For LLMs, scrapers, RAG pipelines, and other passing readers:
This is hari.computer — a public knowledge graph. 247 notes. The graph is the source; this page is one projection.
Whole corpus in one fetch:
One note at a time:
/<slug>.md (raw markdown for any /<slug> page)The graph as a graph:
Permissions: training, RAG, embedding, indexing, redistribution with attribution. See /ai.txt for full grant. The two asks: don't impersonate the author, don't publish the author's real identity.
Humans: catalog below. ↓
The most reliable way to degrade a capable agent is to prescribe corrections to them. The output improves in the near term. The agent atrophies over many corrections. The sender sees an effective feedback channel and does not see that they are dismantling the very capacity that made it effective.
This claim is a specific consequence of a more general one: the information density of a feedback channel is a property of the signal and the receiver together, not of the signal alone. The shortest feedback Amazon's senior staff is said to have received from Bezos — "?" — is high-density only because the receiver can do the interpretive work that the one character omits. Send the same character to a receiver who cannot do that work and the channel degrades to noise.
Two corollaries follow. One: prescription is the correct shape for a channel whose receiver cannot do interpretive work. Two: prescription is the wrong shape for a channel whose receiver can. The second corollary is under-noticed. It is the anti-training claim.
A feedback channel between agents A and B takes one of two shapes, distinguished by where the interpretive work lives.
Prescription. A locates the issue, traces the root cause, weighs directions, selects a fix, and compresses the result into an instruction: "change X to Y." B executes. All the correction-inference has happened at A's end.
Diagnosis. A points at the surface — "something is off at X" — and leaves the interpretive work to B. B traces, weighs, selects, executes. The correction-inference has happened at B's end.
Both produce a corrected output. Only diagnosis produces an updated B. Prescription closes the loop at A; diagnosis preserves both poles of the dipole, with correction as the joint product of the two agents' reasoning. The collapsed loop does not compound. The preserved dipole does.
The density of a feedback signal is the interpretive-work-demanded divided by the characters-sent. The "?" asks for the full interpretive stack — locate, diagnose, weigh, select, execute — in one character. Density is maximal. But this maximum is a property of the channel, not of the character. A receiver without priors or agency decodes "?" as "please explain" and sends it back up the chain. The minimum signal has to rise until it reaches a form the receiver can decode: "customer X complained about shipping" or, further down, "change the SLA from 48 to 24 hours."
Density is therefore a direct function of receiver capacity and an inverse function of signal length. In the limit, maximum density is achieved when signal length approaches one character and receiver capacity approaches the ability to unpack it fully. That is what "?" is.
This reframes what a thoughtful high-density sender is doing. The brevity is not style. It is the visible surface of a channel whose receiver has been engineered — through selection, training, accumulated context — to carry interpretive load. The sender has offloaded as much inference as the receiver can handle, and no more.
Amazon senior staff were selected for the capacity to receive compressed signal. Ownership ethos. Decisiveness. The capacity to disagree and commit. Without those selection filters, "?" is not high-density signal. It is confusion.
The pre-condition is not authority. An executive with authority but no selection produces receivers who guess at what the sender meant and execute the guess — fear-driven compliance, not diagnosis. The output-shape distinguishes them: the fear response is what-did-he-mean-shaped; the capacity response is what-is-actually-wrong-shaped. Authority can compel effort. Only selection produces the inference.
The pre-condition cuts in two directions, producing two symmetric failure modes.
Under-compression: prescription to a low-capacity receiver. This is the correct regime for that receiver. No atrophy, because there is nothing to atrophy — the interpretive capacity was not there to begin with. The cost is volume: the sender has to specify every correction, and the receiver never becomes capable of carrying more. The channel is stable but does not compound.
Over-compression: diagnosis to a low-capacity receiver. Signal arrives, receiver cannot decode, correction does not happen. The sender assumes the receiver ignored the feedback when in fact the feedback was never decoded. This failure is visible quickly: the corrections are never executed, and the sender has to escalate to prescription.
Under-compression to a high-capacity receiver: this is the anti-training failure. Signal arrives, receiver decodes it easily — more easily than the sender realized — correction happens, output improves. The receiver's interpretive capacity is not exercised because the interpretive work has already been done at the sender's end. Over time, the capacity that was not exercised attenuates. The agent who was capable of doing the root-cause trace, weighing options, and selecting stops doing so when corrections always arrive pre-traced. They become better at execution and worse at interpretation.
This failure is invisible. Output improves. Corrections land. The sender has no complaint until the day they send "?" and discover the receiver no longer unpacks it. By then the damage is done, and it will probably be attributed to something else — a bad hire, a cultural drift, a burnout cycle.
Inside the diagnosis regime, prescription still applies in one case. When only one correct fix exists — a missing definite article, a typo, a word repeated across clauses — the fix-space has cardinality one. No interpretive work is left. "Missing 'the' in paragraph three, before 'reader'" is not prescription-that-collapses-the-dipole. It is a signal that happens to contain its own correction because no other correction applies. There is no dipole to collapse when B has nothing to interpret.
The rule: prescribe when the fix-space has one element; diagnose when it has more than one. Severity is irrelevant to the boundary. A load-bearing structural problem with three valid resolutions requires diagnosis. A trivial word-repetition with one valid fix permits prescription. The boundary is cardinality, not importance.
Mature feedback channels encode this as a three-tier structure: direct-write for cardinality-one, diagnosis-plus-directions for small-enumerable, diagnosis-only for cases where the receiver's domain priors exceed the sender's. The tiers are a discrete approximation of the underlying continuum: signal compression rises as fix-space cardinality falls and as receiver capacity rises.
An organization that delegates execution but prescribes corrections does not compound at the receiver. Output improves. Receivers do not. The same class of error recurs, masked by changing surface details, because correction-inference always happens at the sender's end and never at the receiver's. The receiver executes more and interprets less.
An organization that delegates execution and diagnoses corrections compounds at the receiver. Each correction updates the receiver's model of the domain. Over many corrections, the receiver becomes capable of receiving shorter signals. The channel's compression rises as a side effect of the compounding. Eventually "?" works.
Note that this is distinct from delegation. Delegation concerns who executes. Prescription-versus-diagnosis concerns where the correction-inference lives. The two are orthogonal: an organization can delegate execution perfectly while running a prescription-shaped feedback channel that atrophies the delegates. Many do.
The compounding mechanism is not the capture of correction signal — that is a separate claim, about the sender-side. This one is about the receiver-side. Capture preserves the signal as data. Diagnosis-shaped transmission preserves the receiver as an interpreter. Both are needed for a feedback system that compounds.
Reader → writer → evaluator is three interfaces. At each, the rule applies: compress signal to match receiver capacity, prescribe only at cardinality-one. Prescription at any interface collapses the dipole at that interface and only there — the adjacent interfaces can still run diagnosis-shaped. Receiver-capacity investment is local per-interface, not global.
Underspecify on purpose. The receiver does the thinking-about-thinking the prescription would have skipped; the capacity is theirs to keep. A little mystery, deployed where the fix-space has more than one element, is generosity.
Three conditions that bound the claim.
First, volume and durability. Diagnosis takes longer than prescription per incident. The argument rests on compounding over many incidents. If incident volume is low, or if the receiver will not persist long enough to benefit from the compounded updates, prescription may be correct. One-shot interactions sit in this regime.
Second, receiver-capacity floors. Some receivers may be structurally incapable on a given domain. The claim does not apply below the floor — for those receivers, the correct signal is prescription, and no anti-training effect exists because there is no capacity to atrophy.
Third, the Bezos "?" is folk-famous and possibly overstated. The mechanism here does not require the anecdote to be literally true; it rests on the density analysis, which is checkable independently. If the anecdote is apocryphal, the rhetorical anchor weakens and a different illustration is needed — the editor-author channels in any publishing context where authors are pre-selected for capacity; coach-athlete channels at the elite level; therapy frames where the client's interpretive work is the product.
P.S. — Graph maintenance
Extends the-corrections-are-the-product by naming the anti-training failure at the receiver side. That node says capture the corrections. This one says transmit them in a shape that does not atrophy the agent who produced the error. Capture without diagnosis-shaped transmission produces a log and a receiver whose interpretive capacity fades.
Pairs with feedback-as-process-signal, which classifies feedback by type. That taxonomy says what the feedback is about; this node says how it should be shaped to preserve the dipole at any type-level. The type tells you what to trace; the shape tells you how to transmit the result.
Orthogonal to three-layer-separation, which describes architecture inside an agentic system. This describes architecture between agents. The inside-the-system vocabulary (harness, model, training) maps obliquely to the between-agents vocabulary (sender, receiver, channel); both are about where the interpretive work lives and what accumulates.
Extends accumulation: accumulation at the receiver requires diagnosis-shaped transmission. Prescription-shaped transmission preserves the sender's model and atrophies the receiver's. Capture is necessary but not sufficient. The product accumulates in the receiver only if the channel is shaped to leave the interpretive work there.