Seven questions diagnose whether the partner sitting across from you is going to install a bridge onto your DIY prototype or quietly throw your work away and rebuild from scratch. Ask them in a single 30-minute call. For each question, the answer a real bridge partner gives sounds specific and costable; the answer a rebuild-disguised-as-bridge partner gives sounds reassuring and vague. This piece names the seven questions, what good sounds like, what hand-wave sounds like, and why each question diagnoses the gap.
It follows from the DIY-with-AI manifesto, part of the broader idea-to-product manifesto. It sits next to from Claude Code prototype to shipped product: the bridge and the hybrid model: DIY prototype plus professional rebuild. For a 90-minute version of the same diagnostic across green-field engagements, see a field guide to evaluating an AI agency in under 90 minutes.
Table of Contents
Why the handoff vetting problem is different
Generic AI agency vetting checklists assume a green-field engagement — the founder has a slide deck or a Figma file, and the partner is being hired to build from there. The vetting questions that work in that context (“show me a case study”, “give me a fixed-price quote”, “introduce me to two references”) are necessary but not sufficient when the founder already has working code.
The handoff variant has a specific failure mode the green-field checklists do not catch. A partner can say “yes, we can pick up your prototype” and mean any of three different things:
- Bridge — we install six artifacts (eval set, model abstraction, fallback path, observability, runbook, on-call) and your prototype keeps running. Two weeks, $15–25K.
- Refactor — we keep your business logic but rewrite the scaffolding. Four to six weeks, $40–80K.
- Rebuild disguised as bridge — we throw away your prototype, rebuild from scratch, charge you for the rebuild, and tell you the result is better. Twelve to twenty weeks, $120–200K.
The three engagements look similar on a slide and feel similar in a first call. They are not similar. The founder who pays for option 3 loses the prompts they wrote, the domain knowledge they encoded, the speed of iteration they had, and roughly $100K. The seven questions below diagnose which engagement the partner across the table is actually offering.
BCG reports that 74% of companies struggle to scale AI value beyond pilots — and the cited reasons are operating discipline, not capability (BCG 2024). The 2025 Stack Overflow Developer Survey found 76% of developers use AI coding tools regularly while trust in accuracy fell from 43% to 35% year over year (Stack Overflow 2025). Universal use, falling trust. The partner you pick is the variable that determines whether the prototype-to-product gap closes or you pay twice.
The seven questions at a glance
| # | Question | What it diagnoses |
|---|---|---|
| 1 | How do you assess prototype quality? | Whether the partner has a written intake process or improvises. |
| 2 | What of my work gets thrown out? | Whether this is a bridge or a rebuild. |
| 3 | What does it cost to bridge, artifact by artifact? | Whether the partner can price the bridge or only the rebuild. |
| 4 | Who authors the eval set — you or me? | Whether the partner makes you the long-term owner or the long-term renter. |
| 5 | What is your on-call handoff plan? | Whether “we will support you” is a promise or an artifact. |
| 6 | Who owns the prompts I wrote, after you edit them? | Whether your work product survives the engagement. |
| 7 | What timeline guarantees are in writing? | Whether the calendar is a commitment or a wish. |
Each question is the founder’s right to ask. A partner who is annoyed by them is the partner you do not hire.
Question 1: how do you assess prototype quality?
What good sounds like. “We run a 90-minute intake. We grep your repo for hardcoded model identifiers, count your prompt files, look for any test or eval, check how secrets are handled, run the prototype against three of our standard probe inputs, and give you a one-page diagnostic. We do the same intake on every prototype we take on. The diagnostic is a deliverable you keep — it is yours whether or not you hire us.” Bonus: they name a fixed price for the intake itself ($1–2K), separate from the bridge engagement, so the assessment is decoupled from the sale.
What hand-wave sounds like. “We will take a look and see what you have. Don’t worry about it — we are experienced; we will figure it out.” Translation: no written process, no diagnostic artifact, no fixed intake price. The partner is starting the engagement with sunk-cost framing, which is how rebuilds get justified later.
Why this diagnoses. A partner who has done this work before has a checklist. The checklist exists because every prototype has the same set of questions to answer — model abstraction, secrets, eval state, fallback behavior, on-call. A partner without a checklist either has not done this work before or is improvising, and improvisation on intake is where the rebuild justification starts (“once we got in there, it was worse than we thought”). Compare to the field guide to evaluating an AI agency in under 90 minutes, which uses the same intake-discipline test for green-field engagements.
Question 2: what gets thrown out?
This is the single most diagnostic question in the seven.
What good sounds like. “Nothing on day one. We add a model abstraction layer around your existing prompts, but the prompts themselves we treat as your spec. We may rewrite individual prompts later with your sign-off on each one, but we do not rewrite without showing you the diff and the eval-score delta. The application code, the routes, the data shapes — those stay. We are adding the six bridge artifacts, not replacing your work.”
What hand-wave sounds like. “We will start fresh with a clean architecture — your prototype gave us a great spec, and now we will build it right.” Or: “We will use your prototype as a reference but rewrite the parts that need to be production-grade.” Translation: rebuild. The “production-grade” framing exists to make the rebuild sound like a virtue. You will pay for the rebuild and lose the iteration speed you had.
Why this diagnoses. A bridge is additive. A rebuild is substitutive. The partner who answers “nothing gets thrown out” is offering a bridge — six discrete additions, your prototype intact underneath, your prompt edits propagating through the system because the prompts are still the source of truth. The partner who answers “we will rewrite from a clean foundation” is offering a rebuild. The rebuild may even produce a better product. It will cost five to ten times what the bridge would, and you will lose the speed advantage that made DIY worth doing in the first place. See the 80-20 of DIY AI MVPs: what to keep, what to throw out for the founder-side version of the same question.
Question 3: what does it cost to bridge?
What good sounds like. “We charge by bridge artifact. Eval set: $2K. Model abstraction: $1K. Fallback path: $1.8K. Observability: $3K plus vendor pass-through. Runbook: $1K. On-call handoff: $0.7K (the schedule and pager setup; the human time is separately retained). Total: $15–25K depending on your repo’s starting state. We will give you a written quote within 48 hours of the intake, fixed price, no overruns billed to you.”
What hand-wave sounds like. “It depends. Probably somewhere in the $50–200K range. We will know more once we get in there.” Translation: the partner cannot price the bridge because the bridge is not what they do. They price rebuilds, and the price of a rebuild scales with the size of the founder’s checkbook.
Why this diagnoses. Per-artifact pricing is the structural test for whether the partner has done this work before. Bridge artifacts have known scope (each is between 4 and 16 senior-engineer hours; see from Claude Code prototype to shipped product: the bridge for the hour table). A partner who has done this work before has a price sheet because they have done it 10+ times. A partner who has not has no price sheet because each engagement is a discovery exercise — and discovery exercises are how scope creeps. For the cost framing across pilot-vs-MVP-vs-rebuild, see AI feature pilot vs full MVP: cost and risk comparison.
Question 4: who authors the eval set?
What good sounds like. “You do, with our editing. We will give you a template — 20–50 input cases sliced into easy, hard, adversarial, with an expected output for each. We will sit with you for one or two sessions, write 5 examples together, then you write the rest with our review. The eval set is your spec for what ‘correct’ means; we cannot author your spec for you. We can engineer everything else around it, but the spec has to be yours.”
What hand-wave sounds like. “We will handle the eval set. We have an internal evaluation framework we apply to every engagement.” Translation: the partner is authoring the spec for you. The spec is now their property, not yours. Six months later, when you want to switch partners or bring the work in-house, the eval set goes with them and you start over.
Why this diagnoses. The eval set is the founder’s most valuable long-term asset. Prompts change weekly, models deprecate every 12–18 months — the eval set is what survives. A partner who authors it for you has captured the most durable artifact of the system on their side of the engagement boundary. A partner who insists you author it has made you the long-term owner of the correctness contract, which is the only sustainable handoff posture. Anthropic’s “Building effective agents” makes the eval primacy claim directly (Anthropic, Dec 2024). The point is not philosophical; it is operational. See also how to evaluate an AI vendor’s eval discipline before signing.
Question 5: what is the on-call handoff plan?
What good sounds like. “Week 1: we are on the pager 24/7 while we install observability. Week 2: we share the pager with you — you take business hours, we take nights. Week 3 onward: you are on the pager, we are a paid retainer for escalations. The paging tool is PagerDuty or Opsgenie; the SLA is 15-minute response for P0, 1-hour for P1; the retainer is $2K/month for two named engineers with 4 hours of escalation included. The on-call doc is a markdown file in your repo, version-controlled, and we update it together at the end of every escalation.”
What hand-wave sounds like. “We will support you after launch. Just reach out if anything breaks.” Translation: no schedule, no paging tool, no SLA, no retainer math. “Support” is an unfunded promise the partner walks away from in two months.
Why this diagnoses. On-call is an artifact, not a feeling. The artifact has four parts: the schedule (who is on the pager when), the paging tool (PagerDuty, Opsgenie, or equivalent), the SLA (response time for each severity), and the rotation cadence (when responsibility shifts from partner to founder). A partner who has shipped to production has run on-call rotations and knows how each part works. A partner who has not handles on-call as an emergency invoice item.
Question 6: who owns the IP between phases?
What good sounds like. “You own everything you wrote and everything we wrote on your engagement. We will sign a one-page work-for-hire clause stating that all code, prompts, eval cases, runbooks, and infra-as-code produced under this engagement is your property, transferred on payment. The prompts you wrote in Claude Code stay yours; the edits we make to them are also yours. The only thing we retain is the right to reuse non-client-specific frameworks (e.g., our standard observability schema, our intake checklist) on other engagements.”
What hand-wave sounds like. “Our standard MSA covers IP.” Translation: read the MSA. If the MSA says the partner retains a license to “derivative works” or “improvements”, your edited prompts are now partially theirs. The cost of unwinding this is six figures of lawyer fees and a fork of your own code.
Why this diagnoses. The prompts you wrote in Claude Code are domain knowledge encoded as natural language. They are the most concentrated artifact of your founder-product fit. A partner who edits them and then claims any share of the result has captured your founder-product fit on their side of the engagement boundary. The vetting question is the existence of a written, signed, one-page IP clause that names prompts and eval cases explicitly. Standard MSA language is not enough. For the contractual framing of this across the broader agency landscape, see the 7 commitments every AI dev agency should make in writing.
Question 7: what timeline guarantees are in writing?
What good sounds like. “Two weeks for the bridge, calendar-week-precise. Week 1 deliverables: model abstraction, eval set, fallback path. Week 2 deliverables: observability, runbook, on-call handoff. If we miss the calendar by more than three days, you get a 10% credit against the next engagement or a refund of the missed week’s invoice; your choice. We have hit this calendar on 12 of the last 14 engagements; the two we missed were both vendor-onboarding delays we now build into the quote.”
What hand-wave sounds like. “We typically deliver in 4–8 weeks.” Translation: no written calendar, no credit clause, no track record. You will be told at week 6 that “production-readiness took longer than expected” and the invoice will be 60% larger than the quote.
Why this diagnoses. Calendar precision is the signal that the partner has done this work enough times to know how long it actually takes. The credit clause is the signal that the partner is willing to bet their margin on their estimate. A partner who refuses a credit clause does not trust their own calendar; you should not either. For the broader framing across project pricing models, see AI MVP pricing explained: fixed price vs hourly vs milestone.
Running the interview
The seven questions take 25–30 minutes to ask if the partner answers crisply. They take 50–60 minutes if the partner is improvising. Either way, ask them all in one call. Take notes — paper, a phone notes app, whatever you prefer. The notes are the artifact you take to a second partner for the same interview.
Score the answers on a 3-point scale: 2 for “what good sounds like” or close, 1 for hedged-but-substantive, 0 for hand-wave. Out of 14 possible points:
| Score | Interpretation |
|---|---|
| 12–14 | Hire pending references. |
| 9–11 | Hire pending a second 30-minute call to deepen the hedged answers. |
| 6–8 | Pass. The gaps will become invoices. |
| 0–5 | Pass and reconsider whether the partner has done this kind of work at all. |
Two-partner comparison is the right baseline; three is the maximum a founder running this process solo should attempt. Beyond three, the cost of running the interviews exceeds the marginal value of more options.
If you want to apply the same vetting discipline upstream — to the partner you might hire before you start a DIY prototype — see the DIY vs hire decision framework for AI MVPs. The seven-question script here applies to the case where DIY has already happened and the partner is being scoped against working code.
Frequently Asked Questions
Q: Do I need to be technical to run this interview?
No. The seven questions are designed for a non-engineer founder. The partner does the engineering explaining; your job is to listen for whether the answers are specific or vague. The “what good sounds like” examples are the calibration set. If an answer feels close to those examples, mark a 2. If it feels close to the hand-wave examples, mark a 0.
Q: What if the partner refuses to answer in a single 30-minute call?
That is the answer. A partner who has done this work before will give crisp answers because they have given them before. A partner who insists on a longer discovery process before answering is either inexperienced or trying to amortize their sales cost across a longer cycle. Pass.
Q: What if all my partner candidates fail the vetting?
Three possibilities. First, you are talking to the wrong supply pool — generic dev shops rather than partners who specialize in AI prototype handoff. Second, you may be ready for a different engagement model entirely; see the hybrid model: DIY prototype plus professional rebuild. Third, you may not yet be ready for a partner at all — the prototype itself may need 1–2 more weekends of founder work first.
Q: Should I ask these questions in writing or live?
Live, in a single call, with notes. Written answers give the partner time to wordsmith and obscure the hand-wave. Live answers reveal whether the answers are practiced or improvised. Send the questions in advance if you want, but require live answers.
Q: What references should I ask for, in addition to the seven questions?
Two specific kinds. First, a reference from a founder whose prototype the partner installed a bridge on — not rebuilt — within the last 12 months. Second, the eval set from a previous engagement, with the previous client’s permission, to verify the partner’s eval discipline is real. Generic “happy customer” references do not diagnose this engagement model.
Q: Is the bridge engagement always two weeks?
Two weeks is the calendar for a single-feature prototype (one prompt, one or two tool calls, one endpoint). It scales linearly with feature count and superlinearly with agentic step count. A three-feature prototype with multi-step agents is four to six weeks. Ask the partner to scope the calendar to your repo specifically; “two weeks” is a default, not a quote.
Q: What if I am talking to a partner who quotes $5K total for the bridge?
Suspicious on the low end the same way $200K is suspicious on the high end. $5K does not buy six senior-engineer artifacts; it buys a junior engineer two days of effort. The fair price band is $15–25K with a senior US/EU engineer or $4–7K with a senior LATAM or Eastern European engineer. Outside that band in either direction, ask for the artifact-by-artifact cost decomposition.
Q: How does this interview change if I built my prototype in Cursor or Lovable, not Claude Code?
The seven questions are tool-agnostic. The partner’s answer to question 1 (prototype assessment) should mention your specific tool by name — if they default to “we will look at the repo” without acknowledging the tool, that is a small signal of unfamiliarity. Lovable prototypes occasionally need an export-to-code step before the bridge can be installed; ask the partner to price that step explicitly.
Q: What is the right next step after the interview?
Pick the highest-scoring partner. Ask for the written intake quote (question 1) and the written per-artifact bridge quote (question 3). Run the intake with that partner as a paid engagement, separately from the bridge commitment. The intake deliverable is yours regardless. If the intake report matches the partner’s interview answers, proceed to the bridge. If it diverges, you have learned what you needed to learn for $1–2K.
Closing
The seven questions are short because the failure mode they catch is structural, not subtle. A partner who is good at bridging DIY prototypes has done it before, has a checklist, has a price sheet, has a calendar, has signed an IP clause that names prompts explicitly, and has a written on-call schedule with a paging tool and an SLA. A partner who is bad at it has none of those things and will tell you so under questioning. The 30 minutes you spend asking is the cheapest 30 minutes you will spend on the engagement.
If you want a one-page idea review against your DIY prototype before you start interviewing partners, book an idea review with SFAI Labs. The review covers the same intake question the seven-question interview opens with, and the deliverable is yours whether or not you continue with us.
Arthur Wandzel
