Product leadership · Matt Lane · Head of Creator Product, Splice

Scroll to read. Wide diagrams and tables can be swiped sideways.

The narrowest product changed the strategy.

Splice CoSo → Create / Stacks. An AI-assisted song starter built from the sample catalog, with a path back into the creator’s own music software.

Splice's catalog supplied every layer Blank page A promising idea Finished track what we built stays in the creator's DAW
The whole case in one line: we took the first step and gave the rest back.
What I did: proposed the shift, made the boundary call, set the product requirements and evaluation bar the AI had to meet, sequenced the build by risk, and defined success as saves and exports.
What the team did: designed, engineered, built the model, analyzed, launched and evolved the product.
80% reported adoption in 60 days, against an internal ≈30% launch benchmark.
Internal figures · definition and limits →
My work: the earlier CoSo strategy, MVP scope and shaping. Later evolution: Create / Stacks.
01 · Where we started

Powerful — once you knew what to search for.

reject it, search again — over and over Search Audition Download the hard part turn sounds into an idea Your DAW What the product owned nobody's yours
The catalog was excellent. The product ended one step before the hard part — and that step is where momentum either happens or doesn't.
Sample

A short audio clip — a drum loop, vocal phrase, bass line.

DAW

Where you actually produce — Ableton Live, Logic, Studio One.

The business

A subscription to the catalog. Value depended on what you could do with it.

What I did: mapped the path creators actually took — search, audition, drag sounds into the DAW one by one — and the point where they gave up without an idea.
02 · The real problem

Finding sounds is not the same as making progress.

Five sounds you found different tempos and keys · no relationship the assembly nobody helped with A Stack same catalog · one tempo, one key, in context
Search helped creators find sounds. The unresolved gap was assembly — hearing sounds together, in time and in key, before you have an idea worth keeping.
Creator research

Kept surfacing the same blank-canvas friction at the start of a session.

Behavioral signals

Search and browse alone weren't pulling people into repeat creation.

Business pressure

Marketplace subscribers were declining coming out of COVID.

What I did: combined creator research, behavior and business pressure into one diagnosis: finding sounds was not enough to help people start.
03 · The strategic reframe

Move the product one step later.

THE JOB WE HAD Help people find sounds Blank page Sounds in hand the step we added A usable starting point Finished track THE JOB WE WANTED Help people start making something
Same catalog, same creator, same DAW. The product's boundary moves right — into the moment that was empty.
Customer bet

Less friction at the start creates faster, repeatable progress.

Business bet

Repeatable progress makes the subscription habit-forming.

Product principle

Augment the start. Don't replace the studio.

What I did: proposed this shift to leadership and framed the problem it had to solve.
04 · The decision that shaped everything

How much of the creative workflow should Splice own?

Blank page A usable starting point Arrangement Mix Finished track the DAW's job → 01 · Improve discovery too incremental — you still assemble the idea yourself 02 · Guided song starter hand back own the start, return the creator to their DAW 03 · Richer composition tool crosses the line
Three answers to the same question, drawn on the same workflow — showing how much of the creative process each option would ask Splice to own.
The assumption underneath: creators finish tracks in their own DAW. The boundary, the requirements and the success metrics all rest on it; if it were wrong, they would be wrong together. That is why export was non-negotiable.
What I did: made the boundary call — the guided song starter over better discovery and a full composition tool. Appendix A shows the requirements, trade-offs and open questions behind that choice.
05 · What we actually built

One loop, one exit.

What the product owns Splice catalog supplies every layer Pick a seed Generate a Stack up to 8 compatible layers Hear it Save it adjust swap · mute · solo · balance hear it again acquire what you kept export Your DAW stems · samples · project arrangement · mixing · finishing never ours Deliberately outside the box: a full composition environment · social collaboration · unconstrained generation · replacing the DAW
The steering loop is the product; the export is the promise. Everything on the last line was a cut we made on purpose.
What I did: set the hard edges — limited editing, export-first, and no composition environment, social layer or unconstrained generation. Each cut protected scope, creator trust or product identity.
06 · What the AI had to accomplish

I defined what “good output” had to mean for the product.

The architecture belonged to ML / research. My job was to define the product requirements the AI had to satisfy — the bar every Stack was judged against.

Must-have

Worth continuing

The first result had to feel coherent enough to build from.

Must-have

Fast to steer

A creator had to be able to repair an idea without starting over.

Must-have

Catalog-native

Every layer had to remain attributable, acquirable and usable.

Must-have

Workflow-fit

The result had to move cleanly back into the creator's existing production environment.

Each is a Must-have in the decision ledger (Appendix A). Section 08 is what happened when output cleared the surface checks but missed the first one.

What I did: defined the product requirements and evaluation bar. ML / research owned model design and implementation.
07 · How we sequenced the build

We ordered the work by uncertainty, not by feature.

1 · Coherent Stack Can a coherent Stack exist at all? 2 · Fast steering Can users repair it fast enough to trust it? 3 · Catalog in context Do creation and discovery reinforce each other? 4 · Save & reopen Is it worth saving and coming back to? 5 · DAW handoff Does the work continue into production? constrained generation, playback, catalog links swap, mute, solo, add, BPM, mix, undo persistent playback, in-sync preview save, reopen, re-edit, re-save, share stems, samples, mix, project exports If slice 1 had failed, everything to the right of it was wasted work
Ordered by dependency and risk, not by how attractive each piece was. One ≈2-month strategy cycle, then roughly four six-week build cycles.
What I did: wrote the slice plan and ordered it by risk — which is why the failure in the next section showed up early enough to act on.
08 · The thing we got wrong

Creators kept refreshing instead of moving forward.

yes Save · Export · DAW the path we wanted — users hesitated to save Generate Hear it Worth building on? no Refresh refreshing kept sending users back to the start or just leave session over
The output could pass surface compatibility checks and still feel wrong. It could still fail the key creator test: is this worth continuing from?
Known risk

My slice plan asked it first: “Can the first result feel coherent enough to keep?”

My bet

That stronger output quality would answer it. I expected generation quality to be the hard part.

What was missing

Compatibility alone did not establish trust. Creators also needed a quick way to repair and steer an idea.

09 · What that learning changed

Help creators improve an idea without starting over.

yes Save · Export · DAW the path we designed for narrowed by stronger compatibility constraints Generate Hear it Worth building on? not quite swap · mute · solo · balance one short loop Refresh is still there — alongside faster ways to steer.
Stronger constraints improved the starting point. Fast steering let creators hear an adjustment immediately and keep the parts that worked instead of regenerating the whole idea.
What I did: recentered the team on the creator's workflow, not the model's output — capping the layers, keeping editing simple, protecting the path back to the DAW, and measuring saves and exports instead of generation volume.
10 · What happened

Adoption beat our own launch benchmark.

Song assistant · internal launch figure80%
Other Splice launches≈30%

Reported internal 60-day adoption. The denominator, qualifying action and launch cohort are not specified here; this comparison does not establish retention or adoption across all subscribers.

60%

subscription growth in under a year — roughly 500K → 800K subscribers, after a post‑COVID decline.

My assessment: the broader creator-tool strategy helped drive the turnaround. These are business-level figures from my project account; other changes contributed, and this product’s individual effect was not isolated.
What I did: defined success as saves and exports rather than generation volume, with our own launch benchmark as the bar.
11 · What the wedge enabled

Same wedge. More doors into it.

  1. 2022
    CoSo: start from a style

    Assemble compatible catalog layers on mobile.

  2. 2023
    Create: start on the web

    Stacks came to the website, then gained the ability to start from a catalog loop.

  3. 2024
    Start with your own audio

    Match catalog layers to a creator’s uploaded loop.

The same core workflow: compatible layers → fast steering → the creator’s DAW.

Splice Create: catalog search results beside a Stack of compatible layers with export controls
Splice product demonstration · Oct 2023Later public evolution

Mine: the earlier strategy, the boundary, MVP framing and shaping. Not mine: the later surfaces the teams after me built.

My reading of the later expansion: the core workflow had room to grow through more entry points and deeper integration.
12 · What changed how I work

Ambitious outcomes needed unusually narrow scope.

01

Quality is necessary, not sufficient.

accurate output stops accurate output you can steer continues

People adopt systems they can understand, steer and recover from. I now treat control as a first-order requirement, and test it before investing in generation breadth.

02

The right wedge is smaller than you expect.

everything we could have built the part that mattered

Own the step where friction is highest and value is immediate — not the whole adjacent market.

03

Say what you're not building, early.

IN coherent layers light steering save + export OUT composition env social, free generation replacing the DAW

Defining the cuts up front is what let ambition stay high while the build stayed shippable.

Where would you like to go deeper?
Appendix A · Decision ledger

The strongest candidate still had a critical unknown.

Retrospective assessment of the specific concepts considered. Design fit is not a measured result. The current product is the baseline.
RequirementCurrent productDiscovery onlyStack + steeringFull compositionCurated packs
Core goalBlank page to a promising start, fastfrom creator research—~✓✓~
Must-haveA first result worth continuing fromfrom the creator outcome—~?test first?✓
Must-haveRepair an idea without a full DAW workflowfrom the need for fast control—~✓~—
Must-haveCatalog and licensing clarity on every layerfrom the catalog business model✓✓✓~✓
Must-haveStrengthen sound acquisition and subscription valueincremental business value remains a hypothesis—????
Must-haveContinue in the creator’s own DAWfrom the workflow assumption in section 04✓✓✓?✓
Decisionbaselinetoo incrementalTest & buildexceeds scopelimited steering
✓ plausible design fit · ~ partial fit · ? unproven · — not addressed. These are judgments about the concepts considered, not universal limits of each category. An unknown Must-have required a test; it was not a pass. Trust in the first result was the first test.
Why this candidate: it paired a fast start with creator control and catalog continuity inside a credible scope. My reconstructed weighted comparison favored it too: Stack + steering 94; packs 74; discovery 71; full composition 37. These are decision scores, not performance measurements.
Reconstructed after the fact. The choice authorized learning and development; it did not establish that the product would deliver every outcome.
Appendix B · Rotated fit check

What each part was for.

The parts of the chosen product against the six requirements from Appendix A: filled dot means an intended contribution, dot with a question mark means an explicit open question, dash means no contribution was claimed
PartR0 · Start fastR1 · Worth continuingR2 · Fast repairR3 · Catalog clarityR4 · Subscription valueR5 · Into the DAW
Pick a seed — genre or style●—————
AI compatibility / generation system●●?————
Every layer drawn from the real catalog———●——
Steering — swap, mute, solo, balance, undo—●?●———
Browse the catalog in context————●?—
Save, reopen, re-edit————●?—
Acquire what you kept———●●—
Export — stems, original samples, stereo mix———●?—●?
● intended contribution · ●? explicit open question · — none claimed. This maps the chosen candidate’s parts to its requirements. The outlined assumption proved insufficient in early use (section 08); stronger compatibility constraints and faster steering addressed it together (section 09). The slice plan made the unknowns testable.
Coverage isn't sufficiency: a mark says a part claims to help, not that it's enough.
Appendix C · Who did what

One objective, different risks.

FunctionTheir concernHow it shaped the productAttribution
Product (me)Choose the problem, the boundary, the success criteria and the learning sequence.Kept the wedge on creator progress rather than model novelty, and defined the product requirements and evaluation bar the AI had to meet.I framed, led, chose
EngineeringRecommendation quality and workflow complexity could explode scope.Hard edges early: bounded system, ordered dependencies, export-first.Engineering developed
ML / researchCould generation produce layers that are compatible, not just plausible?Model design and implementation that aimed to keep the output coherent and steerable within the product requirements.ML / research owned model design and implementation; I defined the product requirements and evaluation bar
DesignCould this feel magical without becoming opaque or gimmicky?Visible layers, immediate audio feedback, simple actions, recovery.Design created; I partnered
DataWhat counts as success, and what's a vanity metric?Moved the readout from generation volume to continuation signals.Data analyzed; I set criteria
Growth & leadershipIs this big enough to matter, but narrow enough to ship?Framed as a strategic wedge tied to activation and conversion — not an AI brand experiment.I made the case; leadership decided
Content, licensing, brand, PMMCreator trust, royalties, sameness, rights clarity, launch fidelity.Kept catalog provenance and creator control central; set up the launch moment.I partnered; each owned their domain
I ran this as framing and shaping loops rather than a long PRD. When an objection was strong, I turned it into a shaping question — what would have to be true for this concern not to kill the project?
Whose outcomes counted: creators, who needed a faster start they controlled; Splice, which needed subscription value; and the artists whose sounds fill the catalog. Using the catalog as the source of each layer meant every layer stayed an artist's sound, with its provenance intact.
The right-hand column is the attribution: what was mine, and what each function owned.
Appendix D · Evidence and limits

What the evidence can and can't say.

QuestionWhat we looked atResultStatus
Did it work as specified?Built product vs. requirementsCoherent Stacks from a seed, steering, save, export into a DAWUp to eight layers, steering, sharing and a path into the DAW, per Splice's own walkthroughPublic
Did people adopt it?Launch outcomeReported 60-day adoption; cohort and qualifying action not specified here80%, versus ≈30% for other Splice launchesInternal figure, reported by me
Did creators carry the work forward?Creator outcomeSaving or exporting a StackTracked internallyInternal, not shared
Did they come back?Creator outcomeRepeat use, return to the workflowTracked internallyInternal, not shared
Did the business change?Business outcomeSubscriptions60% growth in under a year, ≈500K → 800K, after a post‑COVID declineReported by me; contribution not isolated
How much was this product?AttributionThis product's share of the growth; retention or revenue liftNot isolated: other things changed in the same periodNot claimed
Three claims, kept separate: the launch saw strong adoption against our own benchmark; the business grew substantially over the same period, alongside the broader creator-tool strategy; and neither figure isolates this product's own causal contribution.
Figures come from my project account. Underlying reports, adoption denominator and launch cohort are not reproduced here. Public sources below corroborate product capabilities and chronology, not these internal metrics.
Appendix E · Workflow and system boundary

One canonical Stack had to survive every handoff.

one Stack object, six handoffs — state has to survive all of them Generate Stack editor Catalog browse Save / library Reopen / re-edit Export / DAW must stay true the entire way stable layer roles · compatibility-aware replacement · preserved timing, pitch and mix state catalog provenance on every layer · save-and-reopen integrity · rights-clean export formats The hard part: compatibility has to hold after every edit — not only at first generation.
Product requirements, not architecture — this is the level the decision actually turned on.
Model internals and training data are deliberately left out.
Appendix F · My slice plan

Dependency first, value proof second.

SliceGoalPlanned capabilityKey unknownDepends on
1 · Coherent StackProve the core object can exist.Constrained seed, up to 8 compatible layers, playback, catalog metadata.Can the first result feel coherent enough to keep?Nothing
2 · Fast steeringTurn interesting output into usable workflow.Swap, delete, mute, solo, add, BPM, mix, refresh, undo/redo.Which controls restore trust without becoming a DAW?Slice 1
3 · Catalog in contextConnect creation to discovery and business value.Persistent playback, in-sync preview, compatible insertion.Does contextual discovery beat pure generation?Slices 1–2
4 · Save, reopen, shareTest whether the object is worth keeping.Persist the Stack, reopen, re-edit, re-save, share.Reusable asset, or disposable output?Slices 1–2
5 · DAW handoffProve creation continues into production.Stems, original samples, stereo mix, supported project assets.What export is minimally sufficient and rights-safe?Slices 3–4
This summarizes my slice plan, rather than a dated release log. Timing was one ≈2-month strategy cycle and roughly four six-week build cycles; the five slices did not map one-to-one onto those cycles.
Export mattered enormously — but only once the object was coherent, steerable and worth keeping.
Appendix G · Sources and provenance

Project history first. Public evolution second.

Project historyRole, chronology, decisions, success criteria, outcomes and lessons come from my own work on the project.
On the artifactsAppendices A and B reconstruct the decision logic; Appendix F summarizes the slice plan. These are explanatory artifacts, not dated originals. Diagrams simplify the workflow, and later public features are identified separately.
Splice · How to use CoSo (May 2022)Walkthrough of the older version — stack up to eight layers, swipe to steer, share, and find the layers on Splice to continue in a DAW.
Splice · Introducing Create (June 2023)Launch article — up to eight layers, 12 styles, export to Ableton Live, stems, or original samples.
Splice · Create from any loop (Oct 2023)Product article — start a Stack from any of 1.4M loops while browsing. Source of the product image in section 11.
Splice · Audio uploads (June 2024)Feature update — upload your own idea and match catalog layers to it.
Splice · MobileProduct page — start from a genre or loop, up to eight layers, save, share, export, desktop sync. Checked September 25, 2026.
Deliberately omittedModel internals and training data, confidential roadmap, commercial terms and licensing mechanics, and any private metric not already public or on my résumé.
Public materials give context — they don't rewrite the project history.