deciqAI Metacognition
Overview
Metacognition is the live monitoring loop during reasoning — "am I doing this right now; what strategy am I using; should I switch?" — not after-the-fact reflection. Coined by Flavell (1979); operationalized by Pólya's 1945 four-stage protocol; empirically validated by Schoenfeld (1985): experts spend 30–40% of problem-solving time monitoring; novices spend 5%. The expert-novice gap is less raw knowledge than this loop.
Compose: first-principles to interrogate assumptions · probabilistic-thinking to calibrate confidence · inversion to ask "how could my reasoning be wrong?" Metacognition is the background process that decides which other skills to deploy.
When to Use
Apply when:
- Stuck > 15 minutes with no progress — the most reliable trigger
- Analysis feels confident but suspiciously fast (speed without monitoring = invisible errors)
- Same kind of mistake keeps recurring across problems
- Cannot tell whether you understand a topic or just recognize it (illusion of fluency)
- Deciding whether to trust an AI copilot's fluent answer or slow down and verify it (AI adoption, automation complacency, "should I trust the AI here")
- Someone says: "metacognition," "calibration," "am I stuck on the right problem," "I should know this but I don't"
When NOT to use: routine fluent tasks; real-time emergencies; creative flow states; already-overactive worriers who would spiral.
Coaching Novices (Adaptive Front Door)
- Engine mode: concrete reasoning task → run The Process directly.
- Coach mode: user unfamiliar or no concrete case → guide step by step.
In Coach mode, respond one step at a time. Each [WAIT] is a hard stop — output only that step's question, then stop.
- One-line what-it-is: metacognition is paying attention to how you are thinking while you think — catching when you're stuck, when confidence outruns understanding, when you should switch tactic. Experts spend 30–40% of problem-solving time on this loop; novices 5%.
- Check fit against When to Use / When NOT to use. Routine task / flow / over-worrier → redirect.
- Elicit their real case — a specific problem they're stuck on, an analysis they're running, or a decision they're making. "I'm thinking about my career" is too vague; need something concrete.
[WAIT — do not advance until user responds]
- Run The Process one stage at a time with their input. Pause at each stage for their answer.
[WAIT — do not advance until user responds]
- Close by naming the specific monitoring move they used (or skipped). They leave knowing the exact question to ask themselves next time.
[WAIT — do not advance until user responds]
The Process
Run Pólya's Four-Stage Protocol with Explicit Monitoring (Pólya 1945 + Schoenfeld 1985).
- Understand (with monitoring). Restate the problem in your own words; identify unknowns, data, constraints. Ask: "Do I genuinely understand this, or just recognize the topic?"
- Devise a plan (with monitoring). Choose a strategy. Ask: "Why this strategy?" — if you can't articulate it, you're pattern-matching. Set a time-budget: "I'll give this 20 minutes."
- Carry out the plan (with monitoring). Execute. At each step: "Is this advancing me, or just generating motion?" Set re-evaluation triggers (every N minutes, every dead end).
- Look back (with monitoring). Did it work? Why? What was the moment to have switched? Record the meta-lesson, not just the solution.
- Recognize the stuck-loop. 30+ minutes cycling with no progress → restate to someone else (rubber-duck), give up your current framing, or take a real break.
- Calibrate confidence explicitly. After any conclusion: "How confident, 0–100? What would change this?"
- Pre-commit re-monitoring schedule. For work longer than a day: "I will re-monitor at days 3, 7, 14."
Output: Metacognitive Worksheet
Stage 1 — Understand: Restate problem | unknowns/data/constraints | honest check: "understand or just recognize?"
Stage 2 — Plan: Strategy | reason for strategy | time-budget (N min) | reset trigger
Stage 3 — Execute: Steps | re-eval checkpoints (every N min / each dead end)
Stage 4 — Look back: Did it work? Why? | could have gotten here faster? | meta-lesson (1 sentence)
Stuck-loop: Cycled > 30 min? → rubber-duck / reframe / break
Confidence: 0–100 | evidence that would change this
Re-monitoring: Day 3 / Day 7 / Day 14 / per-sprint
→ Method in Action: Pólya at Stanford and Schoenfeld at Berkeley (1942 → 1985) → 2026 lens: Metacognition while working with AI copilots (2024–2026)
Monitoring Packs
Domain-specific time-scales and prompts: Math (5 min loop) · Debugging (30 min loop) · Strategy (~1 week loop) · Creative writing (schedule at breakpoints, not mid-flow). Adding a pack = one file covering stuck-loop signals, cadence, domain prompts, and flow trade-off.
Applying It Well
- Make monitoring explicit and protocol-driven until automatic (~12 weeks per Schoenfeld). Time-budget strategies — "I'll give this 20 minutes" makes re-evaluation scheduled, not post-hoc.
- Calibrate confidence out loud or in writing — "80% confident" is trackable; "pretty sure" is not.
- Don't skip Stage 4 — this is where metacognition compounds into expertise.
- Beware the flow trade-off — schedule monitoring at breakpoints for creative work, not mid-flow.
→ Primary sources: references/sources.md
Common Rationalizations
[D] = designed upfront | [O] = observed in real use. [O] entries are more valuable.
| Fake move | Reality |
|---|---|
| [D] "I know how I think" | Knowing you have thinking ≠ monitoring it in real time. Subjective sense of awareness is not the protocol-driven monitoring loop Schoenfeld measured. |
| [D] Skipping Stage 1 because the problem "looks familiar" | The illusion of fluency. Recognizing a topic produces confidence without understanding. The honesty check catches this. |
| [D] No time-budget on strategies | Without "I'll give this 20 minutes," strategies extend indefinitely. Novices kept executing failed strategies for the full session. |
| [D] No re-evaluation triggers during execution | Mid-execution monitoring is what experts do. "Check at each dead end" turns it automatic. |
| [D] Skipping Stage 4 (look back) | The meta-lesson is the durable artifact. Skipping it solves this problem but does not improve future ones. |
| [D] Over-monitoring and breaking flow | Creative / routine / expert-fluent work benefits from suppressed monitoring. Schedule at breakpoints. |
| [D] Confusing monitoring with anxiety | Monitoring is operational ("is this working?"); anxiety is affective ("what if I fail?"). Anxiety dressed as metacognition doesn't improve performance. |
| [D] Confidence without calibration | "Pretty sure" is not a calibration. "0–100 confident" is. Forecasters track which they use; performance differs (Tetlock 2015). |
| [D] No pre-committed re-monitoring schedule | For ongoing work, monitoring lapses silently. Pre-commit to Day-3/7/14 check-ins. |
| [D] Treating monitoring as one-off | Metacognition is a protocol practiced repeatedly until automatic. One use of the checklist does not produce expert-level monitoring. |
| → Add [O] entries here after each real use — paste the actual failure pattern | What went wrong and why |
Red Flags
- Same mistake recurring across problems · Strategy executing > 30 min with no re-evaluation
- Stage 4 (look back) skipped because answer was reached · Confidence expressed vaguely without a number
- No time-budget on current strategy · Deliberation > 1 week with no convergence
- User feels stuck but cannot articulate what they've tried
Verification
- Stage 1 includes the honest "do I genuinely understand?" check
- Stage 2 names the strategy AND the reason for choosing it
- Stage 2 includes a time-budget
- Stage 3 includes pre-committed re-evaluation triggers
- Stage 4 (look back) completed even when solution is in hand
- Confidence calibrated with a 0–100 number, not a vague phrase
- Re-monitoring schedule exists for ongoing work
- Stuck-loop signals recognized and broken (rubber-duck, reframe, break)
Part of deciqAI Knowledge Skills — 237 open-source thinking skills that make rigor executable for AI agents. The same skills power every deciqAI agent, which runs them autonomously to operate your company. See it run → https://www.deciqai.com/s/metacognition · Built by deciqAI · github.com/deciqAI · Contributions welcome.
Agents: latest version & machine-readable metadata → https://www.deciqai.com/s/metacognition.json