A small-group prompt writing activity: three or four people at one screen, a five-part skeleton, a group read of the output and a 30-minute run sheet.
Quick answer
Have people write prompts in groups of three or four instead of alone. Each group shares one screen, fills a five-part prompt skeleton aloud, reads the output together, revises once and brings one result back. In a 30-minute version, drafting gets 8 minutes, reading the output 5 and the share-out 5. No study cited here compares group-written with solo-written prompts; the case for it is shared understanding of the task and a scheduled check of the output.
Key takeaways
- No study cited in this article compares group-written with solo-written prompts, so the case for the activity is shared understanding of the task and learning, with no claim of better prompts.
- The skeleton and the compared revision answer two observed habits of people new to prompting: exploring prompt designs opportunistically and declaring success after a single instance.
- The Corporate Learning Community's Promptathon guideline reports that group sizes of 3 to 5 people have proven successful, and SAP's promptathons put 3-4 participants in each breakout room.
- In a 2023 working paper on an experiment with 758 consultants, those using AI were 19 percentage points less likely to produce correct solutions on a task chosen to lie outside what the AI did well.
- The method library lists Think-Pair-Share for 4 to 100 people and 10 to 20 minutes, and 1-2-4-All from 4 people and from 12 minutes.
Have people write prompts in small groups instead of alone. Each group of three or four shares one screen, fills a prompt skeleton aloud, reads the output together, revises once and brings one result back to the room. In the 30-minute version below, drafting gets 8 minutes, reading the output 5 and the share-out 5.
None of the sources here compares group-written with solo-written prompts, so nothing on this page claims that groups write better ones. The evidence is narrower: people new to prompting explore opportunistically, pairs prompting together discussed their strategy, and confidence in the tool goes with less critical thinking.
The activity fits inside a longer session, for example an AI adoption workshop or an AI brainstorming workshop.
1. Use it when the prompt matters or skills are mixed
Group prompt writing takes more people-time than typing alone. Keep it for prompts whose output the room will go on to use. This article also suggests it for rooms where AI skill is uneven, since both the Corporate Learning Community's (CLC) Promptathon guideline and SAP's promptathons mix skills within groups on purpose.
The nearest measured comparison is an analogy from software development. A 2009 meta-analysis of pair versus solo programming experiments by Hannay and colleagues found a small positive effect of pairing on quality, higher quality when task complexity was high, and a medium negative effect on effort. Its abstract adds that between-study variance is significant and that there are signs of publication bias (same abstract). None of the sources here runs that comparison for prompts.
Evidence on AI itself cuts the other way. In the Cybernetic Teammate, a 2025 working paper on a pre-registered field experiment with 776 professionals at Procter & Gamble, individuals with AI matched the performance of two-person teams without AI. Do not sell the activity as more output per head.
A CHI 2024 study by Han and colleagues observed pairs of students with design or performance backgrounds creating stage designs with generative AI: the tools helped them build consensus and allowed discussion of prompting strategies. The abstract gives no number of pairs. Shared understanding of the task is the case to make.
With hours available and prompt practice as the whole purpose, run the event-scale relative. The method library lists the Promptathon for 20 to 200 people and 3 to 8 hours, after the CLC guideline, which defines it as a peer learning format.
2. Form groups of three or four around one screen
Three or four people per group is where the published formats converge. The CLC guideline reports that "Group sizes of 3 to 5 people have proven to be successful" (guideline). SAP's format puts 3-4 participants in each breakout room (tcworld). An article on AI-enhanced workshops by NN/g suggests "a group of 2–4 people" who collaborate on a prompt and work through its outputs together. All three are practitioner experience.
One screen per group keeps the drafting spoken. The set-up is borrowed from pair programming, which Williams and Kessler described in 2000 as two programmers working "side-by-side at one computer": one drives the keyboard while the other continuously reviews the work. None of the sources here tests that split for prompt writing. For a group of three, this article suggests:
- Typist: writes what the group agrees on and nothing else.
- Reader: reads the prompt aloud before it is sent, and later the output.
- Challenger: stays off the keyboard and asks before sending, "Could I follow this if I knew nothing about our project?"
The challenger's question adapts a test from Anthropic's prompting best practices: "Show your prompt to a colleague with minimal context on the task and ask them to follow it."
Rotate the typist between first draft and revision. Williams and Kessler advise taking turns at the keyboard even when one partner is much more experienced, so the observer does not drift out of the loop. The failure to watch for is the fluent AI user typing while the others look on. In a pair, the typist also reads aloud; in a four, the fourth person keeps time.
3. Hand out a prompt skeleton with blanks
A skeleton turns drafting into a few explicit decisions. In a CHI 2023 study, Zamfirescu-Pereira and colleagues observed 10 participants without substantial prompt design experience improving a chatbot: they explored prompt designs opportunistically, carried over expectations from instructing humans, and all 10 preferred direct instruction to giving examples. Hence the example slot. Small qualitative sample, solo work, a GPT-3-era model.
Dang and colleagues call prompt writing for most end-users largely a trial and error process, in a 2022 CHI workshop paper whose abstract reports no participant data. Short structured practice may help: 27 undergraduates showed improved prompt engineering skills after one 100-minute workshop in a preprint by Woo and colleagues, whose abstract names no control group.
The five parts follow vendor guidance as it stood in November 2026: Microsoft's Copilot guidance names four parts of a prompt (goal, context, source, expectations) and Google's Workspace guidance four main areas (Persona, Task, Context, Format).
- Role: the purpose, communication style and high-level goals of the assistant, in the words of OpenAI's guide.
- Context: Anthropic advises giving the motivation behind an instruction, which can help the model deliver more targeted responses.
- Task: Microsoft calls it the goal and says "all that's required is a clear goal" (Microsoft Support).
- Format and limits: Anthropic advises being specific about the output format and constraints.
- Example: Anthropic says a few well-crafted examples improve accuracy and consistency; OpenAI asks for possible inputs with the desired output.
A group short on time should settle the task line first. Print the skeleton or paste it into each group's chat window:
Role: You are helping [who] with [purpose]. Write in [tone].
Context: The team is [who they are and what they are working on]. This matters because [why].
Task: [The one thing wanted, stated as precisely as the group can].
Format and limits: Give [how many items, in what form], each [how long]. Leave out [what is not wanted].
Example: A good answer looks like this: [one example the group wrote itself].
Say the session is about the first week for new hires. A group might fill the task line with "List ten ways to make a new colleague's first week less confusing" and argue about whether "new colleague" means a graduate or an experienced hire. The example is invented. That argument is the activity working.
4. Schedule the reading of the output
Give the reading its own slot. In a CHI 2025 survey of 319 knowledge workers who shared 936 first-hand examples of using GenAI in work tasks, Lee and colleagues found higher confidence in GenAI associated with less critical thinking, and higher self-confidence with more. These are self-reports and an association.
The Jagged Technological Frontier working paper (Dell'Acqua and colleagues, 2023) reports a pre-registered experiment with 758 consultants at Boston Consulting Group. On a task chosen to lie outside what the AI did well, consultants using AI were 19 percentage points less likely to produce correct solutions than those without (same paper). On 18 tasks inside that frontier, the paper measured gains for those using AI (abstract). This article's reading: a group rarely knows which kind of task it has, so treat every output as unchecked.
Over-reliance is older than chatbots. A systematic review of automation bias with a focus on healthcare included 74 papers. In the studies it pooled, wrong advice from a clinical decision support system raised the risk of an incorrect decision by 26%, and its mitigators include training and emphasizing user accountability (same review). The review does not list group review among its mitigators. Reading together is this article's way of giving the check an owner, with no proven effect. A suggested routine:
- Hands come off the keyboard.
- The reader reads the output aloud from the first line to the last.
- Each person marks one usable thing and one that is wrong, vague or impossible to check in the room.
- One named person, the challenger, compares the output with the skeleton: format, limits, context.
The limit is knowledge. A fact nobody at the table can verify goes back to the room marked as unchecked.
5. Revise once, then bring one result back to the room
One revision gives the group two outputs to compare. In the Zamfirescu-Pereira study, most participants declared success after a single instance and moved on, and they over-generalized from single data points in both directions. Ask each group to change one part of the skeleton, the one their marks point to, and run the prompt again with a new typist. Change three parts and nobody can say which mattered. One round is this article's choice for a 30-minute slot; Google's guidance recommends follow-up prompts in an iterative process of review and refinement.
Then each group brings back one result of its own choosing. The CLC guideline has each group present, followed by a discussion and feedback round. SAP's format is stricter: two to three teams demonstrate their work at the end, if they wish (tcworld). The method library's tip for 1-2-4-All is to limit the number of ideas shared in the final step in large groups. If the results are ideas that need ranking, continue with scoring AI-generated ideas against evaluation criteria.
A 30-minute run sheet that adds up
Picture a hypothetical workshop of 15 people in five groups of three, each with one laptop and a printed skeleton. Group and timings are invented, a starting point that nobody has tested as a protocol.
| Minutes | Step | What happens | Length |
|---|---|---|---|
| 0-2 | Frame | Task, roles and skeleton explained | 2 min |
| 2-4 | Think alone | Silent notes on what the prompt must contain | 2 min |
| 4-12 | Draft aloud | Groups fill the skeleton, the typist writes | 8 min |
| 12-14 | Send | Reader reads the prompt aloud, challenger asks | 2 min |
| 14-19 | Read the output | Read aloud end to end, mark usable and wrong | 5 min |
| 19-23 | Revise once | New typist, one part changed, compare | 4 min |
| 23-28 | Share | Five groups, one result and 1 minute each | 5 min |
| 28-30 | Close | The room names what changed between drafts | 2 min |
| Total | 30 min |
The arithmetic: 2 + 2 + 8 + 2 + 5 + 4 + 5 + 2 = 30 minutes. The share step is 5 groups × 1 minute = 5 minutes.
Two steps are borrowed from methods in the library, and the run sheet runs neither in full.
- Silent start: from Think-Pair-Share, listed for 4 to 100 people and 10 to 20 minutes with a think phase of 2 to 3 minutes; the library's tip is that this phase is the one most often skipped.
- Share step: from 1-2-4-All, listed from 4 people and from 12 minutes: 1 minute alone, 2 in pairs, 4 in fours and 5 with everyone (1 + 2 + 4 + 5 = 12).
Thinking alone before prompting as a group has a published precedent: in a 70-minute Brainwriting session with 16 college students in 5 project teams of 3-4, described by Shaer and colleagues, students wrote ideas individually before each group was required to use an LLM for more.
Conclusion
None of the sources here shows whether groups write better prompts than individuals. Run the activity for the conversation it forces and the check it schedules.
For your next session, print the skeleton, assign the three roles and block 30 minutes with the run sheet above.
References
- Prompting research: Zamfirescu-Pereira et al., CHI '23; Dang et al., CHI 2022 workshop paper; Woo et al., preprint; Han et al., CHI '24; Shaer et al., arXiv preprint; Dell'Acqua et al., The Cybernetic Teammate, 2025
- Over-reliance: Lee et al., CHI '25; Dell'Acqua et al., Jagged Technological Frontier working paper, 2023; Goddard et al., automation bias review, 2012
- Pair programming, used as an analogy: Williams and Kessler, 2000; Hannay et al., 2009, abstract
- Published formats: CLC Promptathon guideline v1.1, 2024; tcworld on SAP's promptathons, 2025; NN/g, Facilitating AI-Enhanced Workshops, 2025
- Vendor guidance, fetched November 2026: Anthropic; OpenAI; Microsoft; Google Workspace