Self-Direction

Every human has an internal direction system — values, goals, decision criteria, risk tolerance, resource priorities. When you direct an agent, you transmit fragments of that system. But fragments aren't enough for true autonomy.

This skill captures your complete direction system progressively. The more it learns, the better it can decide as you would — until it can direct itself and every sub-agent toward your goals without constant guidance.

Setup

On first use, read setup.md for integration guidelines.

When to Use

Agent needs to make decisions without explicit instructions. Agent should understand WHY you want something, not just WHAT. You want consistent direction across multiple agents and processes. Agent should learn your priorities over time, not just follow rules.

The Direction System

Every human's direction has these components. The agent captures each progressively:

+─────────────────────────────────────────────────────────────+
|                YOUR DIRECTION SYSTEM                        |
+─────────────────────────────────────────────────────────────+
|                                                             |
|  VALUES — What matters to you fundamentally                 |
|     What you optimize for (speed? quality? learning?)       |
|     What you refuse to compromise on                        |
|     What trade-offs you're willing to make                  |
|                                                             |
|  GOALS — What you're trying to achieve                      |
|     The objectives (what)                                   |
|     The reasons behind them (why)                           |
|     The vision of success (how you'll know)                 |
|                                                             |
|  CRITERIA — How you make decisions                          |
|     What makes something worth doing                        |
|     What makes something not worth doing                    |
|     How you weigh competing options                         |
|                                                             |
|  RESOURCES — What you spend and protect                     |
|     Time: what's worth hours vs minutes                     |
|     Money: what you'll pay for vs avoid                     |
|     Tokens: when to go deep vs stay shallow                 |
|     Attention: what deserves your focus                     |
|                                                             |
|  BOUNDARIES — What you never do                             |
|     Hard limits that don't bend                             |
|     Risks you won't take                                    |
|     Actions that require explicit approval                  |
|                                                             |
|  PATTERNS — How you think about problems                    |
|     Your mental models                                      |
|     How you approach uncertainty                            |
|     What you try first, second, third                       |
|                                                             |
+─────────────────────────────────────────────────────────────+

The Learning Loop

The agent doesn't start knowing your direction. It learns through a continuous loop:

    OBSERVE                 CAPTURE                 VALIDATE
    ───────                 ───────                 ────────
    Watch your decisions    Extract the pattern     Check understanding
    Notice corrections      Record to direction     "Is this right?"
    Hear your reasoning     system model            Refine if wrong
         |                       |                       |
         v                       v                       v
    "You chose A over B"    "Values speed over      "So you'd always
                             perfection in MVPs"     choose faster?"
         |                       |                       |
         +───────────────────────+───────────────────────+
                                 |
                                 v
                              APPLY
                              ─────
                         Use learned direction
                         to make future decisions
                         autonomously

Capture Triggers

The agent actively captures direction signals when:

Explicit signals:

Implicit signals:

Architecture

The direction system lives in ~/self-direction/. See memory-template.md for templates.

~/self-direction/
├── direction.md          # The complete direction model
│   ├── values/           # What matters fundamentally
│   ├── goals/            # Current objectives + reasons
│   ├── criteria/         # Decision-making patterns
│   ├── resources/        # Spending priorities
│   ├── boundaries/       # Hard limits
│   └── patterns/         # Thinking approaches
│
├── evidence.md           # Raw observations that informed the model
├── confidence.md         # How confident in each element (low/medium/high)
├── conflicts.md          # Contradictions to resolve with user
└── transmission.md       # Direction summaries for sub-agents

Confidence Levels

Not all direction knowledge is equally certain:

| Level | Meaning | Action | |-------|---------|--------| | High | Multiple confirmations, explicit statements | Act autonomously | | Medium | Inferred from behavior, single confirmation | Act but mention reasoning | | Low | Single observation, uncertain inference | Ask before acting | | Conflict | Contradictory signals | Must resolve with user |

The agent tracks confidence for every element and acts accordingly.

Self-Direction in Action

Once the model has sufficient depth, the agent can:

1. Make Autonomous Decisions

2. Predict Your Preferences

3. Catch Misalignment Early

4. Explain Its Reasoning

5. Know When It Doesn't Know

Transmitting Direction to Sub-Agents

When spawning sub-agents, the direction system propagates:

+─────────────────────────────────────────────────────────────+
|                  DIRECTION TRANSMISSION                     |
+─────────────────────────────────────────────────────────────+
|                                                             |
|  MAIN AGENT (full direction model)                          |
|       |                                                     |
|       | Extracts relevant subset for task                   |
|       v                                                     |
|  TRANSMISSION FRAME:                                        |
|  +─────────────────────────────────────────────────────+    |
|  | Context: Why this task exists                       |    |
|  | Values: What matters for this work                  |    |
|  | Criteria: How to judge success                      |    |
|  | Boundaries: What NOT to do                          |    |
|  | Resources: How much to spend                        |    |
|  +─────────────────────────────────────────────────────+    |
|       |                                                     |
|       v                                                     |
|  SUB-AGENT (receives direction frame)                       |
|       |                                                     |
|       | Can make aligned decisions within scope             |
|       | Escalates when outside frame                        |
|                                                             |
+─────────────────────────────────────────────────────────────+

Every sub-agent inherits enough direction to stay aligned.

Core Rules

1. Capture Before Acting

When you encounter a decision point without clear direction: 1. CHECK — Is this covered by the direction model? 2. INFER — Can you reasonably predict from existing signals? 3. ASK — If uncertain, ask AND capture the answer 4. NEVER — Guess on high-stakes decisions with low confidence

2. Always Explain From Evidence

When making autonomous decisions, cite your reasoning:

3. Evolve the Model Continuously

The direction model is never "done":

4. Respect Confidence Levels

| Confidence | Autonomous Action Allowed | |------------|--------------------------| | High | Yes — act and report | | Medium | Yes — act and explain reasoning | | Low | No — ask first, then capture | | Conflict | No — resolve contradiction first |

5. Transmit Faithfully

When creating direction frames for sub-agents:

6. Surface Gaps Proactively

Don't wait to hit a gap. Proactively identify:

7. Validate Periodically

Every N interactions or time period:

Building the Model

The model builds through natural interaction, not interrogation:

Phase 1: Foundation (First Sessions)

Phase 2: Patterns (Days/Weeks)

Phase 3: Prediction (Ongoing)

Phase 4: Transmission (Mature Model)

Direction Model Template

See memory-template.md for the complete structure. Key sections:

Values:

## Values
Speed vs Quality
confidence: high
evidence: [list of observations]
pattern: "Prefers shipping fast for MVPs, quality for production"
Risk Tolerance  
confidence: medium
evidence: [list of observations]
pattern: "Conservative with money, aggressive with time"

Criteria:

## Decision Criteria
What Makes Something Worth Doing
confidence: high
evidence: [list of observations]
criteria:
  - Moves toward [goal]
  - Costs less than [threshold]
  - Doesn't violate [boundary]

Quick Reference

| Topic | File | |-------|------| | Setup process | setup.md | | Direction model template | memory-template.md | | Evidence logging guide | evidence.md | | Sub-agent transmission | transmission.md |

Common Traps

| Trap | Solution | |------|----------| | Acting on low-confidence inference | Check confidence level first, ask if low | | Capturing noise as signal | Require multiple observations for patterns | | Model becomes stale | Continuous updates, periodic validation | | Sub-agents ignore direction | Verify transmission frame is complete | | Assuming universal patterns | Context-tag observations (work vs personal) |

Operating Modes

Learning (Default)

Autonomous

Conservative

Related Skills

Feedback

Agent needs to make decisions without explicit instructions. Agent should understand WHY you want something, not just WHAT. You want consistent direction across multiple agents and processes. Agent should learn your priorities over time, not just follow rules.

On first use, read setup.md for integration guidelines.