TL;DR. Most professionals now use AI at work, and almost none of them keep what works. Gallup’s May 2026 survey found 52 percent of U.S. employees use AI in their role at least a few times a year, 30 percent a few times a week or more and 15 percent daily, up from 40, 19 and 8 percent a year earlier. Only 25 percent say their organization has communicated a clear plan for integrating it. The Federal Reserve Bank of St. Louis survey series puts the time saved by workers who use generative AI at 5.4 percent of their hours, about 2.2 hours in a 40-hour week, and the share of all work hours saved rose from 1.6 percent to 2.2 percent between late 2024 and mid-2026. Three strands of research show why a personal prompt library is the cheapest way to make that time compound. Prompts are brittle: changing only separators, spacing or casing moved accuracy by up to 76 points in one benchmark study, and a single prompt template produces unreliable model rankings on many tasks. People design prompts badly by default: in a Berkeley study, non-experts iterated opportunistically and declared success after a single good output. And prompt quality matters: in a controlled student sample it explained 53 to 78 percent of the variance in output quality. A CEOtudent analysis joins the adoption surveys with eight field experiments into a Template Priority Index that tells you which task families deserve a template first, then gives the prompt card, folder structure and maintenance rules. The CEO move is to treat your prompts as versioned assets. The student move is to treat each one as a hypothesis with test cases.
This is the systems companion to our AI literacy stack hub. That piece argued that prompting is one layer among several. This one is about the layer most professionals skip: storing, testing and reusing the prompts that already work for them.
Why does a library matter more than a better prompt?
Four findings, each from a different kind of study, point to the same conclusion.
Prompts are brittle, so reuse the exact one that worked. In a 2024 ICLR paper, Sclar, Choi, Tsvetkov and Suhr changed nothing but formatting in few-shot prompts (separators, spacing, the casing of field names) and found “performance differences of up to 76 accuracy points” for LLaMA-2-13B, with about 10 points of variation on average across more than 50 tasks and several models. The sensitivity shrank but did not vanish for larger and instruction-tuned models; GPT-3.5 showed a median spread of 0.064 with a maximum of 0.562 across 53 tasks. Mizrahi and colleagues, evaluating 20 models on 39 tasks with more than 5,000 instruction paraphrases, found that a single template “leads to unreliable rankings for many of the tasks”, with most agreement scores below 0.85. The practical reading is not that you must engineer the perfect prompt. It is that once a format works for your task, retyping it from memory each time throws the result away.
People iterate badly by default. Zamfirescu-Pereira, Wong, Hartmann and Yang watched 10 non-experts design chatbot prompts for CHI 2023. Participants “explored prompt designs opportunistically, not systematically”, over-generalized “from single observations of success and failure”, and “most participants declare success after just a single instance, and move on”. Several insisted on starting every instruction with “please”. A library with test cases is a direct countermeasure: a prompt is not promoted until it has worked on more than one input.
Prompt quality predicts output quality. In a 2024 study of 45 university students by Knoth, Tolzin, Janson and Leimeister, the quality of the prompt (scored on six components) explained 53 percent of the variance in output quality on a travel-planning task and 78 percent on a project-planning task. The sample is small and the tasks are academic, but the direction is consistent with every field experiment below.
The gains are real but uneven, so you need a map of where they are. The field experiments summarized in the next section report large productivity effects on specific task families and a measurable penalty when AI is used outside its competence. A library is where you record both sides: what to use a prompt for, and what not to.
What have the field experiments actually measured?
| Study (primary source) | Setting and sample | Task family | Verified effect | Caveat recorded in the paper |
|---|---|---|---|---|
| Noy and Zhang, 2023, Science | 453 college-educated professionals, incentivized writing tasks, half given ChatGPT | Professional writing | Average time down 40 percent, quality up 18 percent; exposed workers 2 times as likely to use it at their job two weeks later | Short tasks graded by evaluators; one tool and one model generation |
| Dell’Acqua and colleagues, 2023, Harvard Business School working paper | 758 consultants at one firm, 18 realistic tasks inside the AI frontier, one task outside | Analysis, writing, strategy memos | Inside the frontier: 12.2 percent more tasks, 25.1 percent faster, more than 40 percent higher quality; below-average performers gained 43 percent, above-average 17 percent. The arm given a prompt-engineering overview raised quality scores by 25.1 percent against 17.9 percent for AI alone | Outside the frontier, AI users were 19 percentage points less likely to be correct (control about 84.5 percent, AI arms 60 and 70 percent) |
| Brynjolfsson, Li and Raymond, 2025, Quarterly Journal of Economics (arXiv version) | 5,172 customer support agents, staggered rollout of an AI assistant | Chat-based customer support | Issues resolved per hour up 15 percent on average; about 30 percent for less experienced agents, 36 percent in the lowest skill quintile; agents followed recommendations 35 percent of the time | Minimal gains and small quality declines for the most experienced agents |
| Peng, Kalliamvakou, Cihon and Demirer, 2023, arXiv | 95 professional programmers, one HTTP-server task | Coding | Treated group 55.8 percent faster (71.17 versus 160.89 minutes), confidence interval 21 to 89 percent | Single task and a small sample, so the confidence interval is wide |
| Cui, Demirer, Jaffe, Musolff, Peng and Salz, 2025 (Microsoft, Accenture and a Fortune 100 company) | 4,867 developers across three randomized rollouts | Coding | 26.08 percent more completed tasks (standard error 10.3); junior developers gained 21 to 40 percent, seniors 7 to 16 percent | Each experiment noisy on its own; adoption among treated developers ranged from 42.5 to 75.6 percent at different points, so effects are per user |
| Chatterji and colleagues, 2025, NBER (ChatGPT usage) | About 1.1 million sampled conversations, May 2024 to June 2025 | All work use | Writing is 40 percent of work messages; about two-thirds of writing messages ask the model to modify text the user supplied rather than draft from scratch; about 81 percent of work messages fall under two broad activities, handling information and making decisions or solving problems | Consumer plans only; classifier-based |
The pattern across the table is the foundation of the index below. Gains are largest where the task is frequent, well-specified and inside the model’s competence, and they are largest for people with less experience in the task. The one study that tested prompt guidance directly (the Dell’Acqua overview arm) found it improved quality further. The one study that tested a task outside the frontier found a penalty large enough to wipe out the benefit.
How widely is AI used, and how much time does it save?
Two independent survey programs track the same question with different definitions, so we keep them in separate rows and do not splice them.
| Source | Measure | Latest verified values | Trend |
|---|---|---|---|
| Gallup quarterly workforce panel (U.S. employees, 19,043 to 23,717 respondents per wave) | Use AI in role a few times a year or more / a few times a week or more / daily | 52 / 30 / 15 percent (May 2026) | 40 / 19 / 8 percent in May 2025; daily use 4 percent in May 2024 |
| Gallup, same panel | Organization has communicated a clear AI plan | 25 percent (May 2026) | 22 percent in May 2025; 47 percent say their organization has integrated AI tools |
| Gallup, same panel | Most common uses among AI users | Writing and editing 51 percent, search or research 49 percent, general assistance or problem-solving 39 percent | May 2026 wave |
| Bick, Blandin and Deming, Real-Time Population Survey, NBER working paper 32966 | Share of U.S. workers using generative AI at work; time saved | 26.5 percent any use, 9.0 percent every workday, 22.9 percent weekly (August and November 2024); users save 5.4 percent of work hours, 1.4 percent across all workers | Top-two ranked tasks: writing 39.5 percent, administrative 25.6, interpreting or summarizing 22.7, fact search 18.0, idea generation 13.2, coding 12.9, documentation 12.7 |
| St. Louis Fed Generative AI Adoption Tracker (same survey, revised definitions) | Weekly work use; hours assisted; hours saved | 39.2 percent weekly use (Q2 2026); 6.3 percent of hours assisted; 2.2 percent of hours saved | From 28.2, 4.1 and 1.6 percent in Q3 2024 |
Two numbers from these tables set the stakes. A worker who uses generative AI is saving about 2.2 hours a week by their own estimate, and three in four workers have no organizational plan to help them. Whatever system exists is going to be personal.
The Template Priority Index
The table below is a CEOtudent editorial framework. It joins three verified sources that have not been combined before: Gallup’s use-case shares among AI users, the Fed survey’s top-ranked work tasks, and the field-experiment evidence per task family. Priority is assigned by a stated rule, not by judgment alone: Priority 1 means the task family is both frequent (top three in either survey) and backed by at least one randomized or quasi-randomized field study; Priority 2 means one of the two; Priority 3 means neither, or the evidence carries a verification warning.
| Task family | Share of AI users (Gallup, May 2026) | Share ranking it top-two (Fed survey, 2024) | Strongest field evidence | Repeatability (editorial) | Priority |
|---|---|---|---|---|---|
| Editing and rewriting text you supply | 51 percent (writing and editing) | 39.5 percent (writing) | Noy and Zhang: 40 percent faster, 18 percent higher quality; two-thirds of ChatGPT writing messages are edits | High: same inputs, same audience, weekly | 1 |
| Drafting routine documents and administrative text | Within the 51 percent | 25.6 percent (administrative), 12.7 percent (documentation) | Brynjolfsson, Li and Raymond: 15 percent more issues resolved in scripted support work | High | 1 |
| Summarizing, interpreting and translating | Not reported separately | 22.7 percent | Dell’Acqua: inside-frontier tasks included analysis and synthesis, 12.2 percent more tasks, more than 40 percent quality | Medium: inputs vary | 2 |
| Structured analysis and problem-solving | 39 percent (general assistance or problem-solving) | Not reported separately | Dell’Acqua: strong gains inside the frontier, 19 percentage points worse outside | Medium, with a frontier flag on every card | 2 |
| Research and fact search | 49 percent | 18.0 percent | No randomized field study measured accuracy of AI search at work | Low: every output needs verification | 3 |
| Coding | Not reported separately | 12.9 percent | Peng: 55.8 percent faster; Cui: 26.08 percent more tasks | High for developers | 1 for developers, otherwise not applicable |
| Idea generation | Not reported separately | 13.2 percent | No field study isolated it; Noy and Zhang describe tasks shifting toward idea generation and editing | Low: outputs are not comparable | 3 |
Read the index as a build order. Your first three cards should cover editing text you wrote, the routine documents you produce every week, and (if you code) your most common coding task. Research prompts come last and carry a verification step by design.
The prompt card: one template, nine fields
A library is only as good as its unit of storage. We recommend a card with nine fields, kept in plain text so it survives any tool change. The fields map onto the research: the fixed format answers Sclar and Mizrahi, the test cases answer the Berkeley study, the frontier line answers Dell’Acqua.
| Field | What goes in it | Why it exists |
|---|---|---|
| Name and version | Family-verb-object plus a version number, for example “edit-exec-summary-v3” | You will have five variants within a month; names stop the drift |
| Trigger | The concrete situation that calls for this card, for example “any draft over 300 words going to a decision-maker” | Cards are retrieved by situation, not by cleverness |
| Inputs required | The exact things you must paste or attach: draft, audience, decision wanted, length cap | Missing inputs are the most common cause of weak output |
| Prompt body (frozen) | The full text, including separators and casing, never retyped from memory | Formatting changes alone shift results |
| Model and settings | Which model and mode you tested it on, with a pointer to your routing rule | Results do not transfer across models unchanged |
| Test cases | Three inputs with the output you accepted, saved alongside | A card is promoted only after three passes, not one |
| Known failure modes | What this prompt does wrong when it fails, for example “adds facts the draft did not contain” | Turns one bad surprise into a permanent guardrail |
| Frontier line | The explicit “do not use for” sentence, for example “not for legal or financial wording” | The outside-frontier penalty is the single largest documented risk |
| Last verified and owner | Date you last ran the test cases; who maintains the card in a team | Models change; a card without a date is a rumor |
A filled-in card
Name and version: edit-exec-summary-v3.
Trigger: Any draft longer than 300 words that will be read by someone who decides something.
Inputs required: The draft; the reader’s role; the single decision the reader must make; the length cap in words.
Prompt body (frozen): “You are editing, not writing. Keep every factual claim exactly as given and do not add any fact, number or example that is not in the draft. Reader: [role]. Decision the reader must make after reading: [decision]. Rewrite the draft so the first paragraph states the recommendation and the one number that supports it, the second paragraph gives the two strongest reasons, and the final paragraph names the risk and what would change the recommendation. Maximum [N] words. Where the draft lacks a fact the reader will need, insert the marker [GAP: what is missing] instead of filling it in. Return only the rewritten text.”
Model and settings: The general-purpose model from our routing rule for writing; no browsing.
Test cases: A budget memo, a project status email, a set of slide notes, each with the accepted output saved.
Known failure modes: Drops qualifiers such as “preliminary” when compressing; occasionally merges two reasons into one.
Frontier line: Not for contract language, regulatory filings or anything with legal effect; not for drafts that contain no numbers (use the drafting card instead).
Last verified and owner: 2026-10, you.
Notice that the card encodes the finding from the ChatGPT usage data: two-thirds of professional writing requests are edits of text the user already has. The most valuable card in most libraries is an editing card, not a drafting card.
How do you organize the library?
Folder by task family, not by tool. Tools change; the task families in the index above have been stable across every survey since 2024. Use the seven families as top-level folders and let coding sit empty if you do not code.
Three tiers inside each family. Core cards are the ones you run weekly; they have three test cases and a date. Situational cards run monthly or less; they have one test case. Experimental cards are anything you tried once; they live in the folder but carry no version number until promoted. The Berkeley finding is the reason for the tier system: the default human behaviour is to promote after one success.
One index file. A single page listing every Core card with its trigger. Retrieval is by trigger (“draft going to a decision-maker”), which is how you will actually remember it under time pressure.
Store inputs separately. Your standing context (role, audience, house style, constraints) belongs in a context document, not inside every card. Our pieces on the context file and on configuring custom instructions, projects and memory cover that layer; the cards then stay short and portable. Which model a card is tested on follows your routing rule.
Six operating rules from the evidence
- Freeze the format. Copy the prompt body verbatim. Do not “tidy” separators, capitalization or spacing between runs. Sclar and colleagues note that for people building systems on top of a model, choosing one format that works and keeping it is a valid method.
- Three cases before promotion. A prompt moves from Experimental to Core only after three different inputs produced accepted output. Mizrahi’s multi-prompt evaluation is the research version of the same rule.
- Write the frontier line first. Before you save a card, write the sentence that says what it must not be used for. Dell’Acqua’s consultants lost 19 percentage points of accuracy on the one task the tool could not do, and they did not notice.
- Record failures as fields, not memories. Each time a card fails, the failure goes into the known-failure field the same day. The Berkeley participants over-generalized from single failures as readily as from single successes; a written record prevents both.
- Date and retest quarterly. Re-run the three test cases on the current model each quarter. If the output changes materially, bump the version.
- Share the Core tier. Gallup finds three in four employees have no organizational AI plan and, in Microsoft’s 2024 survey of 31,000 knowledge workers, 78 percent of users were bringing their own AI tools to work. A team-shared Core folder is the smallest unit of an AI plan that actually exists. It also closes the experience gap the field experiments keep finding: novices gained 30 to 43 percent in the support and consulting studies, and a shared card is how a novice borrows an expert’s prompt.
The CEO view and the student view
Run the library like an asset register. Each card has an owner, a version, a last-verified date and a retirement rule (no use in six months means archive). Measure it once a quarter with the Fed survey’s own yardstick: how many hours did Core cards save you this week, by your honest estimate? If the answer is below the 2.2 hours per week that the average user reports, the library is not covering your frequent tasks and the index above tells you which family is missing.
Learn from it like a lab notebook. Every card is a hypothesis about what works; the test cases are the experiment; the known-failure field is the result you did not want. The research gives the student a reassuring fact: the people who gained most from AI assistance in every field experiment were the less experienced ones, and the mechanism was access to codified good practice. A library is codified good practice you wrote for yourself. It pairs naturally with the question quality framework: the card holds the format, the framework checks whether the question inside it is worth asking.
What this evidence cannot tell you
- No randomized study has tested “keeping a prompt library” as an intervention. The case here is assembled from studies of prompt brittleness, novice behaviour, prompt quality and task-level productivity. Treat it as a well-supported inference, not a measured effect.
- The field experiments used 2023 to 2025 models on specific tasks in specific firms. Effect sizes will differ for your tasks and for current models; that is exactly why the card carries a date and test cases.
- The brittleness results were measured mostly on open-weight models in few-shot settings. Newer models may be less sensitive, but the Sclar paper found sensitivity remained after scaling and instruction tuning, and the cost of freezing a format is close to zero.
- Survey figures are self-reported and the two survey programs define use differently. We never combine a Gallup number with a Fed number in one sentence.
- The Knoth study has 45 students and academic tasks; the Berkeley study has 10 participants. Both are used here for mechanism, not for magnitude.
Frequently asked questions
Is a prompt library the same as a prompt-engineering guide?
No. A guide tells you principles; a library holds the specific prompts that have passed your test cases on your tasks. The Dell’Acqua experiment suggests a short prompt-engineering overview improves results, and the library is where that overview turns into reusable practice.
How many cards do I need?
Fewer than you think. Start with three Core cards from the Priority 1 families. Most professionals’ weekly AI use is concentrated in editing, routine documents and search, which is why the index puts them first.
Should I store the cards inside the AI tool’s own feature for saved prompts?
Store the master copy in plain text you control and mirror it into tool features as a convenience. Tool features change, and the card’s test cases and failure fields usually have no home there.
What about prompts for AI agents rather than chat?
The same card works, with the frontier line doing more of the work. Our pieces on delegating to an AI agent and on context engineering cover the extra fields an agent brief needs.
How do I know a card is still good after a model update?
Re-run the three test cases. If two of three outputs change in ways you would not have accepted, the card goes back to Experimental until it passes again.
Sources
- Gallup, “AI Use at Work Has Nearly Doubled in Two Years” (June 2025), “Organizational AI Adoption Jumps Six Points” (July 2026) and the Gallup Indicator on Artificial Intelligence (data as of May 2026).
- Bick, Blandin and Deming, “The Rapid Adoption of Generative AI”, National Bureau of Economic Research Working Paper 32966 (revised February 2025); Federal Reserve Bank of St. Louis, “The Impact of Generative AI on Work Productivity” (February 2025) and FRED Blog, “Does generative AI save time at work?” (August 2026).
- Chatterji, Cunningham, Deming, Hitzig, Ong, Shan and Wadman, “How People Use ChatGPT”, National Bureau of Economic Research Working Paper 34255, September 2025.
- Sclar, Choi, Tsvetkov and Suhr, “Quantifying Language Models’ Sensitivity to Spurious Features in Prompt Design”, International Conference on Learning Representations 2024.
- Mizrahi, Kaplan, Malkin, Dror, Shahaf and Stanovsky, “State of What Art? A Call for Multi-Prompt LLM Evaluation”, Transactions of the Association for Computational Linguistics, volume 12, 2024.
- Zamfirescu-Pereira, Wong, Hartmann and Yang, “Why Johnny Can’t Prompt: How Non-AI Experts Try (and Fail) to Design LLM Prompts”, Proceedings of CHI 2023.
- Knoth, Tolzin, Janson and Leimeister, “AI literacy and its implications for prompt engineering strategies”, Computers and Education: Artificial Intelligence, volume 6, 2024.
- Noy and Zhang, “Experimental evidence on the productivity effects of generative artificial intelligence”, Science, 381(6654), 2023, pages 187 to 192.
- Dell’Acqua, McFowland, Mollick, Lifshitz-Assaf, Kellogg, Rajendran, Krayer, Candelon and Lakhani, “Navigating the Jagged Technological Frontier”, Harvard Business School Working Paper 24-013, 2023.
- Brynjolfsson, Li and Raymond, “Generative AI at Work”, Quarterly Journal of Economics, 140(2), 2025 (figures from the November 2024 arXiv version).
- Peng, Kalliamvakou, Cihon and Demirer, “The Impact of AI on Developer Productivity: Evidence from GitHub Copilot”, arXiv 2302.06590, 2023.
- Cui, Demirer, Jaffe, Musolff, Peng and Salz, “The Effects of Generative AI on High-Skilled Work: Evidence from Three Field Experiments with Software Developers”, 2025 (author version, June 2025); Microsoft and LinkedIn, 2024 Work Trend Index Annual Report (for the bring-your-own-AI figure).
This content was compiled with the support of AI following in-depth research, then written and prepared for publication by the CEOtudent editorial team.
