Version 0 is a hypothesis we can test with three conversations. It is not a locked ICP, a pricing decision, or a claim that the product is ready to sell.
Prepared from the mentor brief of this week, the company materials, and public competitor pages. Labels below mark what is known, what was researched, what is assumed, and what is a recommendation.
EvidenceResearchAssumptionJudgment
This week
The move
Send two learning messages. Do not pitch a memory platform, book a demo, or ask anyone to connect a repository.
Judgment The public site already says “engineering knowledge for AI agents” and invites a demo. The sales materials say the opposite: these are founder-led learning conversations, and the product should not be oversold. The conversations win. A demo ask would test politeness, not the problem.
Evidence Every discovery hypothesis in the sales bank is still unvalidated. There are no recorded customer calls. The only concrete repeated tasks we have are from the founders’ own work, including a set of Tenderly how-tos.
A good conversation produces one recent example: what repeated, what the person did instead, and who else felt it. “Cool idea” is not a result.
Design partners
Three conversations, not three logos.
Two paths are specific enough to act on. The third person has to come from Igor’s own contacts. Inventing a company would fake a relationship.
1. A teammate at SMG Serbia
Research Igor works at SMG Serbia and is described publicly as the internal AI champion: evaluating AI tools and helping others use them.
Judgment This is the fastest place to hear a real repeated how-to, because he already sits inside the workflow. It is a learning conversation, not a design-partner pilot. Swiss Marketplace Group is a large company. Do not connect company code, Slack, or tickets, and do not treat an employer as a logo.
Who. The engineer Igor has already helped more than once when an agent lacked local context. The name is not in the company materials.
Why them. They can describe last month, not a hypothetical. If they cannot recall a repeated how-to, the problem statement is weaker than the internal docs claim.
What to learn. What they did instead — asked someone, searched Slack, re-prompted, or started over — and what they would refuse to share with any outside tool.
Assumption A casual conversation about his own teamwork is acceptable alongside his job. If it is not, skip this one. That choice is his.
2. One former Tenderly engineer, through Milan
Evidence The product note uses Tenderly tasks as the shape of the wedge: recover node-subscriber after a reorg, run tnode locally, point tenderly-server at a local agent, skip a block, bump a protocol version. The founding story is a day lost because the answer already existed in a private Claude session.
Research As of 2026 company profiles, Tenderly is on the order of 80–90 people, with engineering around 30–34, concentrated in Belgrade. That is small enough for a conversation and large enough that how-tos recur.
Who. Not the CEO, and not a cold email. Milan introduces one person who still runs those operational tasks.
Why them. This is the only external team where the repeated work is already written down. If the how-tos no longer hurt, the examples are history, not a market.
What to learn. Do those tasks still get re-solved? Where does the answer live today? Who would have to agree before any tool conversation?
Assumption Milan still has a warm path. If he does not, drop Tenderly rather than forcing an outreach.
3. One founder or head of engineering Igor already knows
Judgment The third conversation should be outside both SMG and Tenderly, so two companies cannot be mistaken for a market. Profile: a product company, roughly 15–80 engineers, people already coding with Cursor, Claude Code, Codex, or Copilot, and a buyer who can take a call without a security review.
Not an agency, not a team that wants “enterprise search,” and not someone Igor has never spoken to. Heapspace and Startit are places he already appears; they are a room, not a customer.
Who. Unchosen. The name has to be someone he can message this week.
What to learn. The same four facts as the other two calls, so the three notes can be compared.
Outreach
What to send
The message asks for one past example. It does not describe Nucipal.
To the SMG teammate
Ćao [Ime],
treba mi 20 minuta tvog iskustva, ne za alat.
Seti se poslednjeg puta kada si radio nešto što je neko kod nas već uradio, a agent to nije znao — lokalni setup, incident, neki interni korak. Šta si tada uradio: pitao kolegu, tražio po Slacku, ili krenuo ispočetka?
Hoću da čujem taj jedan primer. Ne šaljem ti ništa da probaš i ne tražim pristup kodu.
Igor
To Milan, for the Tenderly intro
Milan, treba mi jedna intro, ne pitch.
Hoću da pričam sa jednom osobom sa Tenderlyja koja i dalje radi one operativne how-to stvari: lokalni tnode, node-subscriber posle reorga, skip bloka, bump protokola. Cilj je da čujem da li se to i dalje radi ponovo i gde odgovor danas živi.
Možeš li da me spojiš sa nekim kome je to normalno pitanje? Neću pisati CEO-u i neću tražiti pristup sistemima.
Ako hoćeš, prosledi ovo:
“Igor hoće 20 minuta o jednom konkretnom primeru: poslednji put kada si ponovo sklapao how-to koji je neko već znao. Ne prodaje alat i ne traži pristup kodu.”
Use the same shape for the third person once they have a name. Change only the specific task, if you know one. Leave the product out.
01 · Segment
Who this month’s calls are for
Product engineering teams of roughly 15–80 people who already use an AI coding tool every day, and where one person can agree to a conversation this month.
Evidence Internal materials disagree with each other. The sales one-pager says 50–100 engineers. The design-partner playbook says 15–100, and suggests a first partner around a 70-person B2B company, reached first through a warm Balkan conversation. The lean canvas still marks customer segments as undecided. None of those ranges has been checked with a customer.
Judgment Keep 50–100 engineers as a later commercial hypothesis. Do not use it as the filter for this week. The only team whose repeated tasks we can name, Tenderly engineering, sits nearer 30 people. A team that needs a quarter to approve a call cannot teach us anything before the next check-in.
Industry is not the cut. The cut is daily AI coding, recurring operational how-tos, and a short path to a yes. Crypto infrastructure is where the examples come from. It is not, by itself, the market.
02 · Buyer and user
Start with the person who repeats the work.
The user is the engineer who hits the same how-to again. The buyer is unproven. In a team this size they are often the same person, or one introduction apart.
Evidence The sales guide lists a possible buyer as CTO, VP Engineering, head of platform, head of developer productivity, or an AI-tooling champion. That list is a hypothesis, explicitly unvalidated. The August note from Milan defines the first success as personal: he stops repeating himself, and the same mechanism later stops a company repeating itself.
Judgment For these calls, the champion is whoever already gets interrupted. At SMG that may be Igor, which is useful evidence and a poor sales motion. Do not build the story around a VP Eng purchase. We have not seen a budget, a security review, or a champion who is not a founder.
Role
What we believe
Status
User
Engineer redoing a known how-to, or correcting an agent
Supported by founder stories, not by outside customers
Champion
The person teammates already ask
Recommendation
Buyer
Unknown. Possibly the champion at this company size
Assumption
Blocker
Whoever owns code and Slack access
Assumption, from the security notes
03 · Problem
The answer dies where it was found.
A recurring how-to gets solved in a private AI session, a Slack thread, or someone’s head. The next person, or the next agent run, solves it again.
Evidence The written founding story is a full day lost debugging a service that would not run locally. A teammate had the answer in a private Claude session and mentioned it only after the day was gone. Milan’s later note narrows v1 to that class of task: low-novelty, weekly or monthly, already solved by someone on the team. His own prompt history was measured as heavy with requests a stored how-to could have answered. That measurement is internal. It is not a customer study.
Judgment Lead with that narrow job. An earlier answer in the company Q&A says the MVP is onboarding new engineers. Onboarding can be one example of the same job. It is a worse opener, because a team can nod at onboarding without having felt it this month. “All engineering knowledge is fragmented” is also too wide. It makes Nucipal sound like search.
The problem is not yet priced. We do not know hours lost per engineer. Do not quote the website’s industry figures in a call.
04 · Value proposition
Put the how-to back in the agent.
When someone on the team has already figured out the how-to, the next person should not have to figure it out again. Nucipal keeps that answer and returns it inside the coding agent they already use.
Evidence This matches the wedge sentence in the strategy notes and the product goal in nucipal-core: capture reasoning from AI coding sessions and make it shared. The public homepage is broader. It promises reasoning linked to code, tickets, docs, logs, and discussions.
Judgment Say the narrow sentence. Do not promise, in these conversations, cross-tool citations, permission mirroring, or time saved. Those are on the site. They are not what a first call can prove. The value is still a hypothesis until someone describes a repeated task and says the missing piece was the prior answer, not a smarter model.
05 · Positioning
Beat the workaround, not a category.
The real alternative is what a team does today when the agent does not know the local how-to. Named software competitors matter only after a buyer puts them on a list.
Research April Dunford’s framing, which the mentor sent: competition includes the status quo, a pile of tools, a manual habit, or doing nothing. Positioning against every company that could someday overlap makes the story vague. She also argues that an AI-generated competitor list is a weak input until sales calls show who actually shows up.
Evidence Our own problem note already lists the workarounds: ask on Slack, search a stale doc, grep old pull requests, or re-derive the answer.
Alternative
When it is real
How to talk about it
Re-derive it, or re-prompt the agent
Default. The agent sees the repo, not the last correction.
This is the thing Version 0 has to beat.
Ask a teammate
Works, and spends a senior person’s attention.
Ask how often, and who gets interrupted. Do not insult the habit.
Slack, a doc, an old PR
The answer is somewhere, and nobody trusts it is current.
The gap is recall at the moment of work, not another wiki.
Unblocked
Only if they are already trying it or comparing it.
Do not open with them. See the note below.
Glean, a wiki, an internal RAG bot, Pieces, Mem0
Not on a shortlist we have ever seen.
Leave them out of the pitch. Watch the calls.
Research Unblocked’s own reference, updated 17 September 2026, calls itself “the context layer for agentic software development.” It connects code, chat, tickets, docs, and production tools; offers MCP into Claude Code, Cursor, Copilot, and Codex; cites sources; starts at $29 per user per month billed annually, or $35 monthly; and runs a 21-day trial. Public customers include teams the size of Fetch, TravelPerk, and Cribl. It is SOC 2 Type II. Our June profile said they index existing artifacts and had backed away from ingesting AI-generated answers. Their current public description does not say they capture coding-session transcripts. Treat that June point as stale until someone checks.
Judgment Nucipal should not introduce itself as the agent-memory category. Unblocked already sells a broader version of that sentence, with a trial and security proof we do not have. The difference worth testing is narrower: the answer created inside the coding session, especially the correction, which docs and pull requests often never contain. That difference is a product bet. It is not yet a reason a buyer has used to choose us.
For a team already coding with an AI agent, Nucipal is the shared memory of how this team does recurring work. Unlike Slack, a stale doc, or another prompt against the repo alone, it returns the answer someone already worked out.
06 · Channel
Warm, and only warm, until the job repeats.
This week’s channel is two messages Igor can actually send, plus a third name from people he already knows. Not ads, not a cold list, not the website form.
Evidence The outbound notes say discovery through warm contacts, and they postpone cold outreach until the pain has been heard more than once. The design-partner playbook ranks channels the same way: warm intros, then the Balkan network, then careful LinkedIn. It tells us to skip ads, SEO, conferences, and webinars for now.
Research The Fastlane interview the mentor sent is about volume later: on the order of twenty customer calls a week, notes on past behavior, and changing the product from what people already did. That volume is a destination. It is not the assignment for this check-in. The useful part now is the method. Ask what happened last time. Write it down. Do not ask “would you use this?”
Judgment Igor sends. Milan joins only if a second conversation becomes technical. A meetup talk is a way to meet the third person, not a launch.
07 · Conversion
The yes is a second conversation with a named task.
Call one is 20–25 minutes. Advance only if they can point to a repeated how-to from the last month, name the workaround, and suggest who else feels it.
On the call, ask only this:
When did this last happen?
What did you do?
Who did it interrupt, and how long did it take?
Where would the answer have had to appear for you to use it next time?
Judgment Do not show a deck. Do not discuss price. The sales guide’s “design partners are free” line, and the playbook’s later paid-pilot ranges, are both untested. Price is a later conversation. The mentor’s missing sales-call video can wait until there is a real situation to review.
Evidence The internal path already expects most first calls to die. A rough planning ratio in the design-partner path is about ten first calls to one design partner. Two or three conversations this week are the start of that, not the close of it.
A second call can sketch one workflow. Still no repository access until they ask, and until we can say plainly what would be read, what would be stored, and that we cannot offer SOC 2.
08 · Evidence loop
Three notes, then a decision.
After each call, write five lines. After three calls, keep or change the problem statement. Do not add a feature in between.
Write down
Why it matters
The task, in their words
If it is not a repeated how-to, the wedge is off
When it last happened
Older than a month is a story, not a habit
The workaround
This is the real competitor
Who else felt it
One annoyed person is not a buyer
What they will not share
Sets the first safe scope, if there is one
Judgment Change Version 0 if five conversations produce no repeated how-to from the last 30 days, or if every serious team asks for company-wide search or a security certification before they will describe the work. Keep going if the same task shape shows up in more than one company and the workaround is a person or a re-prompt.
The public site, the 50–100 ICP line, and the “agent memory category” sentence stay as they are until those notes exist. This brief does not rewrite them.
Out of scope
Who not to chase yet
Teams that are not already using an AI coding tool. The session-memory bet has nothing to attach to.
Agencies. Client code makes the data question worse, and the playbook already flags them.
Anyone who needs SOC 2, SSO, or on-prem before a first working session. Unblocked sells that. We should not pretend to.
Buyers shopping for a company wiki or horizontal search. That pulls us into Glean’s conversation.
A long enterprise cycle, including a pilot that would require SMG corporate approval. Learn from a teammate if appropriate. Do not make the employer the first customer by default.
Needs Igor
Three decisions this brief will not invent
Is a learning conversation with an SMG teammate acceptable, or is the employer off limits?
Will Milan introduce one Tenderly engineer, and who?
Who is the third person, outside those two companies, that you can message this week?
Also unread, so not used: the Fathom recording of the session is behind a login, the Notion research prompt is in a workspace this account cannot open, and the customer-profile and business-model attachments were not in the email text. The optional personality tests and the Ominimo interview were not needed to make this call. If the live session fixed a different segment than the one above, that correction should replace this page.