Founders ask Collective 54 this 7 times in our records, 3 of them in 2026. It is usually asked as a technique question, and the useful answer is an ownership question.
Most guidance on prompting treats it as a writing skill, a matter of phrasing and structure. That framing is not wrong, but it explains why so many firms have a folder of clever prompts and no change in output.
In a boutique, a prompt is the point where the judgment of an expert is transferred into something that can run without them. The founder who has seen a hundred client situations knows what a good discovery summary looks like, which risks matter in a proposal, what a weak analysis is missing. A prompt is that knowledge written down in a form the machine can act on, and its quality is almost entirely a function of how much of that knowledge made it in.
Which is why the published position is blunt: the advantage in this era is not the model, which everyone has, but what the founder feeds it. Every firm has the same screen. Your proprietary knowledge is the ammunition.
The most common mistake is the generic instruction. Write me a cold email to chief information officers. Summarize this call. Draft a proposal for a marketing project. Each one asks the model to work blind, with none of the context that would make the output belong to your firm rather than to any firm, and the result is competent, fast and interchangeable.
The second mistake is expecting the model to compensate for something the firm has not decided. A prompt cannot supply positioning that has not been settled, a point of view that has not been formed or a definition of good work that lives only in the head of one partner. It can only make the absence visible faster.
The third is delegation. Founders who treat AI as a black box and hand the configuration to junior staff get mediocre results, because the person writing the prompt is the person whose judgment gets encoded, and junior staff do not yet hold the judgment. The founder or the senior practitioner who owns the work has to be the author, at least of the version that everything else is built from.
The fourth is perfectionism. Firms try to build the complete prompt before using it, when the working method is to iterate quickly against real output and let the system learn.
Collective 54 does not publish a prompt format, and the elements below are inferences from the published material about what the model needs rather than a prescribed syntax.
Context from the firm. The client transcripts, the won and lost deal notes, the frameworks and methodologies, the pricing rationale, the interview scripts, the definitions of the ideal client, the examples of work the firm was proud of. This is what the published material calls a contextual prompt file, and attaching it is the single largest difference between a generic result and one that reads as if your firm produced it.
A defined output. What done looks like, in terms specific enough to check: the sections, the length, the audience, the standard. Delivery professionals already know this move, because acceptance criteria and a definition of done are what keep perfectionism from eating the budget, and the same discipline keeps a prompt from producing something plausible and useless.
A place for judgment. The eighty percent the machine does, which is the drafting, the synthesis, the checking and the variants, and the twenty percent a named person does, which is deciding what is true, what is relevant and what goes to the client. A prompt that implies the output ships unread has been written for a different kind of firm.
A checkpoint. Where the output is reviewed and by whom, before it becomes a proposal, a deliverable or a message to a prospect. Accountability does not transfer to the model. It stays with the person whose name is on the work.
A prompt that works today decays, because the work changes, the model changes and the firm learns things it did not know when the prompt was written. The published material treats prompt libraries as one of the assets that compound over time for early adopters, and compounding requires upkeep.
Give every prompt that matters an owner, the person accountable for the work it produces. Keep the library in one place rather than in the chat histories of individuals, because a prompt that lives in one browser leaves with that person. Version it, and record what changed and why, so that a prompt which used to work and no longer does can be diagnosed rather than rewritten from nothing.
Test against outcomes, not against how the output reads. The question is whether the proposal drafted this way won more often, whether the outreach message produced conversations with people who fit, whether the discovery summary caught the risks that later mattered. Record which prompts preceded work you won and which preceded rework, because that pattern is specific to your firm and is the part of the library no competitor can copy.
Retire prompts. A library that only grows becomes a library nobody trusts. When a service changes or a workflow is redesigned, the prompts that served the old version should be archived, not left to be found by the next person.
And feed the library from the work. The delivery professional who finishes an engagement is the richest source of what should become standard, and turning project artifacts into reusable assets is exactly the kind of capture that used to be skipped because people were drowning. Now it is cheap, and a firm that does it is building the operating knowledge that a buyer will one day pay for.
The person whose judgment the prompt encodes. For prompts that carry the point of view of the firm, the ideal client, the reasons it wins, that is the founder, and this is work that cannot be delegated for the same reason marketing strategy cannot. For prompts that carry the standard of a deliverable, it is the senior practitioner who sets that standard. Junior staff can maintain, test and extend, and should, but the founding version belongs to whoever would have been asked the question if the machine did not exist.
There is no published Collective 54 position on prompt syntax, on specific phrasing techniques, or on which model to standardize on. That is deliberate. Those details change faster than a book can track, and the firms that win on prompting are not winning on technique. They are winning because they fed the model something nobody else has. Treat any prescriptive advice about wording, including advice found on this site, as provisional, and treat the ownership and context principles above as durable.
If the firm has not yet decided what it believes, who it serves or what good work looks like, prompt quality is not the constraint and improving it will only produce generic output faster. Settle the positioning first.
If the prompt is for a genuinely commodity task, formatting, transcription, a first pass at cleaning a spreadsheet, the ownership discipline above is more than the task deserves. Use whatever works and move on.
And if a workflow needs a prompt so elaborate that only one person can operate it, the problem is the workflow, not the prompt. Redesign the work rather than documenting the heroics.
A prompt is a piece of the firm, the judgment of the person who knows the work best transferred into a form the machine can act on, and its quality depends on how much of that judgment and how much proprietary context made it in. The generic instruction fails because it gives the model nothing a competitor could not also give it. An effective prompt carries context from the firm, such as transcripts, deal notes, frameworks and definitions of the ideal client, a defined output specific enough to check, a clear split between the eighty percent the machine does and the twenty percent a named person judges, and a checkpoint before anything reaches a client. Maintenance is the harder half: give each prompt an owner, keep the library in one place, version it, test against outcomes rather than how the output reads, retire what the work has outgrown, and feed it from finished engagements. The founder writes the prompts that carry the point of view of the firm, senior practitioners write the ones that carry a standard, and Collective 54 publishes no position on syntax because the advantage was never the technique.
Because they are generic instructions. Write me a cold email to chief information officers asks the model to work blind, with none of the context that would make the result belong to your firm, and the output is competent, fast and interchangeable. The fix is to attach what only your firm holds, client transcripts, won and lost deal notes, frameworks, pricing rationale and the definition of the ideal client, so the model works as your firm rather than as any firm.
The person whose judgment the prompt encodes. Prompts that carry the point of view of the firm, its ideal client and the reasons it wins belong to the founder, and that cannot be delegated for the same reason marketing strategy cannot. Prompts that carry the standard of a deliverable belong to the senior practitioner who sets that standard. Junior staff should maintain, test and extend the library, but founders who hand the configuration to junior staff get mediocre results.
Give every prompt that matters an owner, keep the library in one place rather than in individual chat histories, version it with a note on what changed and why, test against outcomes such as proposals won and rework avoided rather than against how the output reads, and retire prompts when the service or workflow they served changes. Feed it from finished engagements, since the delivery professional closing a project is the richest source of what should become standard.
No. Collective 54 publishes no position on prompt syntax, phrasing techniques or which model to standardize on, because those details change faster than a book can track and the firms that win on prompting are not winning on technique. They win because they fed the model something nobody else has. Treat wording advice as provisional and the ownership and context principles as durable.
Sources: Greg Alexander, The AI-Native Boutique Firm (Advantage Books, January 2027), specifically The AI Lead Generator for the position that the advantage in Era 3 is not the AI but what the founder feeds it, for the practice of building prompt libraries and contextual prompt files that attach proprietary data, client transcripts, benchmarking data, pricing frameworks, interview scripts and methodologies so the agent behaves like your firm, for the common mistakes of writing vague generic prompts such as a cold email to chief information officers, expecting AI to compensate for unclear positioning, trying to build perfect systems instead of iterating, and relying on junior staff to configure the model, for the founder as chief insight provider and the agent as execution, and for prompt libraries and contextual prompts as assets that compound over time for early adopters. The AI Delivery Professional for the 80/20 division in which AI does the production work and the human owns judgment, truth, taste and accountability, for acceptance criteria and a definition of done as the answer to perfectionism, for continuous quality control passes, and for turning project artifacts into templates, checklists, playbooks and reusable assets. The AI Marketing Manager for the finding that AI helps an uncontrarian firm produce generic faster, and for the twenty percent, vision, conviction, tradeoffs and taste, that the founder cannot delegate. The AI Operations Manager for decision memory and institutional memory as work AI should own so that knowledge survives beyond individuals. Note on scope: the four elements of an effective prompt, the maintenance practices of ownership, single location, versioning, outcome testing and retirement, and the rule that the founding version belongs to whoever would have answered the question if the machine did not exist are inferences used here to organize the source material rather than published Collective 54 positions.
Collective 54 is the private community for founders and executives of boutique professional services firms between $5M and $50M in revenue. Members work these answers against their own numbers.