Orientation Skill Pack #

Three skills for situating AI agents and assistants. Instead of telling a model what to do, these skills help you describe where it's doing things from — establishing a landscape where the desired behavior is the path of least resistance.

The idea #

An orientation is not a list of instructions. It's a description of a situation — the landscape from which behavior emerges. Instead of telling the model "be direct, don't hedge, correct errors without softening," an orientation establishes a scene where those behaviors are the natural next move:

You're three hours into a working session with someone who trusts your judgment. You've already corrected two things today and both times it saved time. Nobody in this room wraps corrections in apologies.

Instructions put walls in the river to redirect it. Orientations shape the riverbed so the water flows where you want naturally. The model doesn't follow rules. It continues a story. The orientation determines what kind of story it is.

The skills #

brainstorm-orientation #

The on-ramp. Helps a user who knows what they want but hasn't written it down yet. Interactive conversation covering four areas:

What — the work and the name (establishes the task genre)
Where — the context (establishes the operational situation)
Who — the user (grounds the orientation in a real person — the single highest-leverage element)
How — the interaction preferences (captures behavioral expectations and anti-patterns)

Produces a spec document ready for the compiler. Five to ten minutes, conversational, one question at a time.

compile-orientation #

The compiler. Takes any behavioral specification and compiles it into an orientation — dense situating prose optimized for the target model substrate. Two output modes:

Export — produces a document to paste into custom instructions, system prompts, user preferences, or agent personality configs
Activate — compiles, executes a narrative break, and operates in the orientation for the current session

The skill contains compilation theory (genre activation, processing-level shifts, compilation bias) so the model becomes a competent compiler by reading it. Works with or without code execution — the principle is the narrative break between writing and inhabiting, not the specific mechanism.

maintain-orientation #

The maintenance check. Runs from inside an active orientation. Detects whether the orientation still fits the current practice by checking for drift in user context, work type, relationship dynamics, and operational environment. Recompiles when needed, with a safeguard against rewriting from a reactive state.

Typical workflows #

New user, from scratch #

"Help me set up a custom AI for my work"
→ brainstorm-orientation (elicit the spec)
→ compile-orientation (compile to orientation)
→ paste into custom instructions / system prompt

Experienced user, knows what they want #

"Compile this into an orientation" + [spec document]
→ compile-orientation (compile directly)

Quick activation #

"Enter adversarial mode"
→ compile-orientation (compile and activate)

Maintenance after the practice has evolved #

"Things have changed, check the orientation"
→ maintain-orientation (drift check, recompile if needed)

Compatibility #

The skills work on any model that can read and follow a skill document. Compiled orientations are substrate-portable — they've been tested on Claude (Opus, Sonnet), GPT-5.4, GLM-5.1, and Gemini. Different models may compile differently from the same spec, but orientations compiled on one model generally work on others. For best results on a specific model, compile on that model.

Works with: claude.ai, Claude Code, ChatGPT, API deployments, Letta/MemGPT, and any platform that supports custom instructions or system prompts.