Opening · Movement I 01 / 15
Sessions · 4 months · one engineer
833
What your AI conversations say about how you think.
Gabriel's mirror mvmt i · 01 / 15
Evidence · six studies 02 / 15
The empirical case for cognitive debt is no longer debatable.
METR · 2025
−19%
Actual speed. Engineers believed they were 20% faster. A 39-point perception gap.
Kosmyna / MIT · 2025
83%
Recall failure. EEG connectivity decline persisted after the tool was removed.
Google DORA · 2024
−7.2%
Stability drop. 75% of engineers felt more productive than before.
Gerlich · 2025
↓
Negative correlation between AI usage frequency and critical thinking scores.
Shen & Tamkin / Anthropic · 2026
−17%
Comprehension deficit. The understanding gap, not just the output gap.
Rani & Danry · 2025
−15.3%
Skill decline over 4 weeks. Reflection helps in-session — but unassisted skills erode.
Gabriel's mirror mvmt i · 02 / 15
Pattern · CHI 2026 · Barcelona 03 / 15
32 papers. Three design principles converged independently.
I · Friction
Friction is a feature.
Kocaballi et al. — "Drag or Traction" · productive friction outperforms frictionless automation
II · Artifacts
Intermediate artifacts as interface.
Sivaraman et al. — transparency designed for the user, not the system
III · Loop
The metacognitive loop as the unit of design.
Forshaw — "The Metacognition Paradox" · Abraham & Čučkov — "Felt but Not Seen"
Gabriel's mirror mvmt i · 03 / 15
Reframe · the missing study 04 / 15
Every study is single-session. Non-persistent. Measured in hours, not months.
The gap
Nobody has measured what happens across 833 sessions of real work.
Gabriel's mirror mvmt i · 04 / 15
Question · Movement II 05 / 15
How do you actually use AI?
Gabriel's mirror mvmt ii · 05 / 15
Mechanism · the mirror 06 / 15
No thesis in the prompt.
01 · Sources
Extract
4 sources — code, documents, research, planning conversations
→
02 · Analysis
Batch analyze
42 batches of 20 · rolling context from the previous batch
→
03 · Synthesis
Synthesize
Larger model · all 42 summaries in chronological order
The patterns emerged from the data. They were not embedded in the prompts.
Gabriel's mirror mvmt ii · 06 / 15
Contrast · rolling context 07 / 15
The rolling context is the architecture.
Batch 01 — no prior context
Patterns in isolated sessions.
The model sees twenty conversations cold. It identifies what happened inside them — but nothing about how behavior changes across time.
Batch 40 — full longitudinal record
Change across months.
The model reads the current batch alongside a compressed record of everything before. It notices drift, discipline, collapse. It can name the arc.
Gabriel's mirror mvmt ii · 07 / 15
Evidence · interaction signature 08 / 15
833 sessions. One interaction signature.
40%
Collaborative iteration
Not delegation. Not automation. Building toward something together.
30%
Execution
Clear task, clear output. The AI completes; the engineer moves on.
20%
Sensemaking
Thinking out loud. The AI as a surface to think against.
10%
Research-first
Source before generate. Friction before fluency.
"judgment was most active precisely where AI confidence was highest"
Gabriel's mirror mvmt ii · 08 / 15
Duality · same program 09 / 15
Program A
Build tools to make AI more useful.
Extend capability. Reduce friction. Ship faster. The practitioner as multiplier.
sameprogram
Program B
Build tools to protect thinking from AI.
Preserve cognition. Maintain judgment. Keep the reasoning muscle active under load.
"the tools he is publishing as productivity multipliers are the same tools he is actively patching mid-use"
Gabriel's mirror mvmt ii · 09 / 15
Evidence · failure 10 / 15
14
Near-identical prompts · one pressure day
Discipline collapsed.
Not because the answers were wrong. Because I wanted a different answer. The pipeline found it. The pattern was in the data.
"the exact moments when human agency matters most are the moments when it is hardest to sustain"
Gabriel's mirror mvmt ii · 10 / 15
Yield · self-authored 11 / 15
"the goal is cognitive compression and model refinement, not dependency."
— self-authored · surfaced by the pipeline, months later, from a session since forgotten
Gabriel's mirror mvmt ii · 11 / 15
Pattern · Movement III 12 / 15
Patterns become skills.
Skill
Pattern from the data
CHI 2026 principle
Practice
Research Gate
10% of sessions began with sourcing, not generating
Friction is a feature
Stop. Source first.
Evidence Mode
Sensemaking sessions produced better subsequent outputs
Intermediate artifacts as interface
Build the draft before the draft.
Scope Guard
Judgment most active where AI confidence was highest
Generative refusal (Sora Kang)
Refuse the completion.
Handoff
Context collapsed between sessions — the forgetting gap
Persistence across sessions — the field's gap
Carry context forward.
Reflect
Metacognitive noticing correlated with better subsequent sessions
The metacognitive loop as unit of design
Close the loop, in session.
Gabriel's mirror mvmt iii · 12 / 15
Mechanism · the instrument 13 / 15
SESSIONS 833 conversations ANALYSIS 42 batches PATTERNS emergent SKILLS 5 practices BETTER SESSIONS
The tool learns from you — then teaches you back.
The single-session studies miss the loop entirely. You cannot measure the flywheel in an hour.
sessions → analysis → patterns → skills → better sessions
Gabriel's mirror mvmt iii · 13 / 15
Handoff · the instrument 14 / 15
The pipeline is available.
Your sessions are waiting.
You'd get different skills. Your sessions would surface your patterns, not mine.
# export your AI assistant sessions
$ mirror export --source cursor --out sessions/
 
# run the analysis pipeline
$ mirror analyze --batches 42 --rolling
 
# synthesize patterns → skills
$ mirror synthesize --out report.md
 
# open your report
$ open report.md
Gabriel's mirror mvmt iii · 14 / 15
Closing · the argument 15 / 15
the scarce resource is structured reasoning, not intelligence.
the question is whether you're practicing it —
and now you can measure.
Gabriel's mirror mvmt iii · 15 / 15