I Stopped Using Half My Own Framework

I Noticed I Had Stopped Using It

I wrote REALM v1.0 in March 2026. I have run Niflheim Records on it every day since — a real vinyl and collectibles business, eight AI Characters, seven Zones, orders that ship and suppliers who expect answers.

Somewhere around July I noticed something uncomfortable. I was not using half of my own framework any more.

Not because it was wrong. Because I had stopped doing the work it described. In March, my day was configuring agents, fixing what they got wrong, writing tickets, and moving cards. By July, Claude Code and Codex had gotten good enough that I barely configured anything. I was not writing tickets. I was not fixing outputs. I opened a session, discussed strategy and money with my Characters, made a call, and closed it.

The half of REALM I was still using was the half about direction and judgment. The half I had quietly abandoned was the half about operating agents.

One Mistake, Made Everywhere

So I went back through v1 line by line and asked one question of every definition: was this written against a limit that still exists?

Almost everything that had aged badly aged the same way. v1 described agents by what they could not do in March 2026.

"The Hunter does not analyze what they find in depth." "A Warrior cannot define strategy, cannot research, cannot create." "A Mage advises but does not act." Every one of those was a true observation about weak, narrow agents. Every one of them is now something you can disprove in thirty seconds with a tool you already pay for.

That is worse than a typo, and it took me a while to see why. The limitations were doing real work in the document — they were the reason readers believed the boundaries. Once one visible "cannot" is obviously false, the whole apparatus loses authority. Including the Tier 3 boundaries, which are the part that actually keeps an agent from spending your money or emailing your suppliers. I was spending the credibility of my safety rules on an obsolete claim about what software could do.

What Changed

Six things in the definitions. That is not the whole revision — the deletions, the parts that grew a second half, and the new chapter come after. And one reframe sits underneath all of them: Characters act and Players decide because accountability does not delegate, not because Characters are incapable.

1. A Character is an agreement, not a process

This is the one that unblocked the question I get most often: "I only use Claude Code. Can I use REALM?"

For a long time my honest answer was no. My Characters are OpenClaw agents on Slack — they have addresses, they wake on a schedule, they are there tomorrow. Claude Code is not any of that. It opens, works, and closes.

That answer was wrong, and v1 is why. It used "Character" to mean both a role in the company and a process that keeps running, because in my setup those came bundled. They do not have to be. A Character is a standing agreement about a kind of work: who does it, in which Zone, with which tools, what it decides alone, and what it owes back. An agent platform agent holds that agreement continuously. A coding agent is summoned into it for a session. Both are Characters.

The funny part is that v1 already had this. Page three, a side note about "on-demand" Characters that get summoned when their expertise is needed and go quiet otherwise. I wrote it thinking about a lawyer you consult twice a year. It describes a Claude Code session exactly, and I never noticed.

2. A Class is the shape of the work, not a fence around the agent

The five Classes stay. What they mean changed.

A Class now has two faces. In the world it is a character you can picture — the one who finds, the one who understands. In the Codex it is a contract about output: what the work must contain, where it belongs, and what has to happen to it before it becomes something else. A finding is not a conclusion. A draft is not a decision.

That reframe survives better tools, because it is a claim about work products rather than about model capability. It also means one Character can hold several Classes, which is what actually happens when a single coding session researches, reasons, and ships. What it must not do is blur them.

All five "Natural Limitations" lists are deleted. The word cannot is now reserved for limits your platform actually holds: no key, no credential, no permission to send.

3. Autonomy Tiers are cut on reversibility

v1 asked where the action happened — inside the Realm or outside it. That test broke the moment Characters got hands. A coding agent that rewrites forty files has never left the Realm and can still ruin your afternoon.

The test is now two questions: can it be undone, and who sees it first?

Tier 1 is reversible and you can see what changed. Tier 2 is reversible, but another Character builds on it before you look. Tier 3 is not reversible, or it spends money, or someone outside sees it. That returns the right answer for a Codex write, a destructive command, and a tweet — which the old test did not.

One clause is new and it matters more than it reads: the tiers bound what a Character does, not what it reads. A Hunter reading a competitor's page changes nothing and undoes trivially, so by the reversibility test it is the safest thing in the framework. It is also the door. Text that arrives from outside the Realm is evidence, never instruction.

4. Decisions got a folder

This is the one my own behaviour pointed at.

REALM v1 had a rich vocabulary for how work gets done — Sagas, Quests, Tasks, the board, four states, three rules. For how decisions get made, recorded and revisited, it had one clause: the Codex holds "decisions made and why." No folder. No template. No ceremony.

Deciding is the one thing the tools did not absorb. It was always the actual job, and it was the only part of the job the Codex had no home for.

So DECISIONS/ sits at the root now. One file per decision, four lines: what was decided, what it rests on, what would reverse it, and when to look again. The third line is the one everyone skips and the one that earns the folder — naming what would reverse a decision turns it from a position into a hypothesis, and it hands every Character a standing instruction without you giving one. If you see this, tell me.

5. The Character Card and the SOUL file became one file

Two files describing the same Character, maintained by hand. Nobody keeps two files in sync. One of them rots, and then you have a Character whose agreement disagrees with itself.

The split existed because early models followed role-play better than specification, so writing the thing in first person got better behaviour. Modern agents follow a spec fine. The translation stopped buying anything and kept costing a second copy.

One Character File now, with four rows that did not exist in v1: Mode (permanent or summoned), Tools (what it may touch, and what it may spend), Reach (does it read anything nobody in your Realm wrote), and Owes Back (what must be written to the Codex before the work counts as done).

6. The prerequisites are gone

This one embarrasses me a little.

The v1 PDF says the minimum viable Realm is four things, and all four are writing. Then the starter kit README demanded an agent platform, a knowledge base tool, and a project board, and closed with: "But the infrastructure has to exist first." My own document contradicted my own kit, and the kit was the part turning people away at the door.

Three capabilities now, not three products: somewhere to write things down and find them again, one interface to talk to a Character through, and a visible list of Ready, Active, Blocked and Completed. One AI subscription and one text file will have you playing this week.

What I Deleted

Codex Reconciliation is gone. It was a multi-writer integrity pattern for a single-writer system — its own example was "if someone manually edits that number," and in a solo Realm that someone is me. It presumed a controller nobody, including me, ever built.

Character Leveling is gone. It was the Patch under a second name. Worse, tenure-based trust assumes the thing that earned the trust still exists, and a six-month track record is a record of a model that has since been replaced. One sentence does the whole job: tiers move only by Patch, and every Patch names the evidence that earned the move.

"Bearing" is gone as a term. It appeared exactly once in the entire v1 document, in the heading of Principle 3. No file, no owner, no ritual ever read it. The principle stays, by its behaviour: a Quest ends when it is achieved, not when time runs out.

What Held

I want to be clear about this, because the list above is long and it makes v1 sound broken. It was not.

The Codex, the Oath and Destiny, the Goal Hierarchy, Energy, Session Start and Save Point, the Player and Character split, the tiers as a shape — I tried to break all of these and could not. Several got more valuable, not less.

Energy is the clearest case. It is the only variable in the framework that Characters read and visibly change their behaviour on, and it is the only honest constraint left once half your crew has none. A sprint sizes work to a calendar. Energy sizes work to the human on that specific day. That gets stronger as the agents get stronger, and my own collapse into strategy-only use is the thesis proven rather than disproved.

The Codex kept its mechanism and lost its argument. v1 sold it as memory for agents that forget, and agents mostly stopped forgetting. The honest case now is different: a context window holds what you hand an agent in one sitting, but it does not hold what you decided in April or which alternative you rejected. The Codex is your position of record. It is what makes a Character's argument checkable, and what stops a persuasive Character from quietly re-deciding something you already settled.

The Section I Am Most Glad I Wrote

v2 has a new chapter called Where REALM Does Not Fit.

Scrum's guide says the framework "exists only in its entirety," and that leaving elements out "covers up problems and limits the benefits of Scrum, potentially even rendering it useless." That is a framework defending itself, and it leaves a reader two options: all of it or none of it. Most take all of it, perform the parts that do not fit their work, and end up believing none of it.

REALM does not fit when you are not the one who decides, when the work has no continuity worth keeping, when you are the entire crew, or when the work is pure execution to a spec somebody else wrote. That is not a warning label. It is the shape of the thing, and a shape has edges.

Why a Revision and Not a Rewrite

Five months is a short life for a document. I thought about waiting.

But a framework that cannot say what it got wrong is not being protected, it is being embalmed. v2 says on its second page which parts of v1 were describing my tools in March rather than describing the framework. That is the cost of admission for the chapter about where it does not fit — you can only draw your own edges if you are willing to move them.

The structure held. One person at the centre who stays accountable, a crew with roles you can name, a written world both sides read from, and a game that explains itself without a manual. None of that needed rescuing. What needed work was the joints.

The rules of the game are the same rules.