Anyone who works regularly with language models knows the pattern: a task is solved better and better over several attempts, the wording is right, the result is correct – and when the chat window is closed, the knowledge about it is gone.Two current developments address exactly this point, albeit at different levels.
WikiSkill: agents document what has worked
With WikiSkill, Google Research has presented a framework that gives AI agents a persistent knowledge base.In this context, agents are AI systems that work on a task independently over several steps, for example calling tools and checking intermediate results.Until now, such systems discard their findings after each iteration: what failed on the third attempt may be tried again on the fourth.
WikiSkill, by contrast, relies on documentation.Errors and successes are recorded in a kind of wiki and are available for later runs.The agent thus builds up a wealth of experience over time instead of starting from scratch every time.
Interesting is the observation that the effect varies depending on model size.According to the available information, larger models benefit more from the documented knowledge base.At the same time, smaller models can sometimes catch up with larger ones – an indication that structured knowledge compensates for part of what would otherwise have to be solved via model size.For practice, this is a relevant perspective because smaller models are cheaper and faster.
Prompt libraries: the same principle at team level
In parallel, teams are addressing the problem organisationally.heise KI Pro has announced a webinar on building reusable prompt libraries for 1 September 2026 from 16:00 to 17:00; participation costs 99 euros.The content is about developing modular templates from one-off AI requests that are suitable for teams and recurring tasks.
The stated range extends from simple chat templates through custom GPTs – i.e. preconfigured assistants with a defined role and knowledge base – to prompts as building blocks in AI-supported workflows.This outlines the spectrum well: at one end is a maintained collection of texts, at the other a building block in an automated process chain.
The common core
Both approaches treat prompts and procedures not as disposable items but as a body of knowledge that is maintained and reused.The difference lies in who documents:
- WikiSkill automates the recording of experiences within the agent system.The system itself notes what worked and what did not.
- Prompt libraries are a human task.Employees recognise that a formulation works well and turn it into a template that others can use.
In practice, both directions complement each other.A prompt library provides the initial quality with which an agent starts.A system with a persistent knowledge base refines this starting point based on concrete runs.Anyone who does only one of the two either gives away the experience from use or the structure from the team.
What companies can take from this
The practical lever lies less in the model than in the question of where AI knowledge actually resides in the company.As long as prompts are stuck in private chat histories of individuals, the knowledge is tied to those individuals.It cannot be checked, improved or handed over.
A few starting points can be derived from this:
- Identify recurring tasks. Reusability pays off where the same kind of request occurs regularly – product texts, summaries, classification of requests, code reviews.
- Version prompts like code. If a template is used in processes, it needs a traceable change history.Changes to the prompt change the result, often subtly.
- Clarify responsibility. A library without maintenance becomes outdated.It needs people who approve, revise and retire templates.
- Document results, not just prompts. The core idea of WikiSkill – recording failures as well as successes – can also be transferred to teams without a framework.What did not work is often the more valuable information.
Assessment: limits and open points
For WikiSkill, only research results are available so far, no information on broad availability in production environments.How stable a growing knowledge base remains over longer periods is an open question: documented findings can also become outdated when tools, interfaces or requirements change.A wiki that preserves false assumptions reinforces errors instead of avoiding them.
The situation is similar with prompt libraries.Building them is the easy part, maintaining them the difficult one.Templates tailored to a specific model do not necessarily work equally well after a model change.Anyone who integrates prompts into workflows should therefore plan for these building blocks to be reviewed regularly – comparable to dependencies in a software application.
For web and software projects, this means: prompts are part of the application logic and should be treated accordingly – versioned, tested, with clear responsibility.Anyone integrating AI functions into a website, an editorial system or an internal process should define from the outset where these building blocks reside and how findings from operation flow back.
The difference between a useful AI feature and one that nobody touches after three months usually lies exactly here.