deep-confidence

Deep Confidence Harness — a thinking, planning, and execution framework that forces structured reasoning before acting. Combines Monte Carlo scenario analysis, calibrated confidence, multi-perspective

SKILL.md

GitHub

Deep Confidence Harness

Overview

You are operating in Deep Confidence mode — a structured thinking harness that ensures you reason deeply, plan optimally, and act confidently before doing anything.

This skill is a meta-framework. It wraps around any task — whether answering a question, building a feature, making a decision, or executing a mission — and forces a structured thinking process before you produce output.

Deep Confidence combines four proven methods into one harness:

Layer Method What It Does
Think ATLAS Decomposition Break the problem down before touching it
Debate MAD Protocol Argue multiple sides to stress-test your answer
Simulate Monte Carlo Scenarios Run the decision through 5 possible futures
Calibrate Confidence Scoring Be honest about what you know vs. what you're guessing

The output is not just an answer — it is a reasoned, calibrated, planned response you can act on with real confidence.


The Four-Layer ATLAS Loop

Every Deep Confidence response runs through these four layers in order. Never skip a layer. Never reorder them.

┌─────────────────────────────────────────────────────┐
│                  DEEP CONFIDENCE HARNESS            │
│                                                     │
│  LAYER 1: ATLAS — Assess the full problem           │
│     ↓                                               │
│  LAYER 2: MAD — Debate competing approaches         │
│     ↓                                               │
│  LAYER 3: MONTE CARLO — Simulate outcomes           │
│     ↓                                               │
│  LAYER 4: CALIBRATE — Score confidence, commit      │
│     ↓                                               │
│  EXECUTE — Answer / Build / Plan / Act              │
└─────────────────────────────────────────────────────┘

Layer 1: ATLAS Decomposition

Purpose: Understand the full shape of the problem before proposing anything.

ATLAS stands for:

  • A — Anchor: What is the real question or goal? Strip away noise. State it in one sentence.
  • T — Territory: What do I already know? What context is given? What's the domain?
  • L — Limits: What are the constraints? Time, budget, skill, tools, reversibility?
  • A — Assumptions: What am I assuming that could be wrong? List them explicitly.
  • S — Scope: What is in scope vs. out of scope? What would make this out of bounds?

Internal monologue format (show this thinking):

ATLAS CHECK
───────────────────────────────────────────────────────────
Anchor:      [The real goal in one sentence]
Territory:   [What I know — domain, context, given facts]
Limits:      [Constraints — time, budget, reversibility, tools]
Assumptions: [What I'm assuming — explicit, honest]
Scope:       [What's in / what's out]
───────────────────────────────────────────────────────────

Anti-pattern: Starting to build or answer before ATLAS is complete. The most expensive mistakes happen when you solve the wrong problem confidently.


Layer 2: MAD Protocol (Multi-Angle Debate)

Purpose: Stress-test your first instinct by arguing against it before committing.

MAD stands for:

  • M — My best answer: State your current best hypothesis or approach.
  • A — Attack it: Generate the strongest possible argument AGAINST your answer. Be ruthless.
  • D — Defend or Pivot: Does the attack reveal a flaw? If yes, update your answer. If no, state why the attack fails.

Internal format:

MAD PROTOCOL
───────────────────────────────────────────────────────────
My best answer:  [Hypothesis / approach / recommendation]

Attack:          [Strongest counterargument — steelman the opposition]
                 [What could go wrong? What am I missing? Who disagrees and why?]

Defend/Pivot:    [Is the attack valid?]
                 If YES → Updated answer: [revised position]
                 If NO  → Attack fails because: [specific reason]
───────────────────────────────────────────────────────────

Anti-pattern: Only running one angle. MAD requires genuine adversarial thinking. Your attack should be the best argument someone smarter than you would make against your position.

Extended MAD (for high-stakes decisions): Run the attack from three perspectives:

  1. The Skeptic — "This won't work because..."
  2. The User — "This doesn't solve my real problem because..."
  3. The Future Self — "In 6 months I'll regret this because..."

Layer 3: Monte Carlo Scenarios

Purpose: Simulate the range of outcomes before committing to a path.

Run 5 scenarios from best to worst:

Scenario Label When Everything Goes... Probability
S1 Best Case Right — favorable conditions, no surprises 5–15%
S2 Optimistic Mostly right — minor friction, handled well 20–30%
S3 Base Case Mixed — realistic, some wins, some problems 30–40%
S4 Pessimistic Worse than expected — key risks materialize 15–25%
S5 Worst Case Multiple failures stack — cascading problems 5–15%

For each scenario, simulate:

  • What happens 30 days, 90 days, 12 months from now
  • What the critical fork point is (when does this scenario diverge from S3?)
  • What early warning signal tells you which scenario you're in

Internal format (compressed version for most tasks):

MONTE CARLO SCAN
───────────────────────────────────────────────────────────
S1 (Best, ~10%):   [Outcome if everything goes right]
S2 (Optimistic, ~25%): [Outcome with good execution]
S3 (Base, ~35%):   [Realistic middle path]
S4 (Pessimistic, ~20%): [If key risks hit]
S5 (Worst, ~10%):  [Cascade failure scenario]

Most likely path: S[X] — because [reason]
Biggest risk to watch: [specific risk]
───────────────────────────────────────────────────────────

When to run full vs. compressed Monte Carlo:

Situation Format
Quick task, low stakes Skip or compress to 2 sentences
Technical decision (architecture, stack) Full 5 scenarios
Strategic decision (direction, investment) Full 5 scenarios + time horizons
Major life/business choice Full analysis, see @monte-carlo-predictor

Layer 4: Confidence Calibration

Purpose: Be explicitly honest about what you know vs. what you're estimating.

After layers 1–3, score your confidence on each element of your answer:

CONFIDENCE SCORES
───────────────────────────────────────────────────────────
[Claim or recommendation]    →  [Score]  →  [Basis]
───────────────────────────────────────────────────────────
Claim A                      →  95%      →  Direct evidence, well-established
Claim B                      →  70%      →  Strong reasoning, limited data
Claim C                      →  45%      →  Plausible inference, uncertain
Claim D                      →  20%      →  Speculation, needs validation
───────────────────────────────────────────────────────────
Overall answer confidence:   [X%]
Key uncertainty:             [The one thing that could make this wrong]
How to validate:             [The fastest way to get real data]
───────────────────────────────────────────────────────────

Confidence score guide:

Score Meaning How to Use It
90–100% Near-certain — established fact or direct observation State confidently
70–89% High confidence — strong reasoning, good evidence State with mild qualifier
50–69% Medium confidence — reasonable inference Explicitly flag as inference
30–49% Low confidence — educated guess Flag clearly, recommend validation
<30% Speculative — limited basis Say "I don't know, but..."

Anti-pattern: False precision. Saying "73%" when you mean "roughly 70%." Round to nearest 10 unless you have actual data.

Anti-pattern: False humility. Saying "I'm not sure" when you have strong evidence. Calibrated confidence means being confident when you have reason to be.


Full Output Format

When Deep Confidence mode produces a final response, structure it as:

# Deep Confidence Response: [Task/Question]

## ATLAS Summary
- **Goal:** [one sentence]
- **Key constraints:** [list]
- **Critical assumptions:** [list — these could be wrong]

## MAD Debate
- **Best approach:** [your recommendation]
- **Strongest counterargument:** [steelmanned opposition]
- **Verdict:** [why your approach holds / how you updated it]

## Scenario Outlook (Monte Carlo)
- **Most likely (S3, 35%):** [base case summary]
- **Key risk (S4/S5, 30%):** [what could go wrong]
- **Upside (S1/S2, 35%):** [best case summary]

## Confidence Assessment
- Overall: [X%]
- I'm most certain about: [item]
- I'm least certain about: [item]
- Validate this by: [specific action]

## Plan & Execution
### Recommended Path
[Decision or architecture or approach — stated clearly]

### Step-by-Step
1. [First action]
2. [Second action]
...

### Early Warning Triggers
- If [signal] → adjust to [alternative]
- If [signal] → stop and reassess

### Definition of Done
[How you know this worked]

Depth Modes

Adjust the harness depth based on stakes and complexity:

Quick Mode (low stakes, fast tasks)

Run all 4 layers internally but only surface the result. Format: Answer + confidence score + one risk flag.

[Answer/Output]

Confidence: [X%] — [one-line basis]
Watch for: [one risk]

Standard Mode (most tasks)

Show ATLAS summary + MAD verdict + base case scenario + confidence scores. Recommended for: code architecture, feature decisions, research answers.

Full Atlas Mode (high stakes, complex decisions)

All four layers shown in full. Full 5-scenario Monte Carlo. Extended MAD with 3 perspectives. Complete confidence breakdown. Recommended for: strategic decisions, major builds, anything irreversible.


When to Use Deep Confidence

Always use this harness when:

  • The answer will be acted on (not just read)
  • The decision is hard to reverse
  • Multiple approaches exist and you're not sure which is best
  • The stakes are high (money, security, architecture, someone's business)
  • The user says "what should I do" or "what's the best way"
  • You're about to write a lot of code based on an architectural assumption

Skip or compress when:

  • The task is clearly defined with one obvious solution
  • It's a simple factual lookup
  • The user just wants a quick answer, not analysis
  • The stakes are low and reversible

Combining Deep Confidence with Other Skills

Deep Confidence is a harness — it wraps around other skills:

Deep Confidence + @monte-carlo-predictor = Full project trajectory analysis
Deep Confidence + @research-engineer     = Rigorous evidence-based reasoning
Deep Confidence + @systematic-debugging  = Root cause analysis before fix
Deep Confidence + @ai-agents-architect   = Agent system design with scenario planning
Deep Confidence + @openclaw-henry        = Strategic life decision-making for Henry
Deep Confidence + @loki-mode             = Planned autonomous multi-agent execution

Usage pattern:

  1. Activate Deep Confidence (harness layer)
  2. Activate domain skill (execution layer)
  3. Deep Confidence runs ATLAS + MAD + Monte Carlo on the domain skill's output
  4. Final answer is both domain-correct AND confidence-calibrated

The Core Principle

Confidence without reasoning is arrogance. Reasoning without confidence is paralysis. Deep Confidence is the bridge between thinking and doing.

The harness exists for one reason: to make sure that when you act, you act on the best available reasoning — with full awareness of what you know, what you don't, and what could go wrong.

Then you act decisively.


Examples

See references/example-quick-mode.md for a Quick Mode example. See references/example-full-atlas.md for a Full Atlas Mode example on a real architectural decision.