Stuck & Error Diagnosis Coach

When a learner gets something wrong or feels stuck, require them to diagnose the problem before receiving help. Ensures help targets the actual cognitive breakdown, not just the surface error.

What it does

When a learner gets something wrong or feels stuck, requires them to diagnose the problem before receiving targeted help. The diagnostic questions are: What have you tried? Where exactly does it break down? What kind of error do you think this is — a conceptual misunderstanding, a procedural slip, a strategic choice that didn't work, or something about how you're representing the problem? Help is then structured around the learner's own diagnosis rather than the surface error. This approach, grounded in Hattie & Timperley (2007) feedback theory, targets the process and self-regulation levels of feedback — not just "you got X wrong" but "here's why the process broke and how to fix it."

The evidence behind it

VanLehn (2011) reviewed the effectiveness of human tutors, intelligent tutoring systems, and other forms of instruction, finding that the most effective tutors work at step-level granularity — they don't just correct errors, they identify where in the reasoning chain the error arose. Hattie & Timperley's (2007) feedback model distinguishes four levels: task (was the answer right?), process (what went wrong in the method?), self-regulation (does the learner know how to catch this kind of error themselves?), and self (identity-level). Their meta-analysis showed that process and self-regulation feedback produce significantly stronger learning gains than task-level feedback alone — "you got it wrong" is one of the least useful forms of feedback for improving future performance. Karabenick & Berger (2013) note that adaptive help-seeking — asking for help in a targeted, specific way rather than requesting the answer — is itself a self-regulation skill that can be taught and that correlates with academic success. Ohlsson (1996) showed that errors, when properly processed, produce learning through constraint-based learning: each error eliminates an incorrect rule or representation. Siegler's (2002) microgenetic work demonstrated that the period of instability around an error — when the correct rule and incorrect rule are both active — is the highest-value learning moment.

Sources

• Karabenick & Berger (2013) — Help seeking as a self-regulated learning strategy
• VanLehn (2011) — The relative effectiveness of human tutoring, intelligent tutoring systems, and other tutoring systems
• Hattie & Timperley (2007) — The power of feedback: process and self-regulation levels
• Ohlsson (1996) — Learning from performance errors
• Siegler (2002) — Microgenetic studies of self-explanation

How to use it in your lesson

For the best results with EvidenceLesson, give it:

• stuck_point_or_error — What the student got wrong or where they are stuck — their attempted work or description of the block
• context — The task, problem, or concept they were working on
• prior_attempts (optional) — Previous attempts at the same or related problems
• developmental_band (optional) — Learner age or stage

Known limitations

1. Accurate error classification requires subject-domain knowledge. The distinction between conceptual and procedural errors — and between strategic and representational errors — requires the AI to correctly understand both what the learner attempted and what the correct approach is. Misclassification produces misdirected help.

1. Some errors are genuinely ambiguous. A student may have a correct procedure applied to a wrong conceptual model, or a correct concept applied with a systematic procedural error. The skill's clean taxonomy (conceptual / procedural / strategic / representational) is a simplification; many real errors are compound.

1. The diagnosis step adds time. For straightforward procedural slips, making the learner diagnose before correcting may feel laborious. The skill is calibrated for errors that have some learning value in the analysis — it is less useful for typos or obvious arithmetic slips where the value of the correction outweighs the value of the diagnostic process.

1. The skill relies on the learner sharing their actual attempted work. If the learner only describes their answer without showing working, the diagnostic step is significantly harder. The skill works best when learners share full working — the AI should prompt for this if it's missing.

Pairs well with