What is a Shared Prompt Review?

A Shared Prompt Review is a short team conversation about how a piece of work was made with AI, held on the work itself and not on a policy. Four things go on the table: the prompt as it was actually typed, the raw output, what a person kept or cut and why, and where the output might still be wrong. Rahim Hirji proposed it on 11 January 2026. The problem it addresses has been measured. The review itself has not, and this page keeps those apart.

The four things that go on the table#

The structure is deliberately small. The prompt, in the form it was used, with no polishing and no retrospective rewriting. The raw output, as the system returned it, before editing. The human intervention: what was kept, what was cut, what was changed, and the reasoning behind each. The critique: where the output might be wrong, narrow or misleading, and one alternative direction the team decided not to take.

The first element carries most of the weight. A prompt rewritten before the meeting is a description of what somebody meant to ask, and the assumptions that went unexamined were in the original. The fourth element is the one most often dropped, and the only part that surfaces a path not taken.

Review normally reaches the artefact and stops#

Most teams are good at reviewing what was produced and have no habit at all of reviewing how it was produced. That gap was survivable when the production was visible in the room. It stops being survivable when a polished, confident draft appears from a process nobody saw, because the discussion then moves to taste, agreement comes quickly, and no learning follows.

The estate's own oversight material says the same thing from the other end. Overseeing work you did not do is the hardest task in the job and the one the automation removes the practice for, which is Bainbridge's 1983 result and the foundation under the invisible work of oversight. A review of the process is one way of keeping some of that practice in the room.

A field experiment shows specialists converging#

The sharpest evidence for the flattening the review is meant to catch comes from Dell'Acqua and colleagues, a pre-registered field experiment with 776 professionals at Procter and Gamble working on real product innovation problems, randomised on AI access and on whether people worked alone or in pairs. Individuals with AI matched the performance of two-person teams without it.

The result that matters here is the second one. Without AI, research and development professionals proposed more technical solutions and commercial professionals proposed more commercially oriented ones. With AI, both groups produced balanced solutions whatever their background. The difference a specialist brings to a room was erased inside a single session. The authors are careful that balanced is not the same as better, and nothing in the study establishes that it is. Graded entry.

Two smaller results point the same way. Doshi and Hauser found AI-assisted stories rated more creative and markedly more similar to one another, with the largest individual gains going to the weakest writers. Graded entry. And Hohenstein and colleagues found that using algorithmic suggestions moved a person's own unassisted sentences, while merely having suggestions available moved nothing (p=0.1801), which puts the loss at the moment somebody adopts another voice's phrasing. Graded entry. A review that puts the prompt on the table is aimed at that moment.

One line from the original essay this page leaves out#

The essay opens with an unattributed appeal to research, claiming that generative AI boosts creativity for only a minority of employees. No study is named, and this estate does not carry a figure or a direction on a claim with nothing behind it. The nearest measured result runs the other way on the individual question: Doshi and Hauser found creativity ratings rose for most assisted writers and rose furthest for the least creative. The collective loss there is in diversity, not in whether individuals improve.

Removing that sentence costs the argument nothing. The case for a Shared Prompt Review does not rest on how many people get better at prompting. It rests on the process being invisible to everyone except the person who ran it.

Nobody has tested whether the review changes anything#

No trial has been run on this practice, and none of the studies above measures a remedy. They measure convergence, invisibility and the difficulty of oversight. Whether a weekly conversation about prompts improves a team's decisions, or merely adds a meeting, is unmeasured, and a reader should treat the recommendation accordingly.

Two specific risks sit inside it. A review of prompts can become a performance, with people bringing the prompt they wish they had written, which is the failure the first element exists to prevent and cannot by itself guarantee. And a standing review of how colleagues use AI can be read as surveillance in a team where disclosure norms are unsettled, which is the setting the practice is proposed for. The Hohenstein suspicion result belongs here too: people suspected of AI use were rated less cooperative and less affiliative at p<0.0001, while suspicion tracked actual use at a correlation of only 0.22. A forum where use is stated openly removes the guessing, and nothing tests whether it removes the penalty.

Running one without it becoming a ritual#

Take one real piece of work that mattered, not a demonstration. Ask for the prompt before the output, so the conversation starts on the framing. Require the reasoning for each cut, because that is where the judgement is. Close on the critique, including the direction not taken, so at least one alternative is on the record. Keep it to half an hour, and stop running it when the norms it exists to build have arrived.

Those five moves follow from the argument above and from the practice as published. None of them has been trialled, and the estate marks that on the face of it rather than in a footnote.

Key sources

• Hirji, R. (2026). The Shared Prompt Review. Box of Amazing, 11 January 2026. The dated first publication of the practice and its four elements.
• Dell'Acqua, F., Ayoubi, C., Lifshitz, H. et al. (2025). The Cybernetic Teammate. NBER Working Paper 33641; Organization Science, June 2026. Graded entry.
• Doshi, A. R. and Hauser, O. P. (2024). Generative AI enhances individual creativity but reduces the collective diversity of novel content. Science Advances, 10(28). Graded entry.
• Hohenstein, J., Kizilcec, R. F., DiFranzo, D. et al. (2023). Artificial intelligence in communication impacts language and social relationships. Scientific Reports, 13, 5487. Graded entry.
• Bainbridge, L. (1983). Ironies of Automation. Automatica, 19(6), 775-779. Graded entry.

Related SuperSkills research#

What happens to a group that adopts these tools without changing how it works is at what AI does to a team and does AI make everyone think alike. The oversight argument underneath it is the invisible work of oversight and meaningful human oversight. On the individual version of the same discipline, the source rule and the human signal.

About this research#

Written by Rahim Hirji, author of SuperSkills (Kogan Page, 2026), keynote speaker on AI and human capability, and founder of The SuperSkills Intelligence Company. The Shared Prompt Review is his, first published in Box of Amazing on 11 January 2026, which is the dated publication the estate requires before crediting a practice to him. One unsourced claim from that essay is named above and left out. The findings are attributed to the researchers who produced them and kept separate from the interpretation.

How this research works  ·  Reviewed quarterly  ·  Found an error? Tell me and it is corrected on the page.