A well-run idea-to-PRD sprint produces 7 named artifacts in 14 days. If a founder cannot count those seven files in their drive by day 14, they did not buy scoping — they bought consultancy theater. This article walks the sprint day by day, names the artifact each phase produces, the founder’s job inside it, and the failure mode if any phase is skipped.
It extends the idea validation playbook within the idea-to-product manifesto. The playbook frames the discipline; this article walks the calendar.
Table of Contents
- Why the 14-day sprint exists
- The 7 artifacts a serious sprint produces
- Phase 1 — Days 1-2: Kickoff and idea compression
- Phase 2 — Days 3-5: Capability probe
- Phase 3 — Days 6-8: Task taxonomy and eval design
- Phase 4 — Days 9-11: PRD synthesis
- Phase 5 — Days 12-14: PRD review, lock, and go/no-go
- What a founder should refuse to sign on day 14
- Frequently Asked Questions
- Closing
Why the 14-day sprint exists
The pre-build stretch is the cheapest place to be wrong. McKinsey’s State of AI in early 2025 shows broad adoption of AI tools (about 78% of organizations report using AI in at least one function) alongside narrow capture of measurable EBIT impact — pilots stall between idea and production because the artifact that defines “what we are building” was never tight enough to ship against.
Two weeks is enough to do the work that, if skipped, eats six to twelve weeks of build. The shape of a serious 14-day sprint is the same regardless of vendor: five phases, seven artifacts, one go/no-go gate at day 7.
The 7 artifacts a serious sprint produces
By day 14, the founder should be able to send these seven files as attachments. Each has a definition tight enough to mark present or missing.
| # | Artifact | Produced by | Day landed |
|---|---|---|---|
| 1 | One-sentence idea + ICP brief | Founder + partner | Day 2 |
| 2 | Capability-probe results (40 cases) | Partner-led, founder-graded | Day 5 |
| 3 | Task taxonomy (5-15 named tasks) | Partner-led, founder-signed | Day 7 |
| 4 | Eval set (50-150 cases, rubric, threshold) | Founder-led, partner-shaped | Day 8 |
| 5 | Failure-mode budget | Founder-led, partner-shaped | Day 8 |
| 6 | The PRD itself (10 sections) | Partner-led, founder-signed | Day 11 |
| 7 | SOW addendum + acceptance criteria | Partner-led, founder-signed | Day 14 |
If any one is missing on day 14, the engagement should not roll into build. The artifact-count test is the simplest filter a non-engineer founder has against vague scoping.
Phase 1 — Days 1-2: Kickoff and idea compression
Artifact produced: a one-sentence idea plus a one-page ICP brief.
Days 1 and 2 are a compression exercise, not a discovery workshop. The founder walks in with a paragraph or three sticky notes; the partner walks in with a structured intake form. Both leave with a single sentence and a single page.
The one-sentence idea names user, trigger, desired outcome, and existing alternative. “Solo independent consultants who currently triage email manually want their inbox prioritized within five minutes of opening their laptop” is an idea. “An AI email assistant” is a category.
The ICP brief is one page: who the user is, their current workflow, what they pay for today, what they would pay for the proposed product, and the evidence the founder has that they would actually buy. Three to five customer-discovery interviews behind the brief is the floor.
Founder’s job in Phase 1: bring the paragraph. Sit through the compression and resist the urge to keep the idea broad. Accepting one sentence is the first scope decision of the project.
Failure mode if skipped: the partner builds against an inferred idea, the founder reviews against a different one, and the mismatch surfaces in week 6 of build instead of week 1 of scoping.
Phase 2 — Days 3-5: Capability probe
Artifact produced: capability-probe results on 40 representative inputs.
Phase 2 is the cheapest place to be wrong. Before any task taxonomy or eval rubric is drafted, the partner runs the central AI capability against 40 representative inputs on a current frontier model — typically Claude Opus 4.8, GPT-5, or Gemini 2.5 Pro from the current Artificial Analysis leaderboard. Inputs are sampled from the founder’s actual data or realistic constructed cases. The founder grades the outputs “useful or better” or “not useful.”
The result is a single number — for example, 33 of 40 useful or better. That number is the ceiling reference for the rest of the sprint.
The probe answers a question every other phase assumes: can a current model actually do the central thing the product depends on? The AI feasibility check walks the probe design in detail. A probe of 8 of 40 says narrow the scope or stop. A probe of 38 of 40 says the AI layer is not the defensibility — workflow, integrations, or distribution will need to be.
Founder’s job in Phase 2: produce or approve the 40 inputs and grade the outputs personally. The grading cannot be outsourced — the founder is the only honest grader on the team. Two to four hours of focused grading.
Failure mode if skipped: the team drafts a PRD against an unverified capability. Six to ten weeks of build later, the central feature is below ceiling and the project ships a half-product or pivots mid-build.
Phase 3 — Days 6-8: Task taxonomy and eval design
Artifact produced: a task taxonomy (5-15 named tasks), an eval set (50-150 cases with rubric and threshold), and a failure-mode budget.
Phase 3 is where most discovery engagements quietly skip the load-bearing work. The partner breaks the product into 5 to 15 named tasks — not features the user clicks, but units of work the model is responsible for. A 5-row taxonomy looks like:
| ID | Task name | Input | Output | User-visible? |
|---|---|---|---|---|
| T01 | Priority classification | Email body + 30-day context | Priority bucket 1-5 | Yes (label) |
| T02 | Rationale generation | Email body + assigned bucket | One-sentence rationale | Yes (hover) |
| T03 | Confidence calibration | Email body + bucket + context | Confidence 0.0-1.0 | No (internal) |
| T04 | Escalation decision | Confidence + user policy | Escalate yes/no | Yes (modal) |
| T05 | Action logging | All of the above | Structured log entry | No (audit only) |
Each task gets its own eval cases. The AI PRD explained walks the taxonomy section in detail.
The eval set sits in version control and ships as part of the PRD. Working size at sprint time: 50 to 150 cases total, a 1-to-5 rubric per task, and a pass-rate threshold per task (“75% of T01 score 4 or 5, zero scores of 1 on the production-critical subset”).
The failure-mode budget answers the question the threshold does not: which failure modes is the founder willing to ship with, at what rates, and with what policy when exceeded? Confident misclassification on a critical subset: 0%. Unwarranted refusal: 5%. Hallucinated rationale: 1%. Latency over five seconds: 5% of requests. The budget forces the founder to distinguish failure modes that matter from those that do not before code is written.
Founder’s job in Phase 3: produce the eval cases. The partner can shape rubric and threshold, but the cases themselves are domain-expertise the founder owns. Eight to twelve hours, split across two or three sessions.
Failure mode if skipped: the build ships against an implicit “perfect” standard, which guarantees a fail. Every quality complaint becomes a vibe argument because nobody signed up to a measurable threshold.
The day-7 go/no-go gate sits at the end of this phase. After the capability probe and the task taxonomy are landed, the founder has enough data to walk away or to commit to PRD synthesis. A serious sprint contract makes this a hard checkpoint.
Phase 4 — Days 9-11: PRD synthesis
Artifact produced: the PRD itself — a 10-section document with eval contract.
Phase 4 assembles the prior artifacts into the PRD. A serious 2026 AI PRD has ten sections: one-sentence idea, AI capability dependency, task taxonomy, eval set, failure-mode budget, no-AI fallback, integration surface, observability minimum, acceptance criteria, and what is explicitly out of scope. The AI PRD explained walks each section.
The partner drafts on days 9 and 10. Day 11 is review: the founder reads start to finish, marks every section agree / agree-with-changes / do-not-agree, and the partner reconciles.
Three sections particularly bear repeating. The no-AI fallback names what the product does when the model fails or is unavailable. The integration surface names every external system with auth model and rate-limit policy. The explicit out-of-scope list over-lists rather than under-lists — the cost of an unnecessary row is zero; the cost of an unstated assumption is a 4-week argument in week 8 of build.
Founder’s job in Phase 4: read the full PRD on day 11. Not skim — read. Every assumption the founder lets pass is one the build runs against.
Failure mode if skipped: the partner ships a PRD nobody read end-to-end. The first time the founder reads it carefully is when something has gone wrong in build — at which point the conversation is adversarial.
Phase 5 — Days 12-14: PRD review, lock, and go/no-go
Artifact produced: the SOW addendum with acceptance criteria — the contractual layer that points at the PRD.
The final three days lock the PRD and produce the SOW addendum that points at it. Acceptance criteria stop being a feature checklist (“user can sort by date”) and become an eval-threshold contract (“AC01: 75% pass rate on the 40-case T01 suite, zero score-1 on the critical subset, by build week 4”).
Day 12 is a final founder read with a different lens: the SOW lens. If I had to litigate this, what is the minimum contractual definition of “delivered”? Every eval threshold and failure-mode constraint becomes a row in the acceptance-criteria table.
Day 13 aligns the build estimate (engineer-weeks, cost, calendar) against the PRD scope. Day 14 is signature.
A founder facing an estimate that does not match the PRD scope should not sign. The right move is to re-scope the PRD, re-shape the estimate, or use the day-7 go/no-go data to decline. The anatomy of the first 14 days of an AI agency engagement walks the engagement-side mechanics.
Founder’s job in Phase 5: sign or decline. If the seven artifacts are present, the PRD is complete, and the estimate matches the scope, sign. Otherwise decline.
Failure mode if skipped: the SOW points at “the PRD” without acceptance criteria. The vendor interprets “delivered” loosely; the founder discovers in week 8 of build that the contract did not say what they thought it said.
What a founder should refuse to sign on day 14
A short list, blunt for a reason.
- A PRD without an eval set. If task taxonomy and eval rubric are missing or hand-wavy, the build is unmeasurable.
- A PRD without a failure-mode budget. Implicit “perfect” guarantees fail.
- A SOW that does not point at a specific dated PRD. The PRD must be the contractual exhibit.
- Acceptance criteria written as feature checklists. “User can click Save” is not acceptable for an AI feature.
- A capability probe missing or below threshold without explicit re-scope. A probe of 18 of 40 with a PRD written anyway means the partner is selling hope.
- A build estimate that does not match the PRD scope. Either the estimate is short or the PRD is overscoped — reconcile before signature.
- Fewer than seven named artifacts in the drive. Count them. If the count is six, ask what’s missing.
Frequently Asked Questions
How long should the idea-to-PRD sprint actually take?
Two weeks of calendar time, with 30 to 50 hours of founder time inside it. Under 10 days skips the probe or the rubric. Over 18 days drifts into a stalled engagement.
Can a founder run this sprint alone, without a partner?
Idea compression and eval-case authoring, yes. The capability probe and PRD synthesis require partner support — current model behavior, eval mechanics, and integration design are not non-engineer terrain. Solo founders tend to skip the probe, which is the worst thing to skip.
What if the capability probe comes back as a no?
That is the cheapest no a founder can buy. Ending at day 5 saves roughly 70% of the engagement spend and the build cost. Walk, re-scope to a narrower capability, or wait for the next model upgrade — all three are correct.
What does the day-7 go/no-go gate look like in a contract?
A clause that says: “At day 7, the founder may terminate with no further obligation, retaining all artifacts through Phase 3 and a pro-rata refund of unused fees.” Vendors who decline this clause are signaling they cannot deliver value before day 14.
How many eval cases should the founder personally write?
Most of them. The founder produces the cases; the partner shapes rubric and threshold. Working size: 30 to 50 cases per critical task, eight to twelve hours across two or three sessions.
What is the relationship between this sprint and the build that follows?
The sprint produces the PRD; the build ships against it. A 6 to 12 week MVP build sits on top of the seven artifacts. The PRD does not change during build except by named change request. If it does, the sprint failed.
Is this sprint priced as a fixed fee or T and M?
Fixed-fee is the right shape. A serious idea-to-PRD sprint runs roughly 30K — consistent with public pricing from operating AI studios. T-and-M scoping drifts because there is no incentive to land artifacts on a calendar.
How do you handle a sprint where the founder’s idea changes mid-sprint?
Pause and re-compress. If the one-sentence idea on day 2 does not match the one on day 9, neither the probe nor the rubric is valid. A two-day pause is correct; pretending the change is small is not.
Closing
A 14-day idea-to-PRD sprint is not a discovery workshop. It is a tightly scoped exercise that produces seven named artifacts on a calendar. A founder who can count those seven files on day 14 has bought scoping. A founder who cannot has bought consultancy theater.
The discipline behind the sprint lives in the idea validation playbook; the document the sprint produces is walked in the AI PRD explained; the Phase 2 feasibility test is walked in the AI feasibility check; the engagement-side mechanics live in the anatomy of an AI agency engagement.
Count the seven artifacts on day 14. The number, not the narrative, is the test.
Arthur Wandzel
