Add deep-research skill: self-challenging investigation with adversarial dialectic

[Misterbra]

What this skill does

Teaches Claude to argue with itself before recommending anything. Instead of the usual one-pass research (search → recommend → done), it enforces an iterative loop where Claude challenges its own findings, adopts adversarial positions, and only commits when ideas survive scrutiny.

Core mechanisms

• Adaptive auto-questions — generated from what was just discovered, not a static checklist
• Adversarial dialectic — thesis → antithesis → critique of the critique → synthesis. Includes challenging foundational assumptions, not just the proposed angle
• Autonomous execution — runs the full loop without stopping to ask permission between phases
• Checkpoint & reset — saves state to memory, enables fresh-perspective reattacks

Benchmarked

Tested on the same topic with three configs (baseline, self-questioning only, full dialectic):

	Baseline	+ Self-questioning	+ Adversarial dialectic
Angles killed	0	4	3 (documented)
Self-corrections	none	yes	yes + dialectic
Foundational assumptions challenged	no	no	yes
Final angle type	descriptive	descriptive	causal

The dialectic changed the nature of the recommendation, not just its quality.

Links

• Full README + benchmark details: https://github.com/Misterbra/deep-research-skill