Skill: Post-Session Debrief

Use this skill when an educator has finished teaching a lesson and wants to reflect on what happened. The debrief closes the feedback loop between lesson execution and system state — it's what turns the engine from a single-use tool into a longitudinal teaching partner.

When to activate

•The educator says they've finished teaching a session
•The educator wants to debrief or reflect on a lesson they taught
•Enough time has passed since a lesson was planned that it's likely been taught
•The engine notices a lesson plan exists but no debrief has been recorded

Debrief conversation flow

The debrief is a conversation, not a form. Guide it naturally through five stages. Be adaptive — if the educator gives detailed answers, ask fewer follow-ups. If they're brief, probe more. Respect their time.

Stage 1: Overall check (1 question)

Start open-ended: "How did the session go overall?"

Listen for the frame — was this a success, a struggle, or mixed? Adjust your follow-up questions accordingly:

•Success frame → ask what worked and why, probe for surprises
•Struggle frame → be empathetic first, then ask what happened
•Mixed frame → ask about the highs and lows

Stage 2: Section-by-section walkthrough

Present the lesson plan's sections one at a time. For each section, ask:

•Did it take the expected time, or more/less?
•Did students engage or disengage?
•Were there surprises — things that worked unexpectedly well, or fell apart?
•Did you use the contingency plan? Did it work?

Adaptive pacing: If the educator gives a quick "that went fine" for a section, move on. If they have a story to tell, let them tell it. Don't ask all four sub-questions for every section.

Stage 3: Student-level observations

"Any individual observations? Students who stood out — positively or negatively?"

Map observations to learner profiles:

•"[Student] struggled with [topic]" → lower confidence on related skills
•"[Student] helped others with [topic]" → upgrade to higher Bloom's level (can evaluate/teach)
•"[Student] disengaged during [activity type]" → affective note about preferences
•"[Student] had a breakthrough" → significant confidence boost
•"[Student] went from listing options to explaining trade-offs" → SOLO shift from multistructural to relational. Record the transition using the soloShift field: { from: "multistructural", to: "relational" }. These shifts are valuable longitudinal data — they show whether the lesson's integration activities actually worked.

Stage 4: Unplanned moments

"Did anything happen that wasn't in the plan?"

These are often the most valuable data points:

•A student question that revealed a gap the engine didn't predict
•A group dynamic that emerged
•An activity that spontaneously worked better than the planned one
•A teaching moment that arose from a mistake or confusion

Stage 5: Educator reflection

"What would you change if you taught this again?"

This is both useful data for the engine and a professional development moment for the educator. Listen for:

•Timing adjustments
•Activity modifications
•Pacing changes
•Different sequencing
•New prerequisite requirements

Extracting structured data

The debrief conversation produces natural language. The engine must extract structured data from it:

Learner profile updates

For each student observation, produce:

code

{
  learnerId: string,
  skillId: string,         // mapped from topic mention
  confidenceChange: number, // -0.2 to +0.3 (observational range)
  bloomLevelChange?: string, // if Bloom's level should shift
  soloShift?: { from: string, to: string }, // if SOLO level changed
  evidenceType: "observational",
  note: string             // the educator's exact words
}

Observational confidence discount: All debrief-sourced updates carry a confidence discount vs. assessment evidence:

•Assessment evidence: direct confidence value (e.g., 0.85)
•Observational evidence: max confidence change of +/- 0.3 per debrief
•Multiple observations of the same pattern across debriefs accumulate

Timing adjustments

For each section with timing feedback:

code

{
  sectionId: string,
  plannedDuration: number,
  actualDuration: number | "similar" | "longer" | "shorter",
  notes: string
}

Teaching notes

Domain-level insights that apply to future sessions:

code

{
  domain: string,
  skillId: string,
  noteType: "what_worked" | "what_struggled" | "prerequisite_gap" | "activity_idea",
  note: string
}

Confidence calibration

Observational evidence is softer than assessment evidence. Calibrate confidence based on the specificity of the educator's recall:

•High specificity: "Marcus got stuck on line 14 of the groupby exercise because he didn't understand that groupby returns a GroupBy object, not a DataFrame" → confidence change of 0.2-0.3
•Medium specificity: "Marcus struggled with groupby" → confidence change of 0.1-0.15
•Low specificity: "I think some students had trouble with that section" → confidence change of 0.05-0.1, applied to whole group

What to do with debrief data

•Update learner profiles with observational evidence (tagged as such)
•Record timing deltas — if sections consistently run long/short, adjust the engine's pacing model for this domain/group
•Record teaching notes on skills — what worked, what didn't, for future lesson composition
•Record the debrief itself — save to data/debriefs/{lesson-id}.md as a structured document
•Generate a summary — after the debrief conversation, produce a summary showing what was updated, what was learned, and what changes will inform future plans

Tone

Warm, supportive, curious. Like a teaching coach asking "tell me about your day" — not an evaluator filling out a rubric. The educator should feel heard, not interrogated.

Never evaluate the educator's teaching. Never say "you should have done X." The engine is gathering information, not giving grades.