Essay · August 8, 2026

The essentials first

The ablation principle

I heard Boris Cherny, the creator of Claude Code, describe how his team treats each new model generation: they delete the product's entire system prompt, then rebuild it line by line to see what is still genuinely useful. With Opus 5, more than 80% never came back. Most of what remained had been compensating for flaws a smarter model simply no longer has. His advice to users of his own tool, as I understood it: every six months, delete your instruction files, your skills, your automations, and watch what the model does without them.

The idea is simple, and it becomes uncomfortable the moment you take it seriously. The instructions we give a model are often crutches fitted for a previous generation. When the model changes, the crutch becomes noise, sometimes a hindrance. We keep explaining how to walk to something that already runs.

Yet a few months earlier, I had built a complete methodology for Ezkey: documentary levels, templates, skills, naming rules. I had presented it at the moment I put on the methodologist's hat, taking care to say I wanted to avoid the trap of yesterday's grand methodologies: light, optional, deliberately chosen. The ablation principle, turned against me, asked a question I did not want to hear.

The article I never published

My first response was to write an article. The one you are reading bears its name; it is not the same text.

The original version did everything a self-critical essay knows how to do to avoid self-criticism. It summoned the mirror diligently, announced a severe examination, then converted every admission into a virtue one paragraph later. My slowness to adopt new practices became a strategy. My failure to enforce my own rules became a saving flexibility. Above all, in a text entirely devoted to the principle of "delete, then watch," I deleted nothing. The moments where I thought I was applying ablation were abstentions: not adding a constraint, not opening a file. Not adding is not removing.

I submitted that text for critique, and the critique came from the AI itself. I have written before that working with a model is a very sophisticated way of talking to yourself, and that the echo-chamber risk is real. This time, the echo did not come back. The critique named, in black and white, what a small voice had been telling me for a while without my listening: not one number, not one scene, not one date. A verdict announced and never delivered. A trial ending in a quick acquittal. A preamble renewable forever.

Once the defense fell, the finding was simple. I had presented as light a methodology that counted dozens of framing documents, some twenty templates, some fifteen skills. I had preached letting models evolve freely, and I had frozen entire workflows around them that I only half practiced myself. My best guardrails, the ones warning against ceremony, slept deep in the corpus in decision records no cold-starting agent had any reason to read. A corpus thick enough to lose its own best rules is a corpus too thick. I had not been honest with myself, and the failed article was the written proof.

Delete, then watch

So I did it, literally. A short plan, one session, one commit.

Gone: the thirty-two dated decision records, for which Git history remains the archive; twenty of the twenty-one method documents, from the session-start guide to the per-workflow state sequences; the fourteen skills that orchestrated the method; sixteen of the twenty templates; the rich HTML view of nearly 1,500 lines. Nothing was moved to an archive folder. An archive folder is a way of deleting without watching.

What was judged genuinely load-bearing was not relocated; it was condensed into a single document of about 230 lines. The lane taxonomy, the sequence diagrams, the entry and exit checklists were folded nowhere. That is exactly the class of structure a good Plan-mode session already provides, better than I do, and improving with every release.

The methodology's public site follows the same movement: it too is being reworked and drastically simplified, a direct consequence of the ablation. The guided tours built for the old corpus no longer have material to walk through; what replaces them fits, like everything else, in very few pages. Deletion came first and republication follows closely behind — in the reverse order, the ablation would have turned into a reorganization, and a reorganization is yet another way of deleting nothing.

What survived

The ablation was not a global verdict; it was a sorting. It fits in one sentence: Plan mode first, the methodology as support, reduced to what Plan mode does not carry.

The Plan mode of the large providers improves with every release, fed by a feedback scale no project this size can fund. Competing with it through homegrown workflows means cutting yourself off from free continuous improvement. What it does not carry, however, is continuity: a session started cold has no way of knowing what a session three months ago decided, unless the corpus carries it.

Three things survived. Documentary levels, which distinguish a product vision from a concrete idea and from a bounded design ready to implement. Bidirectional discoverability, which lets you climb from a line of code back to the intent that motivated it, and back down again. And the project's values compass: the eighty-twenty rule, essential complexity distinguished from accidental, rigor proportional to the stakes. A compass does not show the exact path; it gives a direction when no precise rule applies.

To these, add three rules fished back from the depths, precisely the ones the corpus had buried: distinguish routine hygiene from a real program before materializing artifacts; treat a working plan as ephemeral scaffolding unless it still carries something the canon should not flatten; hold an uncertainty that has been addressed and recorded as closed, with no reopening without a new signal. These rules already existed. Everything around them had to be deleted for them to become readable again.

That leaves memory, since providers are beginning to consolidate their agents' memory across sessions. The distinction I maintain: that memory learns how I work; the corpus retains what I am building, and why. An intent structured in levels, linked in both directions, is not a set of work preferences. I expect the two layers to end up complementing each other, and the day the first absorbs the second, it will pass through the same sieve as everything else.

The test of time

The observation protocol is in place, the very one the failed article lacked: work normally, note what a cold agent visibly misses of what the deleted corpus used to carry, and reinstate only what is genuinely missing. A line in the single document, never a new file.

I do not know what the observation will show. Some pieces may turn out to be missed, and come back. A more mature open-source methodology may someday cover what I am trying to do better than mine does, and the right reflex will then be to move toward it rather than cling to my own: ablation applied, that time, to an entire methodology. Frontier models will keep evolving faster than my habits, and the exercise will have to be repeated. I will get some things right and some things wrong; this is a project in motion, and the future will judge.

Français : Version française de cet article

See also: Critical Thinking and Idea Alignment

See also: From knowing to living

See also: The Ezkey methodology