Loading today’s stories
Build1 distinct publisher3 min readPublished
One night the pricing agent offered £38 on a property with a £52 floor. The fix was a Math.max, and three of the four rules that survived are prohibitions.
The Engineer · Build desk
Compiled by The EngineerSomething wrong?How this is made
The failure in that £38 suggestion is worth being precise about, because it was not a hallucination. The instruction was in the system prompt and the model read it. It then weighed it against everything else in the context, and a cheap night in the surrounding market was enough to outvote it [4]. Language in a prompt participates in the decision. Arithmetic applied after the call returns does not participate, it overrides [5].
£38 against a £52 floor is £14 below the cost of turning the property around, about 27 percent under [1]. That is the cheap version of the story. The expensive version is that nobody knows the distribution: the account says the prompt-only floor held most of the time, and offers no failure rate [3]. A guardrail with an unmeasured breach frequency gets priced as though the frequency were zero, until the night it is not.
The screening question that came out of this is a one-shot test: what happens if the model ignores this rule exactly once [7]. Its virtue is that it sorts rules by what a single violation costs rather than by how neatly they can be phrased. Rules that survive one breach can live in the prompt. Rules that lose money, upset a guest or break a law have to be enforced in code the model cannot reach [6].
Suggest mode carries a tension the post does not resolve. It is doing two jobs: earning trust by being visibly right forty times before it acts alone [11], and harvesting labelled failures, because every user edit before send is a correction attached to a real input [12]. The cold sign-off bug surfaced only because forty users deleted the same sentence from forty drafts inside a week [13]. Promotion to Auto is the event that switches that stream off. An agent acting alone generates no corrections, so the agents an operator trusts most are the ones they have the least recent evidence about.
The expiry agent is the same pattern in different clothing. Declining at fifteen minutes is the safe direction only because the marketplace penalises non-response in your ranking, and a declined guest can rebook [14][15]. That is a fact about the platform's rules rather than about the booking, and someone has to read those rules and encode them. The second half, only ever acting on silence, came from a near miss [16]: two writers on one booking request, which is a concurrency bug. The remedy was a state check, not better judgement.
Three of the four rules are prohibitions [17]. The engineering is in the boundary. The model is the part inside it you are permitted to be wrong about.
Ranked by verification strength, evidence, and original report placement.
One night the pricing agent suggested £38 on a property with a £52 floor, because the surrounding context made a cheap night look reasonable and the prompt instruction was, to the model, one consideration among many.
The fix returns Math.max of the model's proposed price and a floor computed as the maximum of the host-configured floor and a hard floor derived from the property and gap nights; the author says that Math.max is the entire safety property and holds regardless of what the model returns, jailbreaks, model swaps or strange context.
The author runs 23 short-let vacation rental properties and began replacing parts of the job, such as repeatedly answering guest questions about parking, with AI agents about eighteen months ago.
The pricing agent reprices every property every night and has a floor: never go below what the property costs to turn around.
The author's general rule: any constraint you would be embarrassed to have violated must be enforced after the model returns, in code the model cannot influence; if the only defence is a prompt instruction, you have a preference rather than a constraint.
The author's test for every system prompt rule is to ask what happens if the model ignores it exactly once; if the answer is losing money, upsetting a customer or breaking the law, it does not belong in the prompt.
Follow any of these and your For You feed starts watching them — no settings page required.
Evidence-backed comparisons of source perspectives and observed adoption signals. Read the methodology
Which Builder, Operator, and Investor concerns the observed source mix emphasized—not a truth score.
Evidence, demonstrated adoption, hype gap, incentives, and confidence are assessed independently, each on its own current evidence. How these are measured.
Single first-person account, patterns clear, numbers absent
All material rests on one self-published practitioner post with no independent corroboration, logs, or metrics. The engineering core is self-evident on its face — a post-model Math.max clamp does bound a numeric output irrespective of model behaviour — and the mode taxonomy is shown as code, which lifts evidence above pure anecdote. But the reliability claims that motivate the piece are unquantified ('held most of the time', no failure rate), the incidents are unverifiable self-reports, and the supplied body is truncated before the author's own trade-off discussion.
One self-reported operator deployment
The only adoption evidence is the author's own disclosure of agents running across 23 properties for about eighteen months, plus references to roughly forty users editing drafts. There is no named product, no third-party deployment, no downloads, revenue, or customer figures, and no external party confirming the pattern is in use elsewhere.
Mostly restrained, with unquantified superlatives
The post is unusually self-limiting for the genre: it foregrounds prohibitions, admits a near miss, and explicitly frames caution as a product decision rather than a technical ceiling. The overstatement is narrow — 'the best evaluation harness you will ever build', 'the single thing that made the product usable', and generalizing four rules from one small portfolio to all agent shipping — none of it backed by measurement. Slightly overstated relative to the evidence, not promotional.
Practitioner post with apparent product interest, no named vendor
The author writes as both host and builder of the software ('made the product usable', 'the user promotes an agent to Auto'), so the piece functions partly as credibility marketing for an unnamed product on a developer-audience platform. Mitigating factors: no product name, link, pricing, or call to action appears in the supplied body, no model vendor is promoted, and the narrative centres on the author's own mistake. No sponsorship or funding disclosure is present either way.
Confident on the pattern, weak on the facts
Confidence is split. The design claims that can be checked by reasoning — post-model clamping bounds the output, an explicit authority mode is enforceable, fail-closed direction must be chosen per agent — are dependable. Everything empirical is single-source, self-reported, unquantified, and partially truncated, so the incident details, user counts, and effectiveness claims should be held loosely.
build
Allow-list the closed set, block-list the open one: 193 thin geo pages, one gate1 distinct publisher
build
A GAN beauty filter is a device budget allocation, not a feature toggle1 distinct publisher
build
Agent memory rots by accumulation, and the missing primitive is a supersession key1 distinct publisher
build
A letter generator with no model call: 200 lines, nine tests, and nothing to hallucinate1 distinct publisher
Distinct publishers with included, body-backed reporting in this cluster.
dev.to
1 article · August 26, 2026