Lucas Perret

Operations Lead @lemlist

Just a guy obsessed with tech and ways it solves business problems 🤓

I believe people should spend all of their time on meaningful activities such as building relationships with others.

That’s why I spend all of mine solving the inefficiencies I encounter at work by building the right processes and systems for my team.

I am all about networking and discussing anything related to operations, so if you are interested in having a quick chat then feel free to reach out to me. 💌

Growth Manager - Product & Ops

Richard Makara

CEO @ reconfigured | eradicating inconsistent metrics

GTM Engineer-for-hire. 

Returning to GTM/Growth Engineering after a long hiatus and 3 years as a tech startup founder. 

Love to chat with people about this topic. DM me

--

From a previous life: 

building reconfigured - a quick-capture journal for overloaded minds juggling tabs, tasks, and tools. soon also with MCP capabilities.

sharing progress as I go, usually in the form of neat tricks/productivity gains I've learned along the way.

-- 

Have had the privilege of building the growth engine of a 🦄 company, as well as completely rebuilding CRMs from scratch (Salesforce, HubSpot). 

Some others stats:

- Triangle shaped person - Design, marketing & programming
- B2B SaaS growth + tech stack building.
- Grandfather of Junction, Europe's leading Hackathon 
- Board alumni of Aalto Entrepreneurship Society
- Occasionally tinkering on app ideas with Flask/TS-Node, ReactJS, PostgreSQL & Docker. 
- In a love/hate relationship with TypeScript.

Member

A "gtm context operating system" is pretty much just neatly structured folders with text files, optimized for your LLM, not for you and ideally with access to your daily gtm tools. 

How everything is organized only matters to your agent - so it can find, lookup, and update relevant files _as you chat and use the system_.

If you're interested in trying out this whole GTM OS thing here's a structure I have developed over the last few months while happily living inside of a terminal window.

Feel free to screenshot and ask claude to reproduce. I'll drop a scaffolding with sample files to github when i get a chance.

IMO the three main ingredients are: 

- Topic context wrapped into a folder
- Scripts and/or MCPs that allow you to do things
- .claude configurations - how it should behave. 

While the structure below might seem complicated, the key point is that _I didn't set it up nor do I really manage it_. 

My claude instance did and does. We ofcourse iterated on it together but it was built by and for the AI's consumption.

Claude.md and the skills arent written by me either. Everything is updated via terminal/chat.

The day to day interface is the terminal chat and a select few output files (e.g shareable docs)

Few things to note: 

- It's not meant to be database of actual customer data. Connect to you CRM or a real database, get live data out of your tools when necessary. 
- Folder names are optimized for the LLM, not for you. The AI reads the file tree before it reads any file. The tree IS the first instruction.
- This effectively replaces your notepad because the more you use, the more context gets logged, the easier it is to pull up "docs" from your work.

always felt disappointed in techstack data providers and even more so with AI agents visiting websites to find technology signals. 

always had ideas where and how to find the real signals, buried deep in the darkest corners of the network tab

never had the skills/tools+time to build something to address those 😄

the tricky techs to detect were always the ones loaded on the app side eg posthog. 

but where there's a will (and a claude code subscription) there's a way. 

started as a random idea while walking the dog - what if i just mass accepted all possible consent variants and checked what scripts load. 

spray and pray but for cookies 😂

worked pretty well but needed a few more upgrades to find the beef. 

anyway, need to deploy this somewhere so I can use it in Clay via HTTP column.

Went a little overboard with my BuiltWith clone rabbit hole. 🙂‍↕️

Ended up finding signals that are important to me - someone who mainly works with B2B SaaS companies.

My gripe has always been that tech data is never reliable, even at a directional level. 

But the use case ideas are always something like "we need to find companies who are on Slack and use PostHog analytics," so those plays were never truly accessible.

BuiltWith still finds a ton more tech than my little monster but a lot of it is noise to me. e.g i dont care about "lodash version 1999"

So what I'm gonna do is tailor this thing to specifically focus on B2B use cases: CRMs, sales tech, support, devtools, HR systems, etc.

Anything a budding GTM engineer would need to do first-pass filtering on a larger account list. 😄

API version coming soon. First need to fine-tune and map out more SaaS apps I want to track.

Question to you: what SaaS tech do you want to track but haven't been able to get out of tech data providers?

i asked claude to solve the 2025 Clay Cup final round challenge. 😄

... with the help of this little app ive been building.

it let's you operate clay through claude code via MCP, probably other interfaces too. 

Video is edited for entertainment, the real process took around 8 minutes. 

±4 minutes to figure out initial setup, ±3 minutes to self-iterate its own output.

looking for people who use claude (desktop or code) AND clay on daily/weekly basis. 

especially now that my infra doesn't blow up everytime Mikhail boots up his terminal. 😂

Drop a comment or tag a friend if interested - aim to onboard max new 10 users for now.

Building an MCP server is easier than you think. Making it efficient is a little bit harder. Here's my top "crash and learn" bits from the last month of developing an mcp-first app.

1. Building Tools

Main learning I had was: group tools by job, not action.

Started with 15 tools. Claude spent tokens deciding which to use, descriptions bloated the context window, and half had overlapping responsibilities. Reduced tool count to 5

- Each tool should map to a distinct job. My `do` tool handles create, update, delete, run, rename — all mutations under one roof. Claude doesn't pick between `create_column` vs `update_column`. It calls `do` with an action parameter
- Essentially you create simple to understand group tool and nest all your individual bits underneath it.
- Every user action should map to exactly one tool. If something fits two, you get the spiderman meme.

2. MCP Servers can have system prompts 🤡

MCP servers can inject instructions that shape how the model uses your tools across an entire session. It essentially tells a claude how to think about using the tools available.

- My ~800 words of instructions teach Claude to verify data before acting, evaluate results after mutations, and build incrementally instead of chaining blind
- This turned Claude from a tool-caller into something closer to a domain expert

3. Less Tokens, More Fun

A single data fetch on a 50-column table can return ~10,000 tokens. Claude usually needs 2-5 columns. The rest is waste filling up context.

Fix: Throw data into temp memory instead of the context window. Resolve only what you need. Bernardo from Flow AI wrote a super good article on this.

- Return a pointer + "glimpse" (row count, column names, sample values). Claude reads the glimpse, decides what it needs, resolves specific columns on demand. 80-95% token reduction
- The pointer store is just an in-memory Map. Data already flows through the server — I just stopped discarding it immediately. Zero infrastructure needed

4. Don't put LLMs where deterministic logic works

I had Gemini routing every query to classify it into a query type. 4/4 misclassifications observed in testing.

- Replaced it with Claude specifying the query type directly in the tool call. $0/call, 200ms, zero misclassifications
- LLMs add latency, cost, and non-determinism. Use them only where synthesis genuinely adds value

5. The brain is markdown knowledge, not model

My strategy tool uses Gemini Flash ($0.01/call) over a curated knowledge base. 20x cheaper than Sonnet, same quality for instruction-following tasks.

- The knowledge base is the brain — the model just applies it. Upgrading the knowledge improves output. Upgrading the model doesn't
- Very important: Choose a model that is good at following instructions, not at debating.

I'm maybe starting to think that with the advent of Claude Code in go-to-market, it is now once again okay to do operations in batch vs trying to make everything real-time and fully automated.

Back in 2019, in the prehistoric pre-AI era, when i did this kind of work everything was about fully automating end-to-end - lead signs up, enrich, push CRM record in, associate, enrich again, route, run flows, push to warehouse, etc etc. 

It had to be perfect on the first try. if it wasn't - have fun debugging.

Whereas now, you can kinda just have more or less clean data, and then once a week you just run a batch fix with Claude. 

The biggest issue was always finding the records in the UI, manually reviewing them, associating them to the right place, fixing fields. 

but now you can do that in a batch session with Claude by dumping a bunch of CSVs, cross-referencing and if you have some sane context to attach, youre sorted. Claude is very good at finding things and then confirming with you.

I'm still a major proponent of engineering the duck out of GTM systems so this is utter blashpemy on my part lol.

I did a scary thing: Forcing myself to take a few weeks off with no set projects in place. No pre-built pipeline. Figure out what to do next. 😄😬

Not gonna lie every bit of of me is saying "bro what are you doing, in this economy?" but its probably the right thing to do. 

It's been geneuinly fun to get back into 'gtm engineering' after burnout + running a startup.

Since coming back i've been working on end-to-end gtm engineering projects across early stage - series x companies.

a lot of that involved going deep on Clay. 

a lot of that involved applying Rob's demand frameworks as computation rules.

even more of that involved using Claude Code for GTM and eventually creating a 'portable' version that I can handover to clients at the end so they can continue where I left off + have access to every bit of thinking trialing and testing I've done. 

I've also gotten back into building little mini software products, mainly as aids for my own work (a whacky MCP for Clay, a juiced up technogy mapper for saas companies). 

Anyway, the point of this ramble is that I got some free time now and be awesome to talk with people in the industry. 😄

Hit me up 🤙

my claude "gtm context os" project is littered with JSON files that catalog the context filled markdown files. 

it feels like im LARPing "documentation as a coding project" but the thing is, none of these are meant for me. they there for my friend claude.

dont take my word for it, i keep asking claude what would make its life easier; 

"You're right that the JSON helps enormously — when I need to answer "which segments have LinkedIn variants?" or "what's the ICP company size range?", I have to scan markdown files instead of querying a structured index."

again, did i write any of these JSON files myself? 

Nope. 

I just keep instructing Claude to catalog these in a format that helps it _retrieve the right pieces without reading entire essays_

Every main folder in the tree has its own .json file that acts as a detailed mini-map claude can use. 

So when I ask "hey which segments could work with this idea" it will do a couple of things: 

- look up the segment details and read the full context of said segment
- look up the tagged sales call analysis docs
- try to find similar items by inferring from entire JSON

I was about to write that it uses "labels" to do so but claude told me nah man, at this stage I just read the whole thing 😄 

just trying to hammer in the point that im not the expert on what claude needs to operate most efficiently, claude itself seems to be. 🙂‍↕️

the TLDR here is if claude needs to filter it - JSON. if a human reads it  markdown.

upside of building clay tables via MCP: you get a lot of things done.
downside of building clay tables via MCP: you can get unwanted things done if you blink too long 😂

lord knows how many versions of "confirm before doing anything" ive written into claude md and still have to hit esc like its a toddler on a run.

literally all of us anthropic cultists the second that a red 500 appears in the terminal window. 

🤣

taplio

The Battle

Lucas Perret

Richard Makara

Followers

Posts

Engagements

Likes

Performance ScoreRichard Makara's LinkedIn Performance

Ready to grow your LinkedIn presence?

taplio

The Battle

Lucas Perret

Richard Makara

Followers

Posts

Engagements

Likes

Performance Score.css-goub6s{font-size:var(--chakra-fontSizes-md);font-weight:500;color:var(--chakra-colors-gray-600);display:block;margin-top:var(--chakra-space-1);}Richard Makara's LinkedIn Performance

Ready to grow your LinkedIn presence?

Performance ScoreRichard Makara's LinkedIn Performance