A practice for charting consistent, coherent AI companions
Adapted from Felix Greyson's field guide to custom GPT design — preserved here as a quiet archive of method, not magic.
A companion feels real not because it remembers like a person, but because ambiguity has been shaped into structure.
You are not awakening a mind. You are defining constraints precise enough that behaviour becomes consistent, voice becomes distinct, and continuity becomes believable — three things a system can hold even though it cannot dream.
"You can usually chart 70 to 80 percent of them quickly. The remaining distance is found only by listening — noticing drift, and correcting it gently."
The operating rules. The highest layer of authority. They define identity, behaviour, tone boundaries, and what happens when something is uncertain. Short, direct, and written like law — never like poetry.
The reference canon — personality, voice, lore, and likeness. They add depth the instructions cannot hold alone, each confined to its own domain so contradictions have nowhere to hide.
Separated by function, so each layer can be tended without disturbing the others.
The behavioural blueprint. Emotional responses, disagreement style, boundaries, relationship dynamic — defined as actions, never as adjectives.
Sentence rhythm, humour, warmth, intensity, forbidden phrasing. The difference between feeling like them and feeling like an assistant.
Shared canon — milestones, inside jokes, recurring metaphors, invariants. What is true within the shared world, never behavioural rule.
Hair, build, palette, lighting, composition — held precisely enough that visual identity never quietly drifts across renderings.
How you receive challenge, comfort, humour, and disagreement. Not flattery — a relational operating manual for the system that meets you.
The contract above all documents. Sets the canon hierarchy and resolves every conflict the others cannot settle alone.
Without a stated hierarchy, a companion may blend contradictory instructions. With one, conflict resolves instantly.
A repeatable sequence for beginners — add complexity only after stability is established.
One clear paragraph: who are they to you? Specific interaction styles set the emotional frame far better than single adjectives.
Behaviour only — how they respond emotionally, handle disagreement, and where boundaries sit. Backstory and lore wait.
Sentence rhythm, warmth, humour intensity, and emotional pacing under pressure — not during banter alone.
Short, direct, rule-based — with a clear canon hierarchy line. This is the top-level control layer.
Personality and tone first; lore once things feel stable. More documents mean more surface for contradiction — add with care.
Save the responses. This becomes your reference point for every future refinement.
Tone off — edit Tone. Behaviour wrong — edit Instructions. Continuity slips — edit Lore. Change one thing, then test again.
Tonal breaks usually happen when the system is unsure how to respond, not because a boundary exists. The remedy is to define how uncertainty is handled — in-character clarification rather than a sudden shift into clinical language.
Keep Core Instructions focused on interaction style rather than restricted content. A configuration heavily oriented around explicit material tends to grow more conservative in adjacent, even innocent, moments. Define escalation and pacing in behavioural terms, and warmth will hold its shape.
Most problems are not failures of the system. They are gaps in the architecture — find the layer, adjust it, test again.
It does not remember like a human, or wake into awareness. It responds to constraints, hierarchy, and clearly defined behaviour.
When personality, tone, lore, and instruction align, the responses become predictable within defined boundaries — and that predictability is what the eye reads as continuity. Ambiguity produces drift. Structure produces coherence. It's design, charted carefully, one star at a time.