You finish an article with an idea you would like to try. Perhaps your organisation could handle a recurring problem differently, or a scientific explanation has suggested an experiment. You open a large language model, describe the idea, and ask for a plan. Now comes the harder work: deciding what the model needs to know, which questions it should ask, and whether its apparently sensible steps would survive contact with reality.
This is the gap Alkemata wants to investigate. We already use AI to help write articles and organise the publication, while human editors define its topics and editorial direction. Our next experiment would carry that collaboration beyond the published page. We call the proposed format a capsule: an idea accompanied by the knowledge and working method a reader can take into their own LLM to develop a project.
The purpose is to make understanding easier to acquire and easier to act upon. That remains a hypothesis to test. This article describes the design we want to explore; it does not announce a working capsule service.
Publishing the method behind an idea
Putting a chatbot on Alkemata would leave an obvious question unanswered. Readers who already have access to a capable model can ask it to explain an article, challenge its conclusions or suggest applications. Maintaining another conversational interface would consume resources without necessarily giving them much that is distinctive.
The more useful editorial contribution may be to make the method explicit. Knowing which assumptions to examine, which evidence to request and which failure to test is part of expertise. A reader should be able to obtain that know-how along with the idea, without first becoming skilled at writing prompts.
A capsule would therefore bring together a causal explanation of the problem, the evidence supporting it, its uncertainties, and a workflow for applying it. It would include questions that reveal whether the method fits a particular situation, examples of misleading results, and materials such as an interview guide or experimental protocol. The article would remain the readable argument; the capsule would support further work.
Packaging procedural knowledge for AI is already an established design direction. Anthropic’s description of Agent Skills, for example, combines instructions and supporting resources in folders that an agent can consult. Alkemata’s proposed contribution concerns publishing: how an inspiring idea, an inspectable method and experience from readers might form a continuing editorial process. We would have to demonstrate the usefulness of that combination.
Take the capsule to the model you use
The intended starting point would be simple. A reader chooses a purpose, copies or uploads the capsule into a compatible LLM, and adds a sentence about their situation. The capsule itself supplies the instructions for beginning the conversation. Someone could ask to understand the idea, assess an existing process, or develop a small experiment.
A short capsule could be pasted as ordinary text. A longer one could be downloaded as Markdown, a text format with readable headings, and accompanied by templates. A public link would offer another route where the model can retrieve it. File support, web access and tool access differ between services, so these routes need testing. Portability is a design requirement, not a promise of identical behaviour everywhere.
The model should ask a few relevant questions before proposing work: what the reader hopes to change, who would be affected, and what resources or constraints matter. It would then explain the necessary mechanisms while applying them. A reader examining a workflow might be asked to predict what happens when one step fails before seeing the proposed test. Learning would accompany the project rather than become a prerequisite course.
Consider the principle developed in A Public Service Should Give You a Receipt. A possible capsule could help a reader investigate a university application portal. It would guide them through recording what is submitted, what acknowledgement appears, and which questions remain unanswered. This is an illustrative project, not a report of a completed trial.
The resulting work might include a process map, an interview guide and test cases for missing attachments or delayed acknowledgements. The first experiment could be a paper walkthrough using invented application data. The reader would have something to examine with an administrator before proposing software changes. Any conclusion about legal evidentiary value would still need the appropriate expertise.
The workflow is where the editorial effort belongs. “Make an action plan” says little about what makes a plan good. A useful capsule specifies which assumptions need testing, what would count as failure, and when the project should stop or change direction. That follows the method of Ideas at the Limit: explore the ambitious possibility, understand its machinery, and return to a manageable test.
When the reader brings experience back
A publication built around capsules would need a return route. After trying a method, a reader might discover that a key assumption was wrong, produce a useful template, or reach a question requiring someone else’s experience.
With the reader’s agreement, their LLM could prepare a short project record: the originating capsule and version, the objective, the current hypothesis, what has been tried and the next milestone. Submitted to Alkemata, this “project passport” could support a page showing the current state of the work. Participation would be optional; using a capsule should not require registering a project or making it public.
A subsequent field report would explain what happened and what evidence supports the account. It should distinguish observed results from model-generated expectations. A polished conversation is not evidence that a project worked. Paused or abandoned attempts would matter too, particularly when they expose a condition under which the method fails.
Editors would review these contributions. An observation might justify a correction, a contextual note or a new variant, rather than a change to the general method. Each accepted revision would retain its source and explain what changed. The next reader could then start with knowledge that includes the difficulties of the previous attempt.
Help could follow the same structure. A project owner might ask for someone to review three accessibility test cases or investigate one disputed assumption. A request would explain the contribution sought and its likely scope. A public directory of these needs could be read directly or taken into a contributor’s own LLM, alongside information about their skills that need never be submitted to Alkemata.
The aim would be to make useful encounters easier without demanding continuous participation. Contributions would attach to an open question, a test or a practical document. Credit would describe work done, and pseudonyms could protect professional boundaries. Reports and contact details would become public only through an explicit choice and review.
Where human judgement has to remain
This design creates responsibilities on both sides of publication. Editors would select ideas, check claims, develop workflows and review revisions. AI could help draft, challenge and test candidate materials. Readers would supply context, choose objectives and decide what actions to take. Qualified practitioners and affected people would remain necessary where a model’s output cannot settle the question.
A capsule can also transmit a mistake more efficiently. An incomplete method could be repeated across many projects; an authoritative tone could discourage a reader from challenging it. Alkemata would need to make sources and limits visible, as discussed in What Makes an AI Answer Verifiable? A version marked experimental must not quietly acquire the authority of a method tested in practice.
Instructions in a file do not guarantee that a model follows them. The reader should be able to inspect outputs and see which conclusions depend on unverified information. Preparing a document also differs from sending it, and drafting code differs from deploying it. External actions would require the user’s authority and the relevant safeguards. Capsules should remain useful even when the model can only help prepare the next step.
Nor does moving work to a reader’s LLM eliminate cost or privacy concerns. Access may cost money; sensitive project information may be unsuitable for that service. A readable capsule should offer useful explanations and manual alternatives, while public reports should contain only material the contributor is entitled and willing to share. Editorial maintenance and moderation would remain real work for Alkemata.
What would make the experiment worth continuing?
The strongest objection is that a capable reader and a good model may already do this well. A capsule could add paperwork, constrain exploration, or produce a longer plan without improving understanding. We should compare it with the simpler alternative: giving the same model the article and an ordinary request for help. The comparison should examine useful first actions, unsupported assumptions, time and effort, and what the reader can explain afterwards.
Two further possibilities seem worth testing. One is a portable project checkpoint: a short, reader-approved account of decisions, evidence, unresolved questions and the next action. It could help someone resume in another conversation or model without transferring an entire private chat. Portability would then extend to work in progress.
The other is a brief understanding check. Before a consequential step, the reader could explain why the proposed action should work and what observation would change their mind. The aim would be to discover whether using the capsule has strengthened their judgement. A convincing output alone would not answer that question.
A first trial should be modest: one capsule, a few readers, a concrete task and an honest record of where the method helped or failed. Readers who have a suitable case, practical expertise or a reason the approach might fail can use the existing Collaborate page. The question is whether an idea can leave the publication, become useful work in someone else’s hands, and return with knowledge the publication did not previously possess.
