The Workflow Company

We help B2B startups build and scale a consistent pipeline. We use AI and automation for precision in targeting, hyper-personalization in messaging, and ease in scaling.

Founder and CEO

First Round Fast Track

Mentor

Integromat

I help startups integrate Integromat into their operations to streamline and automate processes

Registered Partner

joerhew.com

I partner with early stage startup founders to bring clarity to their strategy and improve their operations. I help companies prioritize what’s truly important by leveraging data, then help implement a culture of quick iterations and learnings through no code. I like to get my hands dirty and execute, but also act as a strategic thought partner to the founders.

Clients include: Karrot (Daangn Market, Series D, backed by DST), Groove (Seed, backed by Resolute)

Fractional Head of Operations

Ops Hacks

Built a community of ops folks that work at early stage startups

Community Builder

Tyltgo (YC S20)

White-label, same-day delivery service for local merchants across Ontario

Co-founder & COO

Uber

(Sensitive figures redacted)

• Lead a team of 4 operations managers to perform continuous marketplace management of over 100+ municipalities across Canada with focus on growth, reliability, and efficiency
• Launched a series of experiments, features, and processes to consistently exceed targets on growth and costs - increased year-over-year deliveries by x% while reducing courier costs by y% 
• Provided managerial oversight to launch walkers as a delivery mode in Canada, which accounted for c. x% of deliveries in downtown Toronto within y months of launch

Head of Marketplace Operations, Uber Eats Canada

• Managed a cross-functional launch team for 3 months in leading the largest single-day launch in the history of Uber Eats across 30 cities in Canada, which increased national population coverage by c. x percentage points
• Built out Reporting & Performance Management function using SQL and GSheet, providing transparency and enabling accountability on organization’s OKRs for key stakeholders
• Conducted a technical analysis and modeling on x million+ historical transactions to determine the optimal pricing structure and project-managed the pilot of service fees in Canada

Business Operations, Uber Eats Canada

• Collaborated closely with the product team to launch a new carpool product, uberSHARE, the first P2P product since uberX the country's banning of uberX in March 2015
• Launched UberTRIP, an hourly rental product, to target tourists and business travelers and increased weekly city gross bookings by x% without incurring any additional costs
• Successfully overhauled the driver partner incentive program, saving costs equal to c. x% of weekly gross bookings on average with negligible impact on churn and engagement

Rides Country Lead, Uber Korea

Startups

1. One Hive: e-commerce app for international customers interested in purchasing Korean fashion items
• Supported UX and UI design of a mobile e-commerce service to market and sell Korean fashion items to international customers
• As business lead, gained 10K+ users and generated revenue within the first 6 months of launching

2. Foodle: mobile social networking platform with algorithmic recommendations for food enthusiasts 
• As product and business lead, designed UX of the mobile app and led marketing efforts across social and offline channels
• Collaborated with UI designer and front- and back-end programmers to launch product on iOS

Co-Founder / Business Lead

Morgan Stanley

• Built detailed financial valuation Excel models, based on assumptions and hypotheses derived from financial, operational, and economic data, to test feasibility and profitability of envisaged transactions
• Advised large Korean and multinational corporate clients on cross-border M&A transactions and other business development initiatives through providing financial and strategic analyses
• Selected transaction experiences include SK hynix’s US$ 248MM cross-border acquisition of Link_A_Media Devices and MBK Partners’ KRW 1.2Tn (c. US$ 1.0Bn) acquisition of a 30.9% equity stake in Woongjin Coway

Investment Banking Analyst

Citi

• Selected transaction experiences include serving as the Global Coordinator and Joint Bookrunner in Hyundai HCN’s KRW 83.6Bn (c. US$ 74MM) initial public offering and SK Networks’ US$ 700MM minority equity investment in MMX Mineração e Metálicos

We use code and AI to build unique, 1-of-1 GTM engines for scaling startups.

Reach out to chat about AI <> GTM.

We are always testing with new tools but here are some of the mainstays we use on a daily basis:
- Claude Code
- Trigger.dev
- Supabase
- Langfuse
- Railway
- Mastra
- Github
- Attio
- Clay

We are a proud member of the 1% for the Planet and are committed to donating 1% of our annual revenues to help protect the planet.

Joe Rhew

Here's a generally accepted best practice that we've gone full contrarian on:

The conventional wisdom on GTM LinkedIn is that you shouldn't get an LLM to copywrite a full email. Instead, get an LLM to write snippets that you sprinkle throughout the email.

Well, we have been letting Claude Sonnet write full multi-email sequences for a few months now. And it works quite well (see attached image - although not every reply is like that; this particular startup has a very compelling product).

The key difference in how we make it work despite going against the accepted best practices - we've modularized our prompt and compose the full prompt at runtime. We've created a prompt each to share context on the company, product offerings, voice & tone, etc. And we pull these in to create one comprehensive prompt before generating an email sequence.

Here are the main benefits of this approach from my pov:

1. You establish a baseline to measure against. You have an email factory function, which ensures consistency and you can iterate on by updating modules
2. You mitigate drift. The latest contexts live in these modular prompts and will always get pulled into generating emails
3. Iteration tempo increases. You have a hypothesis about a new segment - create a new persona prompt and inject it versus the default prompt to test
4. You create evergreen contexts. You can reuse these modules for inbound, outbound, event invites, product launch emails, etc.

There's a lot more that goes into it - in terms of manual and automated tests and evals to ensure the quality of the output - but that's the gist.

Currently humans (i.e. me) are orchestrating the context composition - depending on the campaign, I am hand-picking which prompt modules should be inserted into the overall prompt. But soon enough we'll be able to delegate this to an agent, who will be able to review each individual recipient and compose the right contexts, which is when things will get really interesting.

Anyone else doing this?

GTM Intelligence Orchestration | The Workflow Company

Well done Drew + the Vercel GTM eng team 👏👏👏

This is definitely the direction that every business should go in with AI - bespoke GTM workflows based on hardcoded business contexts. And trying to SaaS-ify this runs counter to the value prop of full customization tailored to your specific business.

And if you're looking for an agency to do this - we've been doing this for our clients for the past year. Reach out for a brainstorming session.

Original tweet here if you want to learn how Vercel rolled out an AI SDR to remove 90% of the manual labour while keeping key metrics at baseline: https://lnkd.in/gc7a5MUj

Screenshot: 11 days. 121 commits. Complete framework migration.

I just finished migrating our GTM automation platform from Mastra to Claude Agent SDK. Here's why and and what I learned.

---

The Problem with Abstractions

Mastra is a solid framework. But GTM moves fast.

When a prospect starts a new job, you want to reach out *that day*. When a competitor launches a feature, you need a campaign *now*. When your client ships a new product, messaging needs to update *immediately*.

With Mastra, every new campaign meant diving into the codebase. Change the prompts. Update the workflows. Redeploy.

I could have built more tooling around it. But I kept asking: why am I fighting the abstraction instead of just... writing markdown?

---

The Unlock: Markdown as Configuration

With Claude Agent SDK, campaigns are just markdown files.

/campaigns/new-job-signal.md
/campaigns/competitor-response.md
/products/latest-feature.md

Client just raised a round? Update one markdown file.
Launching against a new competitor? Create a campaign brief.
Testing new positioning? Write it in plain English.

No code changes. No deploys. Just context.

The agent reads what it needs, when it needs it. That's it.

---

The Takeaway

Abstractions are great until they slow you down.

For GTM, where signals matter and timing is everything, the ability to iterate in plain English—just markdown files—is the unlock.

Sometimes the best architecture is the simplest one.

Thank you to Jacob Dietle, Jiquan Ngiam, and Yu-Sheng Kuo for your advice - I've picked each of your brains before committing to the migration, and it was definitely the right decision.

We just restructured our prompt architecture and cut LLM costs by 85%.

We've been building GTM automation systems — the kind that research prospects, match them to personas, and generate personalized multi-touch sequences. Each generation prompt runs 10k+ tokens of context: product positioning, messaging frameworks, company narrative, matched personas, lead research.

When you're processing thousands of leads per week, that adds up.

Most of that context is identical across every lead in a batch. Same product. Same messaging. Same company story. Only the persona matching and lead data changes.

So we restructured the prompt to put shared contexts first:

- System prompt ← 100% cache hit
- Shared contexts ← 95%+ cache hit
(products, messaging, etc)
- Resolved contexts ← 60-80% cache hit
(AI-matched personas)
- Lead data ← 0% 
(expected for individual data)

Caching: 90% off repeated tokens

Anthropic's prompt cache gives you a 90% discount on tokens that match a previous request - but only if they appear at the start of your prompt. Reorder your context wrong and you get 0% hits.

Batching: 50% off everything

The Batch API runs async (up to 1hr latency) but halves your token costs across the board. For scheduled outbound campaigns, the latency is irrelevant.

The primer problem

Here's where it got interesting. Batch API processes requests in parallel across multiple workers. Without coordination, each worker warms its own cache independently — killing your hit rate.

So we built a primer mechanism: before submitting the main batch, we fire a single request to warm the cache. Wait for confirmation. Then submit the full batch. Every subsequent request hits the warm cache.

Did reshuffling the context hurt quality?

We backtested ~500 emails — same inputs, same model, old structure vs new. No measurable degradation. The model doesn't care if product context comes before persona context. It all lands in the same attention mechanism.

What this feels like

At 15% of the original cost, we can now be more detailed in our instructions and run and feed deep research contexts. Run more experiments.

It feels like abundance of intelligence - not because the models got cheaper, but because we architected a system to make the right trade-offs (latency for cost) for the use case.

Would recommend caching your prompts if the llm generation task you're running is 1) highly repetitive and 2) requires a lot of input tokens. Relatively straight forward to set up. 

Batching, on the other hand, takes a lot more work to get right but the 50% off discount is well worth the effort if you don't need need the results immediately.

We built a person-level research workflow that injects our company context into every enrichment.

Standard CRM enrichment gives you job title, company size, funding round. 

None of it tells you whether this person matters to your business.

Here's what we did differently:

When someone gets added to an Attio list, our workflow fetches their LinkedIn profile and last 10 posts. But before the LLM analyzes them, we inject our company context:
- Target personas
- Messaging and narrative
- Product capabilities
- Ideal customer profile

The analysis doesn't stop at "who is this person." It extends to "who is this person TO US."

Same person, different client context = completely different fit assessment.

For example, the CEO of a CRM company, might score high for one client's developer tools practice. Score low for another client selling to enterprise sales teams.

The key: we pass a client_id with every enrichment request. The system pulls that client's context from a central database before running the analysis.

Why separate this out?

If I hard code our company context into the prompt, I have to update every enrichment workflow when our personas evolve or product expands.

With a central context database, I update the source once. Every downstream enrichment module inherits the change.

This is the direction every CRM enrichment should go. Not just data about the person. Data about the person in context of your business.

We just spent 3 weeks refactoring our GTM enrichment workflows for agent compatibility. The changes were subtle but critical, and they've shifted how we think about writing reusable code in an agentic setting.

Modularization has always been key to writing elegant code. You create modules that can be easily swapped in and out and shared by multiple workflows.

With agentic capabilities and the ability to expose these modules as tools to the agent to call at their discretion, there's an even stronger call to write modular code.

You create primitives - think of these as building blocks that you often chain together in a workflow - to create deterministic workflows. And not only can you share these primitives across those workflows where you pre-define each step; now you can share them with agents and let them use them however they like to meet the user's objective.

MODULES
[LinkedIn Research] 
[Email Finder]
[CRM Updater]
[Sequence Generator]

WORKFLOWS
[LinkedIn] → [Email] → [CRM] → [Sequence]

AGENTS
[LinkedIn] ↺ [Email] ↺ [CRM] ↺ [Sequence] ↺ [CRM again]

Agents can call these tools in whichever order they deem fit and multiple times, to iterate their way towards an objective.

A lot of our work these days is around thinking about the right way to modularize - how to scope each module and where to draw boundaries, such that modules would work well in both deterministic and agentic workflows.

Here's a real example from last week: We have a module that fetches company records from Attio.
Pre-agents, we'd fetch only the 5-6 fields we knew we needed (company name, domain, employee count, funding stage). Clean, efficient, deterministic.

But when we handed this to an agent trying to personalize outreach, it kept wanting more context. The agent wanted to reference the tech stack, recent funding news, hiring signals - properties we'd deliberately excluded because we didn't need them in our scripted workflows.

The fix wasn't adding those specific fields, but rahter returning the full record and letting the agent decide what's relevant. This made the module "fatter" but infinitely more reusable in agentic contexts.

In practice, this means:
- Our Attio retrieval tasks now return full objects, not filtered subsets
- Our Supabase queries use SELECT * more liberally than we used to
- Our Claude Code workflows document all available fields, even "optional" ones

There's an irony here. Good modular code for agents looks "sloppier" than good modular code for workflows. You're returning more data, exposing more parameters, being less opinionated about usage. It feels wasteful - until you see an agent solve a problem you didn't anticipate by combining modules in a way you never would have scripted.

🧠 Context: I received a request for case studies from a potential client
💪 Me: Prompt my Claude Code agent - 
"Look at the most recent email from [Prospect - let's call them John] asking for case studies.

I've sent something similar to:

[List of prospects/clients]

Then look up the meeting transcript with [John] from yesterday to get more context on what he does [in his current role]. I mentioned [specific contexts we should highlight in the case studies], so think about that + other contexts shared during the convo on how to structure the case studies.

Review all of this and let me know if you have any questions before drafting the email response to [John].

Feel free to use multiple subagents/tasks to do this."

🤖 Agent:
1. Parallelized searches for:
A. previous emails to prospects
B. previous emails that mention case study
C. Email from John
D. Transcript from meeting with John

2. Synthesized the request and the available context

3. Drafted email that was 80% of the way there and I dictated my feedback to get it to 100%

--

A few top of mind callouts after doing this:

1. Funnelling all of your contexts into a unified context layer is key to unlocking agentic capabilities. I save my transcripts in a repo, readily accessible to the agent; and the agent has API access to search, read and draft emails (but not send - human in the loop!)

2. Getting the agent to ask you questions to fill in the gaps / find edge cases is a simple hack to getting better output that's more in line with how you've thought about it

3. Given how many times I've had to send case studies via email, it's probably time for me to put them up on our website. Claude now has rich context on this based on all the emails, so this will be a relatively straightforward process <- Involving Claude Code in every task creates a compounding effect where subsequent tasks get easier, thanks to having more context

4. *Most important* Claude Code isn't going to replace you, but someone using Claude Code will. 

I've been thinking about creating a series of walkthroughs on using Claude Code for GTM - would you watch it or is working in the CLI an instant turnoff?

taplio