI ↗
- Make your call before you ask
- Say why you said yes
- Ask why, then say it back
- When it goes round, stop and make the call
- Say what done looks like
- Check the work, not the summary
- Brief your fourth agent like your first
- Predict, then look
- Breathe after it breaks something
- Write one by hand each week
- Write down what it taught you
- Save hard calls for your sharp hours
- Trust it where it has earned it
- Decide before you open the chat
Picked for this example from a larger set. Sense picks from what your own chats show, so your practices won't be anyone else's, and they change as you do.
How one practice works
For the finding: the AI now makes most of your calls.
1 · The lesson
2 · Your words
“My call first. In code I don't know, I name the riskiest part of the plan before I say go.”
In your own words. Sense never picks for you.
3 · Your AI coaches
codex
Plan: add invoice_v2, copy the old invoices across, then switch reads. Proceed?
Your practice: make your call first. What's the riskiest part?
youThe copy locks invoices during month-end close. Run it overnight, in batches. Then go.
4 · What changed
Check, in code new to you · last eight stretchesExample
Check went up after you started: 52 before, 70 over the four stretches since.
Good people to learn from. Your work to grow through.
700+ lessons from 40+ sources: the people building the models, and the people using them hardest. Every one credited and linked.
All 18 featured sources
No endorsement implied.
It lives where you already work.
A line when you sit down. A question when it matters. A number in your menu bar or taskbar. Set up by the AI you already use.
Claude Code
Your number, before your first prompt. Coffee optional.
$ claude
Sense · 71 for your last stretch. Improving./korture for the reading · /improve to work on it
Codex
Your AI asks what you think before it tells you what it thinks.
Plan: copy into
invoice_v2, then switch reads. Proceed?Your practice: what's your call first?youIt locks invoices at month-end. Overnight, in batches.
Antigravity
Something new about you. Said once, then left alone.
Sense · new about you: your fourth agent gets your least attention, and the most rework.youFair. This one gets a proper brief.
Claude desktop
Ask for your reading wherever you talk to AI. Then close the lid.
you/korture
70 out of 100. What went well: you stopped the refactor when it stopped helping.Mac menu bar
Four “try again”s in nine minutes. The menu bar noticed.
5822:40Going round in circles: four “try again”s since 22:31.Notification
Arguably the best Stop of the night.
KORTURE.ai Sense now
Your stretch of work finished at 58 out of 100. Going down. Click to see what went well and what did not.
Notification
Your own practice, working.
KORTURE.ai Sense now
Your practice is working: you said why on 3 of your 5 yeses this week. It used to be 1 in 5.
Also in Codex as $korture, in Claude desktop and Cowork as an extension, and in Chrome for claude.ai and chatgpt.com.
Films: Ketut Subiyanto, HIGHTFILM, Mikhail Nilov, Mixkit, arvizum, Ketut Subiyanto, Hanna Pad Originals linked above. Sense and the tools are rebuilt in code; readings are examples.
Most tools make the AI better. Sense makes you better.
Most tools help the agent do better work. None of them asks whether you are getting better.
Your agent's own reportYour agent
Rules, memory and skillsYour agent
AI code reviewThe codebase
KORTURE.ai SenseYou, the expert
Compare Claude Code /insights, code review, analytics and every alternative
What it isWhat it makes betterWho it serves
Your agent's own reportClaude Code /insightsYour setup: rules to add to
CLAUDE.md, features you aren't using yet.Your agentYour AI maker's review of youClaude ReflectFour “AI fluency” skills, described in words, from one app's chats. No number to watch move, nothing to practise.You, inside one app
Rules, memory and skillsCLAUDE.md, AGENTS.md, Cursor rules, agent memoryWhat the agent remembers about you and your code.Your agent
Engineering analyticsDX, now part of Atlassian, LinearB, Jellyfish, GitHub Copilot metricsAI adoption and throughput, counted across the team.Your manager
Courses, newsletters and threadsWhat you know in general. Nothing checks it against your own work, or whether it held.Everyone at once
A mentor, or a sharp colleagueYour judgment, for real. Rare, and rarely there at 22:40.You, when they're free
Your own gutWhat you trust today. In METR's study it read 20% faster for work that took 19% longer.You, unchecked
KORTURE.ai SenseYour judgment. A number that moves, a profile of you and practices you choose, every line checked against your own words. You feel it, and you keep the record.You, the expertKeep every one of them. Sense sits beside them, and it's the only one aimed at you.
Names belong to their owners. No affiliation or endorsement implied.
Why we built it.
We're SK Prasad and Shishir Modur. Not developers, and not experts in this field.
We build our B2B products with agents. Then SK started asking: am I getting better at this, or is it just the agent?
Athletes have trainers. Pilots have logbooks. With agents, nearly everyone is making the human less necessary. We built Sense for the opposite.
Developers feel this first, because agents reached their work first.
SK and Shishir, with our agents
Your next stretch of work.
Your first reading.
Get Sense Research preview: free, on Apple silicon Macs, and on Windows as a test.
Before you install
Sense joins Claude Code, Codex, Antigravity, Claude desktop, Cowork and your browser. It uses the AI you're already signed into, on a share of its allowance you choose. It starts with your newest work, so your first number comes the same day. Release notes
Can I trust what it tells me?
Every sentence Sense shows you is checked against what you actually wrote, by your own AI. If your words don't back a line, you don't see it.
Your work and your privacy
Your record stays on your computer. Reading uses your own signed-in AI. During the research preview we see only whether the harnesses are working or breaking. The final version shares anonymous analytics only. Never sold to your manager.