Source reading · YC Requests for Startups
Ninety-six prompts, eight editions and several recurring questions. Follow how a thesis changes, then decide what evidence would make a specific problem worth pursuing.
This reading covers the eight labelled collections in the official YC archive, from Summer 2024 through Fall 2026. Each occurrence keeps its edition and original source link. An investor returning to a subject is a reason to investigate its changing premise, not evidence that customers need the proposed product.
Some prompts endure: business records are difficult to replace, people need useful explanations, engineering predictions must be checked and public administration needs accountable handoffs. Other requests depend on a particular technical, financial or policy moment. Keep those moments attached to their editions instead of treating every request as a current recommendation.
The archive moves from tools and assistance toward responsibility for completed work, and from individual agents toward shared context and coordination. That is an editorial reading of these publications. It does not mean each earlier thesis was resolved, that later models can do the work reliably, or that a company should serve every connected category.
Back-office tools, vertical agents, full-stack operators, agencies and service companies propose different places to sit in a value chain. The editorial question is who accepts the completed work and pays for rework. Selling an outcome can change liability, customer acquisition and oversight; it is not merely a different software price.
Shared app memory, internal builders, a company brain and multiplayer sessions ask for different kinds of context reuse. Test the boundary where a rule is approved, an instruction conflicts or access is revoked. These are connected design questions, not interchangeable products or repeated proof of demand.
Simulation, engineering tools, worker guidance and physical-world operating systems all meet real operating constraints. Start with a qualified site partner and a reversible comparison. A model benchmark does not establish field reliability, worker competence or a safe engineering decision.
Government paperwork, consulting and application-processing prompts leave procurement and responsibility with actual institutions. Separate a factual administrative aid from an eligibility, legal or clinical decision. A public benefit is an objective to investigate, not evidence of an accessible budget.
Find a recent instance of the workflow. Who wanted it done, who did it, who accepted the result and who paid? Ask what happened when a document was missing, a system was unavailable or the usual operator was away. Exceptions can reveal a more useful starting problem than the polished version of the task.
Then compare one proposed change with the current method under the same conditions. Count human review, integration, retries and recovery. A shorter model response is not a faster accepted workflow, an appealing explanation is not a faithful one, and an interesting laboratory prediction is not a clinical result. A good stopping condition is part of the experiment: the simpler existing process may win.
Several requests explicitly concern US healthcare payments, government procurement, domestic manufacturing or American defense. An India lens starts by finding the actual institution, buyer and operating arrangement here. The original mechanism may not carry over. A professional qualification, public contract or financial-product permission needs its own local review; the source’s optimism does not provide it.
More portable workflow ideas still need local tests. Which language does a worker use? Who owns a record across organizations? Can the site supply the required data, connectivity and service support? Can the intended buyer fund an outcome? These questions can change the product before its interface or model is chosen. They do not establish an Indian cost, labour or data advantage in advance.
Dates here identify an edition, not an exact publication day or a new verification date. The current official archive still exposes each labelled edition. It is not an immutable historical snapshot, so this review does not claim that every body is unchanged since first posting. Embedded examples and subtopics remain inside their original request; no new occurrence is manufactured from them.
13 source occurrences
This edition mixes new coordination layers with consumer, institutional and physical-world ambitions. Ask who can approve the initial workflow before treating the larger platform or scale claim as a starting market.
Open the original Fall 2026 collection
16 source occurrences
Service delivery, company knowledge and physical production share an operational question: who accepts the result and handles exceptions? Keep semiconductor, hardware, agriculture and space requirements specific to their sites and counterparties.
Open the original Summer 2026 collection
10 source occurrences
Several prompts change how a professional workflow is organized. The first useful comparison can be a cited brief, an accepted administrative result or a reproducible engineering task. Financial and legal premises remain tied to their original context.
6 source occurrences
Enterprise replacement, agent fleets and workforce training require different adoption evidence. A shared enthusiasm for AI does not make their users, budgets or acceptance tests interchangeable.
14 source occurrences
Personal assistance, teaching, voice interfaces and research invite different measures of usefulness. For each, ask whether the person can complete the task, learn without assistance or reproduce the finding, rather than counting fluent output alone.
Open the original Summer 2025 collection
12 source occurrences
Infrastructure spans permissioned context, agent tools, compute and operating workflows. Compare a proposed layer with the simpler existing method and count integration, review and failure-recovery work.
Open the original Spring 2025 collection
9 source occurrences
Public administration and engineering sit alongside broad finance, industrial and employment ambitions. The latter remain source-only when a bounded product, local operating context or qualified evidence is not supplied.
16 source occurrences
The earliest captured edition contains enduring questions about records, tooling, explanation and physical prediction. Its examples and forecasts remain historical source context rather than current evidence about model capability or clinical benefit.
Open the original Summer 2024 collection
Seventy requests have analysis to help frame a first investigation. Twenty-six requests stay source-only: some are broad research or founder-profile invitations; others depend on clinical, legal, military, financial or project premises this review has not independently established. A source-only decision is an evidence boundary, not a judgment that the subject is unimportant.
This reading links the original requests and underlying Atlas entries. It reproduces no source bodies, images, charts, commercial figures or performance benchmarks, and adds no personal founder preferences or first-person commentary. The notes help design an investigation; they do not validate demand, medical benefit, safe operation or permission to enter a regulated activity.
Bharat in Beta