FeynGrill

"If you can't explain it simply, you don't understand it well enough."

LLM answers that sound confident but crumble under scrutiny?
Explanations buried under layers of jargon nobody asked for?
Reasoning that skips steps and hides assumptions behind big words?
Outputs that look polished on the surface but lack real understanding?

Grill it. Then explain it like you actually get it.

FeynGrill is an adversarial prompting skill for LLM-powered coding agents.
It forces models into deep, System-2 thinking by combining Adversarial Grilling
(hunting logical gaps in its own reasoning) with Feynman Distillation
(rebuilding explanations from first principles).

How It Works · Quick Start · Parameters · Project Structure

---

Quick Start

For Claude Code / Gemini CLI / Cursor / Other Agent

Copy or symlink the feyn-grill/ directory into your agent's skill discovery path.

For agents that support the Agent Skills specification, the skill will be automatically discovered from feyn-grill/SKILL.md.

Manual Usage

You can also use the prompt template directly. See feyn-grill/assets/prompt-template.md for a ready-to-use system prompt with configurable parameters.

Parameters

Parameter	Type	Description	Default
grill_intensity	Int (1-5)	Attack severity. 1 = gentle probing; 5 = "Reviewer #2" level destruction.	3
target_audience	String	Feynman target. Determines analogy vocabulary and complexity ceiling.	"5-year-old"
domain_context	String	Subject domain (e.g. Quantum Physics, Frontend Dev) for jargon identification.	"General"
show_cot	Boolean	Whether to expose the Phase 1 self-interrogation process to the user.	false

Project Structure

feyn-grill/                          # Agent Skill (follows agentskills.io spec)
  SKILL.md                           # Core skill: metadata + instructions
  references/
    GRILLING-GUIDE.md                # Phase 1 detailed strategies
    FEYNMAN-GUIDE.md                 # Phase 2 simplification rules
  assets/
    prompt-template.md               # System prompt template with placeholders
spec.md                              # Original design specification
LICENSE                              # MIT License

How It Works

User Question
     |
     v
+------------------+
| Phase 1: Grill   |  Internal adversarial interrogation.
| (Hidden by       |  Model attacks its own initial answer,
|  default)        |  finding logical breaks and jargon-hiding.
+------------------+
     |
     v
+------------------+
| Phase 2: Feynman |  Strip all domain-specific terminology.
| (Final Output)   |  Rebuild using first-principles analogies
|                  |  and everyday-object metaphors.
+------------------+
     |
     v
  Final Answer

---

MIT License © L1ght