PK Wiki

concept · concepts/Metacognitive Strategies.md

Metacognitive Strategies

status: active · created 2026-11-08 · 2 sources · #metacognition #cognitive-science #heuristics #teaching #thinking-and-learning

Small, learnable chunks of knowledge about how to steer your own thinking, such as "look for a problem's deep structure," "consider both sides," or "don't accept the first reasonable-sounding solution." They make good thinking more likely, but they tell you what to do, not how. The how comes from domain knowledge and practice (Willingham — Critical Thinking, Why Is It So Hard to Teach).

Cue, not capability

  • What they do: remind you to act at the right moment. A student repeatedly encouraged to see both sides is more likely to think "I should look at both sides" unprompted. The modest benefit of generic critical-thinking programs is probably this.
  • What they don't do: supply the knowledge to act. Told outright that the band problem resembled the garden problem, most subjects (65%) still couldn't solve it. You may know not to accept the first plausible solution, yet not know how to generate or weigh alternatives.
  • Why they need context to fire: inside a domain there are "more context cues" for which strategy applies. "I'm designing an experiment" triggered control-of-variables seven months later (Chen & Klahr 1999).

How to teach them (Willingham's staged recipe)

  1. Make it explicit. Introduce the strategy with at least two different examples, possibly from learners' own experience.
  2. Label it as a reusable strategy, so it gets a handle that isn't tied to the first example.
  3. Apply it to the content at hand, never in isolation.
  4. Fade the cue. Later, just name the strategy and let the learner work out how it applies. Eventually they pick the right strategy unprompted.

A live instance: the human's "teach as we go" rule

The human's standing instruction to their AI agents (from their global agent config, noted 2026-11-08) is: when making a judgment call that could have gone the other way, state the rule of thumb or first principle behind it in one line, so I can internalize it. This closely matches Willingham's recipe:

  • The maxim is labeled (stated as a rule of thumb) and attached to the content at hand (a real decision in real work), not taught in the abstract.
  • It recurs across many surface variants (different tasks, same principle), which is how deep structure becomes perceptible.
  • Willingham predicts what it can't do alone: the one-line maxim is a cue. Whether the human can apply it next time depends on the domain knowledge the surrounding work builds. A deliberate next step would be the "fade the cue" stage: name the principle and let the human work out how it applies.

Beyond education

  • Working Heuristics — Data, Matching, Product, Org, Tooling and the craft concept pages it spawned are, in Willingham's terms, a library of metacognitive strategies. Each will help most where the reader already has the domain knowledge, and least as a stand-alone rule. See Method Reflection (T1).
  • Falsification Is Asymmetric is itself a metacognitive strategy ("seek the disconfirming case").
  • The same pattern in AI agents. Anthropic — The Components of Agent Memory found that telling Claude what memory is for, not how to use it, worked best for strong models. A fixed rule ("record only failures") cut their gain from memory. A weaker model (Sonnet 4.6) nearly doubled its score when explicitly cued to read its memory and test its theories. Newer models did this without prompting. This mirrors Willingham's recipe: cues help most at the novice stage, and "fade the cue" is right for the expert. Rigid procedure can get in the way of a capable judge. See Agent Memory.

Connections

Sources

The wiki is the source. This door reads it at build time and never writes to it.