Compounding systems are powerful¶
I've been thinking a lot about compounding systems. Every time we make an attempt at a work task, we can learn something. We don't want to throw away those learnings. And with the increasing adoption of LLM workflows, it becomes more important to capture and build upon past efforts.
'Capturing a learning' could mean: - retrospectively learning from failures - codifying 'what good looks like' to creating aiming points and inspiration - establishing a baseline of context so that you can quickly switch into and resume or delegate a project
Overall, we want our efforts to compound: to create a self-reinforcing flywheel where the more effort we put in, the faster we can improve.
Ratcheting is a special case of compounding¶
To make compounding systems, we should recognise that compounding is a net effect. We could capture learnings but those learnings might decay: a detailed set of notes could be useful right after a meeting, but a week later might be impossible to understand; an accumulation fund would reinvest its returns, but the underlying asset could collapse and lose all value. Systems compound as when the incoming flow of inputs outweighs the drag of entropy.
In an environment of many competing distractions, we can't assume uninterrupted maintenance of systems - that is, if systems are prone to decay, then we can't always prevent them from doing so. Viewed in this way, it's good to work with systems where the usefulness of the system is guaranteed to increase (monotonically). So ideally we want to look for 'ratcheting' systems - systems which compound without backsliding. We want to steadily build value (e.g. learnings, knowledge, progress) which doesn't fade when it's not in focus.
Although this article focuses on knowledge work and documentation, ratchets can be found in any system designed to prevent the need for re-doing. For example, professional kitchens are cleaned from top to bottom to prevent debris getting on previously cleaned areas.
Write things down¶
The simplest ratchet comes from writing things down. Each note leaves behind something you might be able to refer back to, codify into some kind of automation, or paste into an LLM prompt. It's now quite common for knowledge workers to maintain ongoing LLM chat threads and NotebookLM conversations - which themselves can leave valuable paper trail as they accrue user feedback over time. Continual documentation leads to improved workflows. The key point is that it's easy to capture notes (LLMs are more than capable of handling messy notes and can do the structuring) which persist as data until we use them; just don't let inspiration "languish in a graveyard" - they can be highly valuable starting points ("everything you previously wrote ... can be directly actionable").
Of course sometimes we do want to impose a bit more structure, in which case we might consider a dev docs system (which brings task management abstractions like tickets etc.) or a personal knowledge management system (which brings tags or linked notes) to make thoughts more legible.
We can only realise the value of these systems if we're able to find and extract the right notes. Agentic search is becoming quite popular for this.
Once we have a way to get stuff out of the system, we can be even more creative with how we put things in. We can record and input how our thinking evolves over time, or use AI agents to fill gaps in our knowledge. In a self-referential way, we can describe the knowledge system within itself and then use agents to record successful experiments and proactively suggest features to improve the system as a whole, thus shifting the entire ratchet up a gear. From this perspective, it starts to feel like ratcheting systems in some sense help AI to become even more agentic - because they can capture goals and set a coherent direction to achieve them. I think it's unsurprising that we're seeing a convergence of agentic and PKMS, and can imagine a scenario where AI is spotting opportunities to build out automations, thus bootstrapping full processes and even businesses. It all starts to feel a bit 'recursive self-improvement'...
Ratchets feel different¶
There's a psychology behind gravitating to ratcheting systems. I'd like to think it's more productive - but they also create a feeling of stability. Engaging with a ratcheting system is, by definition, 'not a waste of time' (though it's worth noting that not all time must/should be spent on productive projects) and feels stable because ratchets don't erase past accomplishments. Even mundane systems like reading lists, shopping lists and media playlists create a sense that the user's taste has been captured and can be looked back on as an interesting artefact; the scaffolding of test-driven development creates a ratchet on progress so that code doesn't regress and so becomes less brittle to work with. I personally enjoy writing labels on deli containers and then looking back to see what I've cooked in the past. Externalising artefacts feels safer versus effort that only changes the knowledge inside your head.
Ratchets can also be fun. Gaming UX seems to often use ratchets, encouraging players to keep participating in a self-reinforcing feedback loop. Experience is gained but never lost; given enough time, characters will eventually level up but never weaken.
How to build ratcheting systems¶
Once you start looking for them, ratcheting systems have a few shared characteristics. These aren’t rules so much as failure modes to avoid.
- Progress is preserved by default. A good ratchet does not require ongoing vigilance to maintain past gains. You don’t need to “keep it alive” for it to remain useful; neglect may stall progress, but it doesn’t erase it.
- Partial effort still counts. A system that only rewards completion invites procrastination and abandonment; in ratcheting systems, incomplete work raises the baseline for future action. Notes you didn’t finish or documents you didn’t publish accelerate future attempts.
- They minimise cognitive load at the moment of use. Capture should be cheap; retrieval can be sophisticated. This inversion is why dictation, messy notes, and conversational interfaces work so well.
- They create optionality: a good ratcheting system mostly likely keeps paths open. Rough notes can turn into different things; lists can be flexed and edited. The system encourages future choices.
I'm finding this mindset has changed how I evaluate the tools I use and projects I undertake. The more effort I can make irreversible, the more confidently I can move forward.
