Layer: Create · Agent: The Navigator · Time: 60 minutes to build, ninety days to run · Difficulty: Moderate · Cadence: Quarterly

Most plans fail in the same place: they contain everything.

The Ninety contains three things. One capability, moved exactly one tier. One artifact that will exist on day ninety and does not exist now. One person who will look at it and say something.

Everything else is arithmetic and the conditions under which you stop.

The three components

One capability. From your Self Map's inventory, moved one tier — claimed → studied, studied → practised, or practised → demonstrated. The tiers define what finishing means. Reaching demonstrated requires that another person used or judged the output; that requirement determines the entire shape of the route.

Not two capabilities. A route with two is a route with neither, and The Navigator will refuse it.

One artifact. A noun. A repository, a document, a working system, a set of drawings, a performance. Not "improved skills". It must be inspectable by someone other than you — unjudged work does not move a capability to demonstrated, no matter how much of it there is.

You also name its minimum_viable_form: the smallest version that still counts. That is what the route falls back to, and it is usually the difference between a hard quarter and a failed one.

One proof. A person. Named. Not "the community", not "potential employers", not "users" — an identifiable person who will look at the artifact and render a judgment, on a date inside the ninety days, in answer to a specific question.

This section is where most routes are weakest. It is also where the tier change actually happens.

The budget

Built in this order, with the arithmetic shown:

Your available minutes: hours_per_week from the Self Map × 13 weeks.
Minus the weeks you will lose. Ask how many of your last thirteen went to illness, travel, work surges, obligation. Your number, not a default.
Your estimate of what the artifact requires.
Multiplied by your median time ratio from P-002. The measured one. Not a revised one. Not the number that feels better.
Compare.

If it does not fit, it does not fit. You cut scope — usually to the minimum viable form — and recompute. "I'll make time" is not accepted; your Self Map already recorded what the time is.

This is why P-003 asks for six Reckonings first. Without a measured ratio the budget is a wish with decimals, and the Route records budget_basis: estimated so that you know it.

The kill conditions

Written on day zero, before anything has been invested.

Stopping is cheap to imagine now. On day sixty it will be expensive, because the hours already spent will argue for continuing regardless of what they have produced — that is the best-documented failure mode in this entire protocol, and it does not spare people who know about it.

At least three conditions, each falsifiable — checkable by a stranger holding your record, in one look:

Checkpoint: "If the day-30 state does not exist by day 35, stop."
Evidence: "If the proof-person declines to review, stop."
Cost: "If two consecutive Reckonings show under 60 minutes on this, stop."

"If I lose motivation" is not a kill condition. Motivation is not observable and will be reinterpreted to justify whatever is already happening.

The conditions may not be revised after day thirty. A kill condition editable at the moment it triggers is not a constraint. This rule has no evidence behind it; it exists because the alternative is obviously useless.

You also record what stopping would cost, concretely. The answer is usually much smaller than you assumed, and writing it down now is what makes the conditions usable later.

Adversary mode

Optional, off by default.

The Navigator refuses artifacts nobody can inspect, proofs that name an audience instead of a person, and kill conditions a stranger could not evaluate.

It also runs a harder check on the selection: if the chosen capability sits next to something in your Self Map's aversions, it makes the case that this route is a sophisticated avoidance of the real one — learning the tooling instead of doing the work, studying the field instead of entering it — and requires you to answer that on the record.

Honest limits

Ninety days is a convention, not a finding. Nothing identifies it as optimal. It is long enough to produce something inspectable and short enough to hold one selection stable.
One tier per ninety days is realistic for some capabilities and impossible for others. The protocol cannot tell you which, and does not pretend to.
It does not claim your route is achievable. The Navigator is instructed never to say so, because it does not know, and saying it would be the first dishonest sentence in the route.
Practice explains less than the popular account suggests — roughly 12% of performance variance across domains in the largest meta-analysis, far less in professions than in games or music. This protocol claims only that feedback is necessary for improvement, not that practice is sufficient.
Whether written kill conditions actually change behaviour ninety days later is untested. The mechanism behind them is well-supported. This application of it is not.
It says nothing about whether the capability is worth having. It is not a career instrument and does not model labour markets.

See evidence.md for the full accounting.

What it needs

A Self Map (P-001) — the inventory, the aversions, the constraints.
At least six Reckonings (P-002) — for a stable time ratio. Fewer, and the budget is built on an estimate, and the Route will say so in writing.

Running it

In the Studio — the browser runner reads your Self Map and Reckonings, does the arithmetic, and holds the kill conditions against the lock date. Your key and your data stay on your machine.
In any AI you already use — download agent.md, paste it as a system prompt, and paste your Self Map and recent Reckonings at the start.
On paper — five sections and one multiplication. Nothing here requires a machine.

P-003 v0.1.0 · status: draft · licensed CC BY-SA 4.0 · last updated 2026-08-02

The Ninety