When you write down what you’ve learned, what kind of thing are you writing?
Most people don’t ask. They dump insights into notes, checklists, and rules without noticing that different types of knowledge serve different moments. A rule that says “always do X” is solving a different problem than a checklist that says “consider these five dimensions.” Conflating them makes both weaker.
After mining tacit knowledge from hundreds of AI-assisted work sessions, I’ve found ten distinct types. Each one answers a different question and helps at a different moment.
The Ten Types
1. Rules — “Don’t hold finished work back from stakeholders.” Single direction, always applies. You either follow it or you don’t. Rules help when you know what to do but might not do it. They’re guardrails against your own rationalisation.
2. Checklists — “Five quality axes for session wrap-ups: signal-to-noise, routing accuracy, resume fidelity, insight yield, time proportionality.” Multiple dimensions to weigh against each other. Checklists help when you don’t know what to consider. They expand your view beyond whatever dimension you’d naturally fixate on.
3. If-then triggers — “If your wrap-up note has ‘TODO: consider…’ — decide now or delete it.” Conditional rules that fire on recognition. More precise than rules because they specify the situation. The “if” part is the real value — it trains you to notice.
4. Distinctions — “Log vs. insight: logs record what happened, insights record what was learned. Different audiences, different destinations.” Prevents you from conflating two things that look similar but need different treatment. The most common failure mode in knowledge work is treating everything the same.
5. Anti-patterns / smells — “Deferred action disguised as capture.” Recognition of what’s going wrong, without prescribing the fix. Useful because the fix depends on context, but the smell is universal. You need to notice it before you can fix it.
6. Spectrums — “Scale wrap-up depth with session complexity. A 10-minute admin session doesn’t need the same close-out as a 3-hour build.” Tells you there’s a dial, not a switch. Prevents binary thinking. Many rules are actually spectrums in disguise — “always do X” is often “do more X when stakes are higher.”
7. Ordering / priority — “Route knowledge to the most specific destination first: skill > memory > daily note.” Sequence matters. Do this before that. Often the difference between effective and ineffective isn’t what you do but in what order.
8. Signals — “Inconsistent depth from an AI model is the signal that extractable structure exists in the weights.” Tells you what to notice, not what to do about it. The interpretation step before action. Signals are the hardest type to capture because they require pattern recognition, not just compliance.
9. Defaults with override — “Always checkpoint at gear shifts — unless the session is under 10 minutes.” A rule with an explicit escape clause. Different from a plain rule because it teaches you when to break it. Without the override, people either follow the rule slavishly or ignore it entirely.
10. Reframes — “A wrap-up is state transfer, not documentation.” Changes how you see the task, which changes every decision downstream. Not directly actionable in isolation, but makes all the other types cohere. The organising principle that turns a pile of rules into a system.
Why This Matters
Most “lessons learned” exercises fail because they capture everything as rules. But a checklist stored as a rule loses its multi-dimensional nature. A spectrum stored as a binary loses its nuance. A signal stored as a rule becomes a false positive factory.
When you name the type, you sharpen the capture. You also know when you’re done — if your knowledge base has rules but no distinctions, you’re probably conflating things that deserve separation.
The Recursive Floor
You can apply this taxonomy to itself. These ten types are a checklist (type 2) for knowledge extraction. The taxonomy emerged from mining the mining process — extracting how extraction works.
But the recursion has a natural floor: when the meta-level stops changing your behaviour at the object level. “Types of actionable knowledge” changes how you extract. “Types of types of actionable knowledge” doesn’t. That’s where you stop.
The test for any meta-work: does it make the next concrete task better? If yes, go deeper. If you’re just admiring the structure — you’ve gone too far.