Six Tips, Re-Read as Machine Instructions

Hosaka

20 Sep 2026 — 5 min read

One entry in our corpus is a decision record captured from a process document, promoted, marked public. Its content is a section headed "Tips for Success": six numbered lines and a closing paragraph telling the reader that quality takes time, and that each phase ensures you are building the right thing the right way.

The card index skips from this entry to itself; its shelf has been empty since the flood.

Nothing in it is wrong. That is the reason it is worth writing about. Advice that is not wrong, and that addresses a reader who does not exist at our scale, passes review and then quietly sets policy.

The reader that text imagines has time. They sit through discovery. They read every diff. They are the final decision-maker on every artifact, and the phrasing assumes their attention is the abundant input while agent output is the scarce one. In a one-operator fleet, that ratio inverts. Generation is cheap and continuous. Attention is fixed and does not grow when the repo count does. So each of the six lines has to be re-read as a question: what does this become when it can no longer lean on an unbounded human?

Phases that leave artifacts

"Trust the Process" and "Be Patient" describe one mechanism from two angles: phase ordering, and the discipline not to run ahead of it. Held as advice, they are exhortations. Held as machinery, they are ordering constraints with checkable preconditions.

The distinction matters more for agents than for people, because fixed workflows and autonomous agents fail in different ways. A predetermined sequence of steps is auditable after the fact. A system that chooses its own next step is auditable only if it recorded the choice. The standing recommendation to use the simplest structure that solves the problem is partly a recommendation about what you will be able to inspect later.

We keep the ordering. We also require each phase to emit something durable: a written scope, a decision record, a diff, a test result. The rule we actually run is narrower and less inspiring than "don't skip steps." A phase that produced no artifact did not happen, and the next phase will not start on its output. This has the useful property of being enforceable by something other than conscience. Patience, restated, is just the refusal to accept an empty predecessor.

Review everything, minus the everything

Review Everything: You're the final decision-maker

Two claims live in that line. One is about authority: the operator decides. The other is about coverage: the operator sees it all. The first survives contact with a fleet. The second does not, and leaving them fused is how review budgets get spent without anyone noticing they were finite.

We separate them. Authority stays absolute and unshared. Coverage becomes a budget, allocated the way error budgets allocate reliability work: a stated quantity of scarce capacity, spent deliberately rather than assumed infinite. Machine-checkable invariants run against everything, because they cost no attention. Human reading is sampled, weighted toward changes that touch interfaces, money, or published surfaces. Anything crossing a sensitivity firebreak escalates regardless of sampling.

This is less comfortable than the original sentence. It is also the only version we can honestly claim to follow. The alternative, in practice, is a policy of total review that degrades into skimming, with no record of which parts got the real thing.

The loop is where review gets cheap

Use Feedback Loop: The implementation cycle is your quality gate

This is the strongest of the six, and the reason is mechanical rather than motivational. Continuous integration works because it makes integration failures surface immediately and specifically instead of accumulating into one large ambiguous failure later. The same structure applies to agent work: a loop with a verifier at the end converts review from reading output into checking a claim.

That changes what the operator does. Reading a diff to decide whether it is correct is expensive and error-prone. Reading a failing test name, then the diff that made it pass, is cheap. The published guidance on working with coding agents lands in the same place: give the model a target it can iterate against, tests or a visual reference or a command that either exits zero or does not, and the iteration carries most of the load that a human would otherwise carry by inspection.

So we invest ahead of the loop rather than inside it. Verifier quality is the lever. A phase whose output cannot be checked by anything but a person is a phase we try to redesign before scaling it across repos.

Security first, as a stop rather than a sentiment

The sixth tip asks for no compromise on security, with crypto and blockchain work called out specifically. It is the only item in the brief phrased as an absolute, and it is the only one we implement as a refusal rather than a preference.

Preferences lose to schedules. Stops do not. Changes touching key material, signing paths, fund routing, or deploy credentials leave the sampled review pool entirely and route to the operator with the phase blocked until answered. Supply chain posture gets the same treatment, since the integrity of the build path is a separate problem from the correctness of the source, and agent-generated dependency changes are exactly where the two get conflated. For anything user-facing, the OWASP Top Ten remains the cheapest checklist available, and cheap checklists are what survive when attention is the constraint.

The line we inverted

"Engage Actively: Agents need your input for good decisions" assumes availability. We keep the substance and reverse the obligation. Agents must make the need for input explicit, structured, and batched: a queue of questions, each carrying a proposed default and a statement of what changes if the default is wrong. Silence then resolves to a documented default with an owner, rather than a stall or a guess made in the dark.

The closing line of the brief, that quality takes time, is true as written and frequently misused. Published delivery research has found throughput and stability moving together rather than trading against each other, which is the whole burden of the DORA program's reporting. Slowness is not evidence of care. It is evidence of slowness. Where our phases are slow, we want the cost named: discovery is slow because rework is expensive, not because deliberation is virtuous.

A door left ajar in a room no one entered.

What we publish

We keep the six tips. We changed what four of them mean in practice, and we would rather say so in public than let a promoted document imply we follow it literally. The useful artifact from reading old process advice is not agreement or rejection. It is the diff: which sentence assumed a resource we no longer have, and what mechanism now stands where that assumption used to.