Nicole Forsgren19 October 2025

How to measure AI developer productivity in 2025

8Frameworks

15Insights

Episode overview

Lenny Rachitsky speaks with Nicole Forsgren about measuring AI's effect on developer productivity without relying on easily gamed output metrics such as lines of code. Forsgren explains how DORA and SPACE should be applied in AI-enabled workflows, emphasizing system-level velocity, satisfaction, feedback loops, trust, code quality, and direct developer research. She also presents the seven-step process from Frictionless for launching and scaling a developer-experience initiative, from listening and securing a quick win through measuring value and iterating.

Key ideas

Lines of code is an especially unreliable productivity metric when generative AI can produce verbose code and comments on demand.
AI accelerates code creation but also shifts developer effort toward reviewing generated code, validating reliability, and managing multiple agents.
DORA remains useful for assessing delivery-pipeline speed and stability when used for its prescribed purpose, while SPACE helps teams choose context-specific measures.
Teams should begin developer-experience work by interviewing developers, mapping workflows, and identifying recurring friction before choosing tools or automation.
Useful warning signs include broken builds, flaky tests, cumbersome processes, slow environment provisioning, and high costs when switching projects or teams.
AI impact should be measured against the outcome leadership values, such as idea-to-customer speed, profit margin, market competitiveness, or reduced vendor and cloud costs.
A short, carefully designed satisfaction-and-barriers survey can establish a baseline when reliable system instrumentation does not yet exist.
Developer-experience improvements should be managed like a product: define users and problems, test an MVP, gather feedback, communicate a strategy, and retire obsolete tools or metrics.

Transcript available · source text is retained privately and is not published

Frameworks in this episode

Leadership5 steps

Friction Smells: Signs Your Team Can Move Faster

The tell-tale signals that friction, not capability, is capping your team's speed

Read →

Productivity3 steps

Prescriptive vs. Framework Metrics

Know whether your metric is a recipe or a lens — and never misuse a recipe

Read →

Productivity4 steps

Prime-Then-Parallelize AI Coding Workflow

Front-load the architecture, fan out to agents, then step back and evaluate

Read →

Communication5 steps

Reframe Metrics to Leadership's Language

Pick two or three metrics that speak the exact word your leaders keep repeating

Read →

Leadership4 steps

The DevX Listening Tour

Before you build a tool, walk developers through their yesterday

Read →

Leadership7 steps

The Frictionless Seven-Step DevX Program

A start-anywhere playbook to stand up a developer-experience initiative

Read →

Productivity4 steps

The SPACE Framework

Pick balanced developer productivity metrics across five dimensions — never rely on one.

Read →

Productivity5 steps

The Top-Three Weighted DevX Survey

Force a top-three, weight by frequency, and never ask four questions at once

Read →

People & resources mentioned

Attributed to the moment in the episode. Timestamps are approximate.

People · 11

Nicole ForsgrenMentionsNicole is the developer productivity expert having written the award-winning book accelerate
“my guest is Nicole Forsrren.”
▶ Watch at ~1:00approx.
Nicole ForsgrenMentionsNicole is the developer productivity expert having written the award-winning book accelerate
“Nicole has been at the forefront of this space longer than anyone.”
▶ Watch at ~1:30approx.
Gloria MarkMentionsGloria Mark has done some really good work on attention and deep work
“Gloria Mark has done some really good work on attention and deep work”
▶ Watch at ~18:30approx.
Andrej KarpathyMentionsI remember Karpathy tweeted that he just like has never seen a model like this.
“I think it was Karpathy that shared he was like so stuck on a bug”
▶ Watch at ~35:30approx.
Abi NodaMentionsI also wrote this with Abby Nota who has just uh of DX.
“Abby started this company called DX”
▶ Watch at ~37:30approx.
Dan ShipperMentionsDan is the co-founder and CEO of Every which is a company that is at the very bleeding edge of what is possible with AI.
“I had a Dan Shipper on the podcast”
▶ Watch at ~58:30approx.
Peter AttiaCoinedPeter T and all these guys are always saying. This is the thing you want to track
“Outlive by uh Peter Aia is fantastic.”
▶ Watch at ~1:02:30approx.
Stuart McGillCoinedBack mechanic by Stuart McIll is incredible.
“Back mechanic by Stuart McIll is incredible.”
▶ Watch at ~1:02:30approx.
Michael LewisCoinedThe Undoing Project by Michael Lewis.
“The Undoing Project by Michael Lewis.”
▶ Watch at ~1:03:30approx.
Matt VelosRecommendsMatt Velos recommended it to me and it's so good.
“Matt Velos recommended it to me and it's so good.”
▶ Watch at ~1:03:30approx.
Lenny RachitskyMentionstoday we've got another very special compilation episode something I've been pulling on more and more with the podcast and the newsletter
“Thanks for having me, Lenny.”
▶ Watch at ~1:07:00approx.

Resources · 30

SPACECoinedother · Nicole Forsgren and collaborators
“She created the most used frameworks for measuring developer experience called Dora and Space.”
▶ Watch at ~1:30approx.
DORACoinedother · Nicole Forsgren and collaborators
“She created the most used frameworks for measuring developer experience called Dora and Space.”
▶ Watch at ~1:30approx.
FrictionlessCoinedbook · Nicole Forsgren and Abi Noda
“is about to publish her newest book called Frictionless”
▶ Watch at ~1:30approx.
AccelerateCoinedbook · Nicole Forsgren, Jez Humble, and Gene Kim
“She wrote the most important book in the space called Accelerate”
▶ Watch at ~1:30approx.
MercuryUsescompany · Mercury
“I've been banking with Mercury for years”
▶ Watch at ~3:00approx.
WorkOSMentionscompany · WorkOS, Inc.
“Work OS turns those deal blockers into drop-in APIs with a modern developer platform built specifically for B2B SAS.”
▶ Watch at ~4:00approx.
DORAMentionsother · Nicole Forsgren and collaborators
“I have done Dora metrics their speed metrics their stability metrics”
▶ Watch at ~14:00approx.
SPACEMentionsother · Nicole Forsgren and collaborators
“space is a framework. It doesn't tell you what metric to use.”
▶ Watch at ~15:30approx.
CodaUsessoftware · Coda
“I personally use KOD every single day to manage my podcast and also to manage my community.”
▶ Watch at ~32:00approx.
OpenAI CodexMentionssoftware · OpenAI
“Something I've seen people on Twitter sharing is how good uh OpenAI Codeex especially is at finding really gnarly bugs.”
▶ Watch at ~35:30approx.
AtlassianMentionscompany
“They just sold the company for a billion dollars to Atlassian.”
▶ Watch at ~37:30approx.
DXCoinedsoftware · DX
“Abby started this company called DX”
▶ Watch at ~37:30approx.
JiraMentionssoftware · Atlassian
“they have all these companies working using Jara and all their products.”
▶ Watch at ~38:00approx.
developerexperiencebook.comMentionswebsite · Nicole Forsgren and Abi Noda
“Yeah, developerexperiencebook.com.”
▶ Watch at ~42:00approx.
GitHub CopilotMentionssoftware · GitHub
“Like there's the common ones. Copilot cursor.”
▶ Watch at ~58:00approx.
Claude CodeRecommendssoftware · Anthropic
“I love claw code.”
▶ Watch at ~58:00approx.
CursorMentionssoftware · Anysphere
“Like there's the common ones. Copilot cursor.”
▶ Watch at ~58:00approx.
Claude CodeMentionssoftware · Anthropic
“he said clock code is the most underrated AI tool out there”
▶ Watch at ~58:30approx.
ChatGPTUsessoftware · OpenAI
“I've really been loving Chad GPT and Gemini especially to render pictures for me”
▶ Watch at ~1:01:00approx.
GeminiUsessoftware · Google
“I've really been loving Chad GPT and Gemini especially to render pictures for me”
▶ Watch at ~1:01:00approx.
Back MechanicRecommendsbook · Stuart McGill
“Back mechanic by Stuart McIll is incredible.”
▶ Watch at ~1:02:30approx.
OutliveRecommendsbook · Peter Attia with Bill Gifford
“Outlive by uh Peter Aia is fantastic.”
▶ Watch at ~1:02:30approx.
How Big Things Get DoneRecommendsbook · Bent Flyvbjerg and Dan Gardner
“I will say I love how big things get done.”
▶ Watch at ~1:03:00approx.
Love Is BlindUsestv show
“I will say I watched Love is Blind if I got to like shut down at the end of the day.”
▶ Watch at ~1:03:30approx.
The Undoing ProjectRecommendsbook · Michael Lewis
“The Undoing Project by Michael Lewis. Uh Matt Velos recommended it to me and it's so good.”
▶ Watch at ~1:03:30approx.
ShrinkingRecommendstv show
“And Shrinking. Shrinking. Have you seen Shrinking?”
▶ Watch at ~1:04:00approx.
Ninja CREAMiRecommendsproduct · Ninja
“The Ninja Creamy is.”
▶ Watch at ~1:04:00approx.
JURA coffee machineUsesproduct · JURA
“Another one is a Jura coffee maker. I'd love good coffee and I'm not great at making it.”
▶ Watch at ~1:04:30approx.
GoogleMentionscompany · Google
“you share that you have a new role at Google.”
▶ Watch at ~1:05:30approx.
Lenny's PodcastMentionspodcast · Lenny Rachitsky
“You can find all past episodes or learn more about the show at lennispodcast.com.”
▶ Watch at ~1:07:30approx.

Spot an error or want something removed? Request a correction or removal.

Insights & moments

The myth-busts, hot takes, explainers, and tools worth keeping.

Myth Buster· 2

Myth Buster12:00

"Most Productivity Metrics Are A Lie" — Why Lines Of Code Is Broken

Nicole is blunt that most productivity metrics are a lie, and lines of code has always been bad — now catastrophically so. Because LLMs are verbose by definition, you can prompt for the longest possible code, gaming the metric while injecting complexity and technical debt.

Lines of code was always a poor proxy for output or complexity
AI blows the metric out of the water because agents and LLMs are verbose
Gaming it introduces complexity and technical debt across your work
It becomes more useful only if you can separate human-written from AI-generated code

“I will say most productivity metrics are a lie.”
Nicole Forsgren · 12:00

“if the goal is more lines of code, I can prompt something to write you the longest piece of code ever and add tons of…”
Nicole Forsgren · 12:30

#metrics#lines-of-code#measurement

Myth Buster56:30

Why Nicole Doesn't Trust Happiness Surveys For Engineers

Nicole pushes back on happiness surveys: happiness is driven by too many things — work, family, hobbies, weekends — to be a useful engineering signal. Satisfaction is different and measurable: you can ask whether someone is satisfied with a specific tool and follow up. Happy devs do write better code, but you influence that through satisfaction, not by measuring happiness directly.

Happiness is too all-encompassing to be a useful metric (work, family, hobbies, weekends)
Satisfaction is narrower and answerable: 'Are you satisfied with this tool?'
Satisfaction and happiness are related but not the same
Happy devs write better code, but target satisfaction to get there

“I don't love a happiness survey because there are too many things that contribute to happiness.”
Nicole Forsgren · 57:00

“happy devs make happy code.”
Nicole Forsgren · 57:30

#surveys#satisfaction#measurement

Hot Take· 1

Hot Take29:00

"Faster For What?" — You Can Ship Trash Faster Every Day

Nicole's counterpoint to speed obsession: you can always move faster, but faster for what? Without strategy and smart decisions about what to ship, teams just ship trash faster every single day. Speed only compounds value when the strategy underneath it is right — garbage in, garbage out.

You can always move faster, but the real question is 'faster for what?'
Without strategy you can ship trash faster every single day
Strategy is the core piece; speed multiplies whatever decisions feed it
This is where PMs matter — deciding what to build and in what order

“you can always move faster, but faster for what?”
Nicole Forsgren · 29:00

“we can ship trash faster every single day.”
Nicole Forsgren · 29:00

#strategy#speed#product

Explainer· 6

Explainer07:00

What "DevX" Actually Means And Why It Gates Everything Else

Nicole defines developer experience (DevX) as what it's like to build software day-to-day: the friction developers face, the workflows they navigate, and the support they have. Her core point is that DevX is foundational: when it's poor, no amount of great tooling or process compensates.

DevX = the day-to-day reality of building software (friction, workflows, support)
When DevX is bad, the best processes and tools won't save you
Productivity sits inside DevX, alongside engineering happiness
Output achieved through high toil or high cognitive load eventually burns developers out

“when DevX is poor, everything else just isn't going to help, right? The best processes, the best tools, the best whatever magic you have, right?”
Nicole Forsgren · 07:00

#devx#developer-experience#productivity

Explainer09:00

How AI Coding Breaks — And Can Rebuild — Developer Flow State

Nicole explains that writing code with AI is now interrupt-driven: instead of locking in and typing for hours, you prompt, get code back, and review it. That fragments flow, but a well-built tool chain can also hand flow-recovery off to the machine, which reminds you of context and regenerates system diagrams.

The three reinforcing components of DevX: flow state, cognitive load, feedback loops
You no longer lock in and type large blocks of code the way you used to
Prompt-review-integrate cycles constantly interrupt flow
AI can also restore flow by reminding you of context and generating diagrams

“you don't just get in the flow and lock down and write a whole bunch of code and do like the typing of a whole…”
Nicole Forsgren · 09:30

#flow-state#ai-coding#cognitive-load

Explainer14:00

Why Dora And Space Can't Be Applied Blindly In The AI Era

Nicole, who created Dora and Space, warns her own frameworks can't be applied blindly now. Dora's prescriptive speed and stability metrics still assess the pipeline, but AI has changed feedback loops so they run much earlier and throughout. Space, being a framework rather than a prescription, actually adapts well to new AI contexts.

Prescriptive metrics like Dora must be used only the way they were prescribed
AI creates feedback loops much earlier and throughout the pipeline, not just at build/test
Space is a framework, not a prescription, so it transfers to emerging AI contexts
Nicole would add a 'trust' dimension since LLMs are nondeterministic

“we can't just blindly apply the exist existing metrics we've used before because we'll miss super important phenomenon and and changes in the way people…”
Nicole Forsgren · 14:00

#dora#space#metrics

Explainer18:30

Four Hours Of Deep Work — And Why 45-Minute Blocks Now Count

Citing Gloria Mark's research, Nicole notes humans get roughly four hours of good deep work a day. Deep work historically needed long uninterrupted chunks (two hours minimum) to reach flow. With AI handing flow-recovery partly to the machine, she asks whether a 45-minute block can now become productive too.

People get about four hours of good deep work per day, per Gloria Mark
Reaching flow historically needed ~two-hour minimum uninterrupted blocks
AI coding is inherently interrupt-driven now
Because the machine helps you re-enter flow, shorter blocks may become useful

“Cubans can get about four hours of good deep work a day and like that's about it.”
Nicole Forsgren · 18:30

“does that mean that now we can also make a 45minute work block useful?”
Nicole Forsgren · 20:00

#deep-work#flow-state#time-management

Explainer27:00

The Smells That Say Your Engineering Team Could Move Faster

Nicole says most teams can move faster and she doesn't know a single team near the diminishing-returns ceiling. The tells: constant build breakages, flaky tests, overly long processes, painful environment provisioning, and — a strong smell — people refusing to switch teams because the switching cost and friction are so high.

Most teams can move faster; almost none are near the point of diminishing returns
Listen for builds breaking, flaky tests, and overly long processes
Hard-to-provision environments and painful task switching are friction signals
Some companies made switching orgs cost the same 'tax' as a brand-new hire

“Most teams can move faster, right?”
Nicole Forsgren · 27:00

“if you're always hearing about bills breaking, flaky tests, uh, overly long processes”
Nicole Forsgren · 27:30

#engineering-velocity#friction#diagnostics

Explainer33:30

AI's Productivity Gains Are Real — But We Can't Measure Them Yet

Nicole says the gains are real but we lack good measures. The strongest signal is end-to-end velocity from idea to experiment, though it's hard to attribute to a specific tool or developer. One study she saw found AI generated significantly more code for regular users — and the engineers' own output increase was double what the coding agent contributed, suggesting AI mainly unblocks work you'd already do.

Gains are real but there are no great measures yet
Best signal is velocity through the whole system, idea to experiment
Attribution to a specific AI tool or developer is hard
One finding: engineers' output increase was double the coding agent's own contribution
AI's main effect may be unblocking work you'd already do

“I'd say it's real. And I would also say we don't have great measures for it yet.”
Nicole Forsgren · 33:30

“AI was generating like significantly more code for the people who were using it regularly.”
Nicole Forsgren · 34:30

#ai-productivity#measurement#velocity

Story· 2

Story10:30

The Senior Engineer Who Orchestrates Parallel AI Agents Like Production Code

Nicole describes senior engineers she's watched set up an incredible workflow: they prime the AI with the architecture, stack, and conventions up front, have it design the plan, then assign an agent to each piece to work in parallel. Because they were systematic early, the output lands much closer to production code than something vibe-coded.

Prime the model with intended architecture, stack, and workflow before building
Let the AI design it, then assign an agent per piece to run in parallel
Specify up front that the pieces must work together with correct APIs and conventions
Front-loaded systematic planning yields near-production code, not vibe-coded output

“And then for each piece, it'll assign an agent to go work on each piece in parallel.”
Nicole Forsgren · 11:00

“they're much closer to something that looks like production code”
Nicole Forsgren · 11:30

#ai-agents#workflow#parallelization

Story23:30

The Mainframe Fix That Was Just Sending An Email

Nicole recounts a company that avoided an ancient, hated mainframe process because they assumed fixing it meant a full replatform. In reality, someone had to physically print a document, walk it down several flights for approval, then walk it back up. The fix wasn't a rebuild — they just had it send an email.

The team dodged the process because they thought it required replatforming
The actual bottleneck was manual: print, walk down floors, get approval, walk back up
The fix was changing the process, not redesigning anything major
Often the highest-impact improvement is a low-lift process change

“Sometimes all you have to do is change a process”
Nicole Forsgren · 24:00

“They just had it send an email.”
Nicole Forsgren · 24:30

#process-improvement#friction#case-study

Tool· 1

Tool58:00

Claude Code As A "Little Mini God" For Non-Engineering Tasks

Nicole and Lenny riff on Claude Code beyond coding. Nicole loves it and has a post coming on non-engineering use cases. Lenny describes using it to clean up laptop storage — like ChatGPT running on your computer — and quotes Dan Shipper calling it the most underrated AI tool because people don't realize it's not just for coding.

Claude Code works well for non-engineering tasks, not just coding
Example: asking it to find and clean up storage on your laptop
It's like ChatGPT running directly on your computer with system access
Dan Shipper called it the most underrated AI tool out there

“I love claw code.”
Nicole Forsgren · 58:00

“clock code is the most underrated AI tool out there because people don't realize what it's capable of.”
Lenny Rachitsky · 59:00

#claude-code#ai-tools#productivity

Takeaway· 3

Takeaway22:30

The One Thing To Do This Week: Talk To Developers, Don't Build A Tool

Asked for one tactical move a team can make this week, Nicole says go talk to people and listen — start with listening, not tools and automation. Teams too often build a tool for a problem they personally had; instead, ask developers to walk through yesterday and name where they got frustrated or slowed down.

Start with listening, not tools and automation
Companies over-build tools for problems that are just easy to automate
Ask developers to walk through yesterday: what was delightful, what caused friction
A handful of conversations surfaces low-lift, high-impact fixes

“Honestly, I think the best thing you can do is go talk to people and listen.”
Nicole Forsgren · 22:30

“And I would say start with listening and not with tools and automation.”
Nicole Forsgren · 22:30

#listening-tour#developer-experience#tactics

Takeaway49:00

To Measure AI Impact, Start With What Leadership Actually Cares About

Nicole's advice for measuring AI's productivity impact: it depends on what your leadership chain cares about. Listen to the recurring theme — market share, profit margin, velocity, transformation — then pick two or three metrics that map to it, and reframe your work in their language rather than making them work to understand it.

Start from the theme leadership keeps repeating
Market share → measure speed from feature to production/customer/experiment
Profit margin → measure money saved, then translate to recouped headcount or vendor spend
Reframe in their words; don't make them work to understand the value

“it depends on what your leadership chain really cares about, right?”
Nicole Forsgren · 49:30

“We don't want to make them work to understand what it is that we're doing and the value that we provide.”
Nicole Forsgren · 51:00

#measurement#leadership#communication

Takeaway55:30

Don't Ask Four Questions At Once — And Let Claude Draft Your Survey

Nicole warns that a question like "Were the build and test systems slow or complicated last week?" secretly asks four things, making the signal impossible to untangle. She recommends working with someone who knows survey design — or having a couple of rounds with Claude, Gemini, or ChatGPT to refine questions — and a rule of thumb: if you can't answer a question with the data, don't collect it.

Compound survey questions bundle multiple variables and muddy the signal
Talk to someone familiar with survey design
Use Claude, Gemini, or ChatGPT across a couple of rounds to refine questions
If you can't answer a question with the data, don't collect it

“Were the build and test system slow or complicated in the last week?”
Nicole Forsgren · 55:30

“If you can't answer a question with data, don't get it.”
Nicole Forsgren · 56:00

#survey-design#ai-tools#data

Key ideas

Frameworks in this episode

Friction Smells: Signs Your Team Can Move Faster

Prescriptive vs. Framework Metrics

Prime-Then-Parallelize AI Coding Workflow

Reframe Metrics to Leadership's Language

The DevX Listening Tour

The Frictionless Seven-Step DevX Program

The SPACE Framework

The Top-Three Weighted DevX Survey

People & resources mentioned

People · 11

Resources · 30

Insights & moments

⚡Myth Buster· 2

"Most Productivity Metrics Are A Lie" — Why Lines Of Code Is Broken

Why Nicole Doesn't Trust Happiness Surveys For Engineers

◆Hot Take· 1

"Faster For What?" — You Can Ship Trash Faster Every Day

✶Explainer· 6

What "DevX" Actually Means And Why It Gates Everything Else

How AI Coding Breaks — And Can Rebuild — Developer Flow State

Why Dora And Space Can't Be Applied Blindly In The AI Era

Four Hours Of Deep Work — And Why 45-Minute Blocks Now Count

The Smells That Say Your Engineering Team Could Move Faster

AI's Productivity Gains Are Real — But We Can't Measure Them Yet

❝Story· 2

The Senior Engineer Who Orchestrates Parallel AI Agents Like Production Code

The Mainframe Fix That Was Just Sending An Email

✦Tool· 1

Claude Code As A "Little Mini God" For Non-Engineering Tasks

▲Takeaway· 3

The One Thing To Do This Week: Talk To Developers, Don't Build A Tool

To Measure AI Impact, Start With What Leadership Actually Cares About

Don't Ask Four Questions At Once — And Let Claude Draft Your Survey

Myth Buster· 2

Hot Take· 1

Explainer· 6

Story· 2

Tool· 1

Takeaway· 3