1 of 24

Deck Club · November 6, 2026

PREPARE FOR THE

BEAR!

Welcome to Startup Grind Princeton. Tonight I want to talk about what I think is the most important thing happening in the world right now — and what it means for all of us in this room. The title is Prepare for the bear! That's not hyperbole. The bear is real, it's moving fast, and it doesn't care how smart you are. Let's get into it.��I likely used Claude, ChatGPT, Perplexity, and other AI tools for various parts of this presentation.

***

Two campers and a bear.��Two campers are walking through the woods when a huge brown bear suddenly appears in the clearing in front of them. The bear sees the campers and begins to head towards them.

The first camper drops their backpack, digs out a pair of joggers, and frantically begins to put them on.

The second camper says: “What are you doing? Joggers won’t help you outrun that bear.”

“I don’t need to outrun the bear,” the first replies. “I just need to outrun you.”

Image by chatGPT

2 of 24

2

sg

DECK CLUB · NOVEMBER 6, 2026

Spec-Driven Development

How ethical AI raises productivity, and how to measure it

Thanks, Gael Lambert, Princeton Digital Laboratory!

THE PREMISE

10–20%

of a project's value is the code

Sean Grove (OpenAI), as cited in the deck

WELCOME (about 2 minutes)�Good morning and welcome to Deck Club at the NJ AI Hub. Thanks to the Hub team, and Jordan, for hosting us again.��The norms: this is a discussion, not a lecture. Disagreement is welcome, kindly. Speak from your own seat: founders, engineers, investors, lawyers and people at big companies all see this differently, and that's the value of the room. Keep turns short so more voices get in. And you do not need to have read the deck. I'll give a ten-minute overview first, and most of today's questions are about your own experience.��This month's deck is local. It's "Specification-Driven Development: How Ethical AI Raises Productivity" by Gael Lambert of the Princeton Digital Laboratory, an independent research group here in Princeton. Gael gave it as a talk to the PMI New Jersey chapter, the project managers' association. It's licensed CC BY 4.0, so we can show and share it freely with credit.��The number on screen is the deck's starting provocation: Sean Grove of OpenAI argues the code is perhaps 10 to 20 percent of a project's value. The rest is the specification, the structured statement of what you intend to build. If that's true, AI writing the code matters less than who writes the spec.

3 of 24

3

sg

THE HOUR

One hour, five moves

Didn't read it? You're in the right place.

1

Ten-minute overview

2

Opening hand vote

3

Four discussion threads

4

Open floor

5

Re-vote, then what's next

4 of 24

4

sg

THE THESIS

People write and agree the specification first. AI produces against it.

The value is in the spec, not the code.

Lambert, slide 8

5 of 24

5

sg

WHO WROTE IT, AND HOW

A project manager's view of AI

THE AUTHOR

Gael Lambert, PMP since 2010

25 years shipping software, incl. Accenture and IBM

THE METHOD

A talk to PMI New Jersey

Draws on a 138-source mapping study

Independent analysis, unaffiliated with the vendors named. Lambert, slides 2 and 19

OVERVIEW, PART 2: WHO AND HOW (about 1 minute)�The author is Gael Lambert, a researcher and founding member of the Princeton Digital Laboratory, an independent research organization based here in Princeton that studies how the physical world gets modeled digitally. Per the deck, Gael has been a PMP since 2010, was educated at Oxford and Harvard, and has spent 25 years delivering software around the world for firms like Accenture and IBM.��The method: this is a talk deck for the PMI New Jersey chapter, so the audience was project managers, not engineers. It draws on Gael's systematic mapping study of SDD, 138 sources, published on Zenodo as version 0.2. The lab has also released its AI agents, called Shirley, under the MIT license.��Two things worth flagging neutrally. First, the effort splits for agile, SDD and vibe coding are labeled illustrative, not measured. Second, the deck closes with hypotheses, not results. It's a framework and a research agenda more than a study of outcomes.

6 of 24

6

sg

THE STRUCTURE

The deck in four parts

1

Methods, from waterfall to vibe coding

2

What SDD is, and who's behind it

3

Productivity: P = V·Q / T

4

Hypotheses and what to measure

7 of 24

7

sg

WHERE WE STARTED

Waterfall: half the effort was testing

BROOKS, 1975

1/6

of the budget went to coding

Source: Lambert, slide 4, after Brooks, The Mythical Man-Month (1975)

8 of 24

8

sg

FOUR WAYS TO RUN A PROJECT

Same five activities, shifting weight

WATERFALL

Long cycle, test-heavy

AGILE

Short cycles, repeated

SDD

Spec-heavy, AI builds

VIBE CODING

Prompt, ship, test later

Agile, SDD and vibe-coding splits are illustrative, not measured. Lambert, slides 5–10

OVERVIEW, PART 5: FOUR METHODS (about 2 minutes)�The deck's frame: every method runs the same five activities, requirements, design, code, test and release. What changes is the weight on each and how often you go around.��Waterfall: one long pass, heavy on testing.�Agile: the same five activities in one short cycle, repeated. Each iteration is a small waterfall.�Vibe coding, which the deck calls an extreme agile: prompting takes the bulk of human effort. Code and release shrink because the model does them. Design is done by the AI and is never documented or reviewed. And testing moves to after release, so in effect your users are the test.�SDD sits in the middle. The deck pictures the methods as gears: smaller wheels turn faster and cover the same distance. SDD is a smaller, faster wheel than agile, but unlike vibe coding it puts real human effort back into requirements and design.��Again, the deck is explicit that only the waterfall split comes from a source (Brooks). The agile, SDD and vibe-coding proportions are the author's illustrations.

9 of 24

9

sg

WHAT IS SDD?

Three big names, one idea

OPENAI

10–20%

of value is the code

Sean Grove, "The New Code," June 2025

GITHUB

4

phases in Spec Kit

Open-source toolkit, Sept 2025

IBM

3

levels of SDD

IBM Think topic page

Source: Lambert, slide 8

OVERVIEW, PART 6: WHO'S BEHIND SDD (about 1.5 minutes)�The deck shows this isn't one person's idea. Three examples:��OpenAI: Sean Grove's talk "The New Code" at the AI Engineer World's Fair in June 2025 argues code is perhaps 10 to 20 percent of a project's value. The rest is structured communication of intent: the specification.��Microsoft's GitHub released Spec Kit in September 2025, an open-source toolkit for AI coding agents with four phases: specify, plan, tasks, implement.��IBM's Think explainer says the spec is written and agreed before development begins, and describes three levels: spec-first, spec-anchored, and spec-as-source, where the spec effectively becomes the source code. IBM also has tooling for it, called Bob.��Note these are the vendors with the most to gain from AI coding. That's a fair thing to bring up in thread three.

10 of 24

10

sg

HOW IT RUNS

From charter to test plan

Charter

Plan

BRD

FSD

DTD

Tests

At every step, a person and the AI draft the document together, and a person signs it off.

BRD business requirements · FSD functional spec · DTD technical design. Lambert, slide 9

11 of 24

11

sg

THE PRODUCTIVITY MODEL

Productivity has a quality term

P = V·Q / T

V

value delivered

Q

quality

T

time

If Q goes negative, more speed only makes things worse.

Source: Lambert, slides 12–14

12 of 24

12

sg

THE SIDE EFFECT

Knowledge is part of quality

1

Write context for the AI

2

It becomes docs the team can read

3

The AI explains it back

4

Team knowledge raises Q

…and around again

Source: Lambert, slides 15–16

13 of 24

13

sg

THE RESEARCH QUESTION

Three hypotheses, not yet results

H1 · VIBE CODING

Fastest

worst ROI

H2 · NO AI

Slowest

most expensive

H3 · SDD

Middle

best ROI

The deck states these as hypotheses and presents no data for them yet. Lambert, slide 18

14 of 24

14

sg

TAKEAWAY

What to measure tomorrow

T

Cycle time

In progress to done

V

Throughput

Issues closed per week

Q

Reopened rate

Moved back from done

Q

Escaped defects

Bugs in released work

KT

Learning log

What did we learn?

Record a baseline from your tracker before you change anything. Lambert, slides 21–22

OVERVIEW, PART 11: THE PRACTICAL TAKEAWAY (about 1 minute)�The most usable slide in the deck. Five things any team can pull from Jira or Linear next week:�- T, cycle time: status-change timestamps from in progress to done.�- V, throughput: issues resolved per sprint or per week.�- Q, reopened rate: issues moved back from done.�- Q, escaped defects: bugs linked to issues that were already released.�- KT, knowledge transfer: keep a learning log. What did we learn this sprint?�The deck marks cycle time and throughput as easy to get, and the two quality measures as harder. Its broader quality scorecard adds change failure rate and review findings per document, and for knowledge transfer: time for a newcomer to first accepted work, glossary terms agreed and in use, decisions recorded with their reason, and questions answered from documents.��The rule: record a baseline before you change anything, or you'll never know whether AI helped.

15 of 24

15

sg

THE CLOSING CLAIM

Without quality, speed can multiply a negative.

SDD puts effort back into the spec, and keeps a person accountable for it.

Lambert, slide 23

16 of 24

16

sg

OPENING HAND VOTE

For a startup, is spec-first worth the slower start?

A

Yes, write the spec first

B

Depends on the stage

C

No, ship and fix later

Hold that thought — we'll come back to it.

17 of 24

17

sg

THREAD 1 · EXPERIENCE

When AI work looked done but wasn't

When has AI output cost you time later, and who caught it?

CONTEXT: The deck says quality can go negative: work that looks finished but costs more to fix than it was worth. Let's ground that in real examples.��STARTUP & VENTURE LENS: For founders, negative-quality code is hidden debt that shows up in diligence and in the first big customer. For investors, it's a question of how you'd even spot it.��FOLLOW-UPS:�- Engineers and startup employees: "Was it the code, or the design decision nobody reviewed?"�- Service providers (lawyers, accountants, marketers): "Same question outside code. Where has an AI draft looked done and wasn't?"�- Investors: "Has AI-built product changed what you ask in technical diligence?"��TEST THE DECK: "Is AI work more likely to be negative-quality than a rushed junior hire's, or just faster to produce?"��QUIET ROOM BACKUP: "Raise a hand if you've had to redo something an AI produced in the last month." Then ask one hand: "What gave it away?"

18 of 24

18

sg

THREAD 2 · PREDICTION

If the spec is the product…

If code is 10–20% of the value, what does a senior engineer do in 2029?

CONTEXT: The deck cites Sean Grove's claim that code is perhaps 10 to 20 percent of a project's value, and IBM's "spec-as-source" level, where the spec effectively becomes the code.��STARTUP & VENTURE LENS: If specs are the asset, what's the moat? Can a competitor with your spec rebuild your product in a weekend? What does that do to valuations of software-only startups?��FOLLOW-UPS:�- Founders: "Who's the first hire now: an engineer, or someone who writes great specs?"�- Startup employees and engineers: "Is spec writing a skill you're building on purpose, or by accident?"�- Investors: "Would you ever ask to read the spec in diligence?"�- Service providers: "Lawyers, contracts are specs. Does this make your job look more like engineering, or engineering look more like yours?"��TEST THE DECK: "Grove's 10–20% is one person's estimate. Does it match your experience?"��QUIET ROOM BACKUP: "Finish the sentence: in three years, the most valuable person on a software team is the one who..."��CALLBACK: In September's AI 2040 deck, Bill Fishkin projected coding as one of the first jobs where AI adoption would roughly double by 2026. Does SDD change who does that work, or just what they do?

19 of 24

19

sg

THREAD 3 · TEST THE DECK

Midpoint, or waterfall renamed?

Is SDD a real middle path, or waterfall with a new name? What would change your mind?

CONTEXT: The deck places SDD between agile and vibe coding, and its document chain (charter, plan, BRD, FSD, technical design, test plan) will look very familiar to anyone who lived through waterfall.��STARTUP & VENTURE LENS: Startups adopted agile to escape heavy documents. Is this a return of big-company process, and who benefits from selling it? Note the three backers on slide 8 are OpenAI, GitHub/Microsoft and IBM.��FOLLOW-UPS:�- People from big companies: "Does this look like what your company already does, with AI added?"�- Founders: "Would you put a BRD and FSD in front of a three-person team?"�- Engineers: "Is a prompt file in your repo already a lightweight spec?"��TEST THE DECK: "The deck's effort splits for agile, SDD and vibe coding are illustrative, not measured. What evidence would you need to believe H3, that SDD has the best ROI?"��QUIET ROOM BACKUP: "Show of hands: who has written a BRD in the last year? Keep your hand up if an AI helped."��CALLBACK: Last month Steve Blank argued AI killed the MVP, because his teams showed up on day one with finished products. Is SDD the answer to that, or the opposite of it?

20 of 24

20

sg

THREAD 4 · SO WHAT FOR STARTUPS

Spec vs. speed at seed stage

Where in a startup's life does writing the spec first pay off, and where does it slow you down?

CONTEXT: Startups win on speed, and vibe coding is the fastest path to a demo. The deck argues that without quality, speed multiplies a negative.��STARTUP & VENTURE LENS: Timing. Maybe vibe coding is right for finding product-market fit and SDD is right for scaling, regulated markets, or the first enterprise contract. Where's the switch point, and does an investor care which one you're using?��FOLLOW-UPS:�- Founders: "What's the first moment you'd want a written spec: first customer, first hire, first audit, Series A?"�- Investors: "Would you fund a team's spec-writing time, or tell them to ship?"�- Startup employees: "Who on your team would own the spec today?"�- Service providers in regulated fields (health, finance, legal): "Does SDD map onto what your clients already have to document?"��TEST THE DECK: "The deck's last slide says start measuring cycle time, throughput, reopened rate and escaped defects. Which of these does your startup actually track today?"��QUIET ROOM BACKUP: "If you had to give one rule to a first-time founder using AI to build, what would it be?"

21 of 24

21

sg

OPEN FLOOR

What the deck leaves out

No data yet: hypotheses only

How do you put a number on V?

The cost of writing specs

When vibe coding is right

OPEN FLOOR (about 10 minutes)�Open it up: "What's missing from this deck? What would you add?" The four cards are prompts if the room needs them:��1. No data yet: the deck states hypotheses, not findings. What would a fair test look like?�2. Value is hard to measure: the formula needs V, but the scorecard only measures T, Q and knowledge transfer. How does your team measure value?�3. The cost of the spec itself: who writes it, how long it takes, and what happens when it's wrong.�4. Where vibe coding is the right call: prototypes, throwaway tools, internal scripts. The deck mostly argues against it.��Also fair game: the title says "ethical AI," but the deck spends little time on ethics beyond accountability. Is keeping a person accountable for the spec the whole of the ethical case?

22 of 24

22

sg

RE-VOTE

For a startup, is spec-first worth the slower start?

A

Yes, write the spec first

B

Depends on the stage

C

No, ship and fix later

Did anything you heard move you?

23 of 24

23

sg

WHAT'S NEXT

Keep the conversation going

THU, NOV 19 · 6 PM

Fireside chat with Rob Ott

NJ AI Hub, 619 Alexander Rd

FRI, DEC 18 · 10 AM

Next Deck Club

NJ AI Hub, 619 Alexander Rd

Nominate a deck

Networking until noon

CLOSE (about 4 minutes)�Thank you, all of you. I especially appreciated [SPECIFIC POINT] and the back-and-forth on [SPECIFIC THREAD]. And thanks to Gael Lambert and the Princeton Digital Laboratory for a deck we could share freely. The mapping study and the Shirley agents are at princetondigitallabs.org.��Coming up:�- Thursday, November 19, 6 to 8 PM at the NJ AI Hub: AI x The Relevance Advantage. A fireside chat with Rob Ott, executive coach and life-science entrepreneur, on his new book about what happens to hard-won experience when the tools change underneath it. It pairs nicely with today's thread on what senior people do when the code writes itself. RSVP on our Startup Grind page.�- Friday, December 18, 10 AM to noon at the NJ AI Hub: our next Deck Club.��If you've come across a deck you'd like the group to take on, use the nomination form here on site, or come find me.��Thanks for starting your Friday with us. Stick around, grab a coffee, and keep the conversation going. Networking until noon.

24 of 24

24

sg

ALSO COMING UP · WED, DEC 2, 2026

NJx Venture Summit

NJ startups pitch to VCs, angels and industry leaders

8 AM–6 PM · Montgomery Innovation Hub, Skillman · Presented by SG Princeton

5

pitch tracks

17

investor firms

303

July sign-ups

njxventuresummit.com

Startups: apply to pitch

NJx VENTURE SUMMIT (about 1 minute)�One more date for your calendar. On Wednesday, December 2, Startup Grind Princeton is one of the presenting partners, with 1435 Capital Management and Cogent Connections, of the 5th NJx Venture Summit. It runs 8 AM to 6 PM at the Montgomery Innovation Hub, 23 Orchard Road in Skillman.��What it is: New Jersey startups pitch to venture capitalists, angel investors and industry leaders, across five pitch tracks, each paired with a panel:�1. University spinouts: research, ownership and company formation�2. Public investment in New Jersey's innovation economy�3. Early-stage companies: legal and right-sizing considerations�4. The NJ early-stage investor's perspective�5. Women building and backing ventures��There's a keynote at 9, a startup lunch at noon, and a networking reception from 4 to 6.��The site lists 17 investor firms, including Alumni Ventures, Techstars, SOSV, Plug and Play, NJ Angels and TiE New Jersey. The summer edition in July drew 303 registrants and about 40 startups.��For today's room:�- Founders: you can apply to pitch through the form on the site.�- Investors: there's an investor ticket that includes startup introductions and a firm bio in the event guide.�- Everyone else: general admission includes the full day and the reception. Early-bird pricing is on for a limited time.��Register at njxventuresummit.com.��Source: njxventuresummit.com, checked November 5, 2026.