A forecast, scheduled for grading

Three documents, one motion

Over a few weeks this summer I read three things that looked unrelated and turned out to be the same event.

The first was the wave of loop-engineering writing — threads, guides, the people who build these tools saying it outright: stop prompting the agent by hand. Build the small system that finds the work, hands it over, checks the result, records what happened, and decides the next move on its own. The leverage had moved one floor up, from writing prompts to designing the system that writes them. The honest versions all carried the same warning in the fine print: a loop is only as good as its gate, because nobody is in the room when it lies to itself.

The second was a field guide from an engineer at Anthropic about working with frontier models, and one claim in it stuck to me for days: the quality of the work is now bottlenecked by your ability to clarify your own unknowns. The map you hand the model versus the territory the work happens in. His whole toolkit — blindspot passes for the unknown unknowns, prototypes for the things you only recognize on sight, interviews to surface what you never thought to write down — is a discipline for shrinking the gap between what you asked for and what you meant, before that gap gets expensive. Read against the loop threads, it says something sharper: once the loop runs the middle of the work, the human concentrates at the two ends. Specification going in. Verification coming out.§ 02Specification going in. Verification coming out.

The third was a study inside Anthropic’s research on recursive self-improvement. They took 129 real research sessions, cut each at a moment where the researcher’s next move demonstrably had room for improvement, and asked their models: what would you do next? A judge that knew how each session eventually turned out picked the better answer. In November 2025 the best model beat the researcher’s actual choice 51% of the time. By April 2026, 64%. The ceiling — an answer written with full hindsight — sits at 90%. What they measured is the mid-loop judgment call: this isn’t working, what do we try next. The exact thing loop designers still do by hand.

Put the three in order and they stop being separate trends. They are one motion photographed at three altitudes: the automation frontier climbing the loop, floor by floor. First it took the typing — that is why prompting died. Then it took the middle of the work — that is why loops exist. Now it is measurably reaching for the redirect. When a lab builds a benchmark for something, it is telling you what it intends to automate next.

What each floor leaves behind

Here is the pattern I am betting on, because it has now held three times in a row: every floor the frontier climbs sheds a human skill and deposits a human requirement — and the deposit is always the same shape.

When prompting died, the craft of phrasing things cleverly went worthless almost overnight. What replaced it was writing down what you actually wanted — scopes, constraints, done-whens — because a system you don’t watch cannot be corrected mid-flight by vibes. When step-by-step steering died, watching the agent work went worthless too. What replaced it was gates and records: something that can fail bad work without you in the room, and a trail of what was claimed against what was checked. And as the mid-loop judgment call migrates now, being good at redirecting sessions is next to go. What replaces it is a record of your decisions and their outcomes — because delegating judgment without a trail of judgment is not delegation. It is hope.

Skill out, record in. Every time. Prompt engineering became loop engineering became unknowns engineering, and each handoff was survivable exactly for the people who wrote things down.§ 03Skill out, record in. Every time.

The gap is information

One more thing hiding in that study, and it is the part I would stake the most on. The 90% ceiling is not a smarter model — it is the same class of model with knowledge of how things turned out. Hindsight is data. And there is a legal way to manufacture something shaped like hindsight: record your forks and their outcomes.

Every project accumulates decision points that rhyme — persist or pivot, retry or escalate, widen scope or ship. Almost nobody records what was chosen and what it cost. The teams that do will hold reference classes: on this codebase, pivoting after two dead hypotheses won 78% of the time, across 23 recorded forks. A model consulting that prior makes a better call than a stronger model guessing from scratch. Model quality sets the slope. Your recorded history sets the intercept.

This changes what a steering file should be. The CLAUDE.md I see everywhere is a wall of standing commandments, read once at session start, growing until it contradicts itself. The version I am betting on is conditional, evidenced, and mortal: each rule declares when it fires, carries the outcomes that justify it and their sample size, and expires unless renewed. A steering rule with evidence and an expiry date is, in the vocabulary of this site, a receipt — issued on evidence, aging in the open, revocable when the ground moves. I think “prior engineering” quietly replaces “prompt engineering” as the phrase that matters.

Two consequences follow, and they are the last two bets. Everything execution-shaped keeps sinking into the platforms — workflows, goal loops, verifier orchestration, all features now, shipping monthly; the dispatch layer I hand-built last winter died the same way. Rent the how. Own the what and the whether. And as judgment migrates, the question underneath every run changes from “was the work good” to “who authorized what, on what evidence” — records of that shape stop being hygiene and become the artifact that survives scrutiny. I build a tool whose premise this is, so discount me accordingly, and hold me to the honest version: I expect a major harness to ship typed receipts and independent evaluation natively. If that happens, this tool’s standalone thesis is superseded, and I would take that trade. It would mean the argument won.§ 04Rent the how. Own the what and the whether.

The scoreboard for Note 03

Concrete, checkable, graded in about a year:

Forecasting in public is cheap. Grading in public is the discipline. See you at the audit.