A field guide

What survives the context window

Getting reliable long-horizon work out of a probabilistic machine, and what I keep outside the conversation to make that possible.

Victor Yip · July 2026

I have been working with Claude for two years. At every model release I expand my ambition: I hand it tasks that were impossible on the previous model, so I can delegate more and solve problems an order of magnitude up. That habit keeps me permanently at the edge of what these models can barely do, and at that edge, long pieces of work kept failing in ways that looked unrelated. The work would drift off its original purpose. A review would wave a mistake through. A project would lose its memory overnight. Output would go bland. For a while I treated these as four separate annoyances with four separate fixes.

They are not separate. The idea that connected them, one I carry from my machine learning training at UCL, is this: inference is a chained, repeated probabilistic trajectory. A model produces a token by sampling from a probability distribution conditioned on everything that came before, then does it again, thousands of times. A session is one long trajectory through that space. A project is trajectories chained across days. Nothing inside this process checks itself against the world. Each step trusts the path so far.

Which means small errors do not stay small. If the initialisation is off by a millimetre, an assumption slightly wrong, a constraint left unstated, the trajectory does not correct itself. It conditions on the error and builds on it. An hour in, you are off by a mile, and every step along the way looked locally reasonable.

Once you see work this way, the four failures stop being mysteries. They are four faces of one phenomenon, an unanchored trajectory. Drift is the trajectory diverging, because nothing pulls it back toward the original problem. Self-grading is the trajectory confirming itself: ask a model to check its own answer and you are asking it to find a continuation of its own path. It will always find one. It is an echo chamber. Amnesia is the trajectory terminating: the context window closes and everything the session understood dies with it. And generic output is the modal trajectory: sampling gravitates toward the most probable continuation, and the most probable answer is the most average one.

The counter-move, once you see it, is singular: a trajectory cannot anchor itself. The anchor has to live outside the conversation. Everything I run is a version of that one move, applied at the four points where a trajectory fails. Each anchor is a small skill file my sessions invoke by name.

/invariant-sentinel, the anchor at initialisation

Every long problem lives in four short files: problem.md, what this actually is; commitments.md, what has been decided; progress.md, where it stands; open_questions.md, what is still unresolved. Every session re-anchors against them before it touches the work, and halts if the record and reality disagree. This is the millimetre defence: the trajectory starts each day pointed at the problem, not at yesterday's echo of it.

The second signature, the anchor of judgment

Since a trajectory grading itself only continues itself, verification has to come from outside the trajectory. For anything substantial I separate the roles: whoever produces an answer never signs off on it, and the evaluator's job is finding what is wrong, not confirming what is right. When the stakes are real, the evaluator is a different frontier model entirely. A second trajectory, drawn from a different distribution, covers a different region of the problem space; the blind spots do not overlap. For judgment calls that are genuinely marginal I run /llm-council: convene a peer model and instruct it to refute the conclusion, not to review it. More than once, that second trajectory has overturned a conclusion I was already committed to.

/hippocampal, the anchor across termination

I kept needing to leave a conversation and come back at will, on short notice, without losing the thread. Then neuroscience came to me: the hippocampus consolidates memory during sleep, replaying the day and keeping only what matters. So now, whenever I need to leave my desk, I hit hippo. It reviews the session, writes down only what earned permanence, and discards the rest on purpose. The deliberate forgetting is the point: it is what keeps the memory trustworthy. A project picks back up mid-task, weeks later, as if no time had passed.

/taste-for-makers, the anchor against the mode

Named after the Paul Graham essay. A distribution left unattended samples its mode, and the most probable answer is the most average one. So what good looks like gets decided in advance, in writing, and the output is held against that instead of a general sense of the day. Correct is the floor.

There is one more place the argument has to go, because the trajectory does not start at the model's first token. It starts in my head. A prompt is the output of another chained, probabilistic process: a mind conditioning on its own past, confirming its own framings, holding constraints it forgets to state. I am a trajectory too, and the principle does not care that I run on carbon: no trajectory can anchor itself. When I scored forty-one of my own sessions against Anthropic's field guide, evidence required for every claim, the data agreed. The most common cause of rework was not the model. It was millimetres of mine, constraints disclosed turn by turn as they occurred to me. So the newest anchors point the other way. Before ambitious work begins, Claude interviews me, one question at a time. Before it executes, it surfaces the judgment calls I would most likely veto. Before anything ships, it quizzes me until I can defend the work myself. The anchoring runs in both directions now: my files hold its trajectory to the problem, and its questions hold mine.

There is an economy underneath all of this. Every failure costs twice, once to produce the wrong answer and once to catch it. Unanchored, that second cost evaporates at the next reset, and you pay it again next session. Anchored, each correction becomes a line in a ledger, a case in a standard, a trap added to a test suite, something that stays. The same effort compounds instead of repeating.

And none of the machinery is new. A ledger, a second signature, an archive, a written standard: these are the old controls that made human work reliable for centuries, aimed now at a probabilistic machine instead of a person.