The DIY AI builder's self-assessment: 9 questions before you start

The DIY AI builder’s self-assessment is a 9-question, 27-point test that decides whether you should ship an AI MVP yourself, ship it with a part-time senior reviewer, ship the prototype yourself and hire out the hardening, or hand the whole project to a partner. Most published frameworks ask whether you can code. That is the wrong question in 2026. Code is the cheap part. Eval thinking, cost reasoning, scope discipline, and handoff timing are the expensive parts — the skills that decide whether your DIY ship is a working product or a six-month rebuild bill.

It sits within The DIY-with-AI manifesto and the broader idea-to-product manifesto. It pairs with 5 founder profiles where it works — that piece tells you who can DIY; this one tells you how ready you personally are.

Table of Contents

How to score yourself in 10 minutes

Read each of the 9 questions below. Score each 0 to 3 against the rubric. Add the scores. Look up your band in the table at the end. Grade like a hiring manager, not like a friend.

Each question covers one competency that breaks a DIY-AI build in a distinct, expensive way. The 2025 Stack Overflow Developer Survey reports 76% of professional developers using AI coding tools daily, with steepest growth among novice and adjacent-to-engineering builders (Stack Overflow 2025). GitHub Octoverse 2025 documents 20M+ AI-assisted developers, fastest growth in solo-founder accounts (Octoverse 2025). This assessment answers which of those builders ship working products and which ship half-built ones they pay a partner to finish.

Question 1: can you read code without writing it

Cursor and Claude Code write most of the code; the founder reviews. If you cannot read what the agent produced, you cannot tell when it is wrong.

The rubric.

• 0 — Never. You cannot tell a function from a variable.
• 1 — With help. You can step through code with an AI explaining each block, but cannot review independently.
• 2 — Slowly. You can read a 200-line file in a stack you have used and tell roughly what it does.
• 3 — Confidently. You can review a pull request in JavaScript, Python, or TypeScript and catch obvious bugs — wrong variable name, missing await, off-by-one, a hardcoded credential.

A 1 is the danger score. Builders who score 1 ship code they have not read, because the agent’s confidence reads as a substitute for review. A 2 is the floor for shipping the project classes named in 5 founder profiles where it works.

Lift the score this week. Pick one open-source repo in your planned stack. Read three files, ask Claude Opus 4.8 or GPT-5 to explain each, then read three more without help.

Question 2: are you fluent with git and deploy

A working AI MVP is a deployed AI MVP. Git is the unbypassed bottleneck.

The rubric.

• 0 — You have never used git on the command line.
• 1 — You can clone, commit, push from a UI like GitHub Desktop, but a merge conflict stops you.
• 2 — You can branch, merge, resolve conflicts, open a PR, and have deployed to Vercel, Render, or Fly.io.
• 3 — You can also rebase, write a CI config, roll back a bad deploy, and use environment variables correctly — you know where secrets live and how they are injected at build time.

Most DIY-AI builds break for the first time at the deploy step. The agent writes the code; the founder discovers the env-var problem at midnight on demo day. A score of 2 is the floor for shipping to anyone outside your laptop. The DIY-with-AI manifesto names the 4-hour terminal tax — git and deploy is where most of it is paid.

Lift the score this week. Deploy a small Next.js app to Vercel through a main-branch deploy. Break it on purpose — push a syntax error, watch the deploy fail, fix it. Five iterations is the threshold.

Question 3: do you think in evals, not features

This is the highest-impact 2026 competency and the most commonly missing one. An eval is a representative input the AI must handle correctly; an eval suite is 80 to 150 of them with a regression gate.

The rubric.

• 0 — You have not heard the word “eval” in this sense.
• 1 — You have heard of evals but think they are optional. The plan is to “test it” by hand.
• 2 — You can describe the difference between deterministic tests and LLM-as-judge evals. You have 20+ inputs drafted.
• 3 — You can wire an eval suite into CI, set a regression threshold, and refuse to ship until it passes. You have read stop scoping AI projects in features, scope them in evaluations and the eval-first build playbook.

A 1 is the most expensive score in the test. It ships a prototype that demos well, fails silently on real inputs, and burns customer trust before the founder notices. Every DIY-AI build needs an eval suite before the first paying user.

Lift the score this week. Write 30 inputs and correct outputs in a Google Sheet. Wire an LLM-as-judge using GPT-5 or Claude Opus 4.8 to score each response. You now have a v0 eval suite.

Question 4: can you reason about cost per query

Token economics are not optional. A DIY build at $0.08 per query × 1,000 queries per day costs $2,400 per month; the same product at $0.40 per query costs $12,000. Either is fine — if you priced for it.

The rubric.

• 0 — You do not know what a token is.
• 1 — You know tokens exist and that input and output are billed separately. You have not modeled cost per query yet.
• 2 — You can estimate cost per query within 30% accuracy for your planned features.
• 3 — You have a spreadsheet that models cost per query across three model choices (Claude Sonnet 4.6 vs. GPT-5 vs. Gemini 2.5 Pro), at three usage scales, with input and output tokens separate. You have read decoding cost per query.

A founder shipping a $20-per-month SaaS to 50 customers needs a per-query cost under roughly $0.04. The 1-score founder ships the product, gets traction, then discovers the unit economics do not work — and the customers are watching.

Lift the score this week. Build one spreadsheet: five features, input tokens, output tokens, cost per query per model, multiplied by planned daily volume. Read decoding AI project TCO for the full cost lineup.

Question 5: do you have scope discipline

The most predictable DIY-AI failure mode is overscoping. The tool feels like a force multiplier, so the founder ships features that were not in the plan.

The rubric.

• 0 — You change your product plan more than once a week.
• 1 — You have a plan but find it hard to say no to a “quick” feature the agent can produce in an afternoon.
• 2 — You have a written PRD with a short feature list, a hard launch date, and you have already killed one feature that did not fit.
• 3 — You cut ruthlessly. You can name three features you will not ship in V1, and you have read the AI build trap and the AI feature scope test.

Scope discipline reads as a character trait but is a learned skill. Cursor and Claude Code make adding a feature feel free. The discipline is the muscle to say “yes, the agent could build that, no, we are not shipping it.”

Lift the score this week. Write a 1-page PRD. Pin it to the wall. Every new idea goes on a separate “V2” page. Keep V1 alive.

Question 6: do you have a security mindset

You will handle API keys, customer data, maybe payments. The agent will happily print a .env value into a log or commit a credential — unless you stop it.

The rubric.

• 0 — You do not know what an API key looks like.
• 1 — You know to use a .env file but have not tested that secrets stay out of git.
• 2 — You use .env, have configured .gitignore, know how secrets are injected at deploy time, and have rotated a key once as practice.
• 3 — You also use a secrets manager (Doppler, 1Password, Vercel env), know what RLS is in Supabase or equivalent, and have read your deploy provider’s threat model.

A leaked OpenAI or Anthropic key on a public GitHub repo will be scraped within hours; the bill is the founder’s. The 1-score builder ships at least one leaked credential in the first month. The 2-score builder ships none.

Lift the score this week. Run gitleaks against your repo. Configure Vercel or Render to inject secrets at build time. Practice rotating a key.

Question 7: do you have 10-15 focused customer-time hours per week

DIY-AI is fast. It is not free. The 1-to-8-week timelines in the profile guide assume focused time.

The rubric.

• 0 — Less than 5 focused hours per week. Day job, family, weekends full.
• 1 — 5 to 10 focused hours, mostly at night. Distractions constant.
• 2 — 10 to 15 focused hours. You can block 3-hour sessions twice a week.
• 3 — 20+ focused hours. Pre-product full-time, or employer’s blessing to use work hours.

A 0 or 1 means the project drifts past 8 weeks without a meaningful release — the profile guide’s explicit failure threshold. Every interruption costs a full re-context-load with the agent.

Lift the score this week. Audit your calendar. Move two 3-hour blocks to a known time on a known day. Protect them. If you cannot find them, you are not ready.

Question 8: do you know when to hand off to a real engineer

Almost no published self-assessment asks this question. It is the most expensive missed signal in the DIY-AI failure pattern.

The rubric.

• 0 — You believe DIY is forever.
• 1 — You will hire eventually but have no rule for when.
• 2 — You have a hand-off rule, e.g., “at 100 paid users” or “at 8,000 lines of code.”
• 3 — You have three named hand-off signals (paid-user count, codebase size, founder’s calendar share spent on engineering), and you know what a clean hand-off looks like: a CLAUDE.md at the repo root, a 2-page architecture document, and a list of the five most fragile parts of the code. You have also read when to fire your AI development agency for the inverse pattern.

The 0-score founder ships V1, gets traction, refuses to hire, and produces a codebase the next engineer cannot inherit. By month 9 the rebuild tax is larger than the original partner engagement would have cost.

Lift the score this week. Pick three hand-off signals. Write them down. Pin them next to the PRD.

Question 9: are you comfortable swapping models

Frontier models shift every 2 to 3 months. Prompts that worked in April stop working in June. The cost-optimal model for one feature is not cost-optimal for another.

The rubric.

• 0 — You will pick one model and never change it.
• 1 — You will switch when the new model is hyped. No data informs the choice.
• 2 — You run your eval suite against the new model, compare scores, and switch only when quality is at parity and cost is meaningfully lower (or vice versa).
• 3 — You also know different features in the same product can run on different models — classification on a cheap model, reasoning on a frontier model — and you have read AI model selection 101.

Model lock-in is a quiet tax. A founder who cannot swap pays 2x to 4x more than necessary at scale and is exposed when a deprecation lands.

Lift the score this week. Pick one feature. Estimate cost per query against Claude Sonnet 4.6, GPT-5, and Gemini 2.5 Pro. Decide which ships and write the reason.

Score interpretation: 0-27 to a decision

Add the 9 scores. Find your band. The recommendation is honest, not flattering.

Score	Band	Recommendation
24-27	Full DIY	Ship it. Set a 4-to-8-week timer, scope ruthlessly. Optional: a 4-hour-per-month senior reviewer as drift insurance.
18-23	Hybrid DIY	Ship it, with a senior reviewer 4 to 8 hours per month. Most 2026 DIY-AI founders land here. See the best idea-to-product partners for solo founders.
12-17	Partial DIY	Build the prototype yourself in 2 to 4 weeks. Hand hardening to a partner. See AI feature pilot vs full MVP cost and risk.
6-11	Full partner engagement	Do not DIY V1. The expected rebuild cost exceeds the partner cost. Start at the idea-to-product manifesto and the AI MVP economics playbook.
0-5	Not now	Score yourself again in 60 days after lifting the lowest 3 scores. The fastest path to ready is the rubric itself.

The two most common landing bands in 2026 are 18-23 (Hybrid DIY) and 12-17 (Partial DIY) — not full DIY. Most founders who think they will DIY end up in hybrid mode by week 4. The honest plan is to start there.

What to do next

The MoFu path from this assessment is to translate the score into a 6-to-12-week plan with the AI MVP Scoping Worksheet — the worksheet maps each score band to a concrete week-by-week scope, an eval threshold, a cost budget, and a hand-off trigger.

If your score is 18 or above, read the DIY-with-AI manifesto and pick a companion guide: Claude Code for non-developers if your stack is terminal-first, what is Cursor if you want an IDE, or the profile guide to confirm archetype fit.

If your score is 11 or below, the next read is anatomy of a $75K AI MVP — it shows where the budget lands in a partner engagement so the comparison to a DIY rebuild is honest.

Frequently Asked Questions

Why 9 questions and not 12 or 20? Nine is the smallest set that covers the failure modes without redundancy. Each question kills DIY-AI builds in a distinct way. Adding more would create noise — readers run out of grading discipline past 10 items.

Is the score deterministic, or do I retake it? Retake it every 60 days pre-launch. The “lift the score this week” task on each question is designed to move a 1 to a 2 in two weeks. Founders who score 11 in March often hit 18 by May with deliberate effort.

Which question matters most? Question 3, the eval-thinking question. It is the most commonly missing and the most expensive miss. A founder who scores 3 on everything else and 0 on evals ships a product that demos well, silently fails on real inputs, and burns customer trust.

Can I outsource the questions I score low on? Partially. Question 1 (code reading), Question 3 (evals), Question 5 (scope), and Question 8 (hand-off) are non-outsourceable — they decide whether you can lead the project. Question 2 (git/deploy), Question 4 (cost), Question 6 (security), and Question 9 (model swapping) can be partially outsourced to a senior reviewer.

Does the score predict whether my AI idea is good? No. The score predicts whether you can ship the idea you already have. For idea quality, see anatomy of a high-confidence AI idea.

What if I score high but I do not have an idea yet? Start with the idea, not the DIY decision. The idea validation playbook covers the front-end work; this assessment is for the build phase only.

How does this differ from a generic “should I be a solo founder” quiz? Generic quizzes ask about resilience, sales chops, and capital. This test asks only about the technical and process competencies that decide whether your DIY build is a working product or a rebuild bill.

Why is “I can code well” not one of the 9 questions? Because writing code is not the rate-limiter in 2026 — reading and reviewing it is. Code generation is the cheap step. The 9 questions are about every other step.

Should I redo the score with a senior engineer beside me? Yes, once. Score yourself, then walk a senior engineer through the same rubric. A 3-point gap on any question is a flag to retake the assessment in 60 days.

Closing

DIY-with-AI is a real route to a shipped product in 2026 — for builders who can read code, deploy a repo, think in evals, reason about cost, scope ruthlessly, treat security as work, find 10+ focused hours a week, plan the hand-off, and swap models without panic.

Most founders score in the hybrid band. The honest plan is to grade yourself, lift the lowest three numbers this week, and then translate the result into a 6-to-12-week scope with the AI MVP Scoping Worksheet. The fastest move from this article is to score yourself in the next 10 minutes, not to read another framework.