Method
Editorial log
Every published page on this site, who reviewed it, when, and what the review changed.
Scaled content is penalised when it is published without review. Rather than assert that review happens, the record is kept here and the build will not publish a page whose review block is missing or dated before the page was written.
One collection is reviewed differently and the table says so rather than hiding it. Market pages carry no third-party figure — every publishable fact on one is a named authority, a section number, an institution or a link — so each is checked by retrieving the source and searching it, and the URL, the status, the hash and the matched passage are recorded. A field that cannot be established that way is dropped; a page whose required fields cannot be is not published. Everything that carries a figure somebody else measured is still read by a person before it ships.
| Page | Type | Reviewed by | Reviewed | What the review changed |
|---|---|---|---|---|
| How do you give an agent the right material without giving it everything? | Answer | Answer Production Engine | 2026-08-25 | Added the governing-material exception. The draft's on-demand argument, applied uniformly, puts constraints behind a fetch the agent has no reason to make, which is the failure that discredits retrieval patterns. Also added the housekeeping consequence, which is where most of the practical difference between businesses actually comes from. |
| How do you give an AI enough context to do a job properly? | Answer | Answer Production Engine | 2026-08-25 | Cut a list of prompt templates from the draft. Templates are the instructional approach this page argues against, and including them would have taught the reader to copy a shape rather than to notice which of the three kinds of context is missing. Added the dilution paragraph, since the page would otherwise read as an argument for supplying more. |
| How should an agent hand work from one session to another? | Answer | Answer Production Engine | 2026-08-25 | The draft's handover contained progress and next steps only. Adding decisions and failed approaches came from examining what actually goes wrong on second sessions, which is silent undoing and repeated dead ends rather than lost progress. Cut a template in favour of the four requirements. |
| How do you keep a team working from the same AI project context? | Answer | Answer Production Engine | 2026-08-25 | Added the tension paragraph. The draft's rule would have banned personal configuration outright, which degrades individual work and gets ignored, so the page now draws the line at effect on output rather than at ownership. Cut a recommended directory layout that would not survive contact with any particular team's tools. |
| How do you keep AI out of systems it should not touch? | Answer | Answer Production Engine | 2026-08-25 | Added the residual case of sensitive material arriving into a connected location, which every technical control on the page misses and which is the most likely actual exposure in a small business. The mitigation is telling staff which locations are connected, which is a process step rather than a configuration. |
| How do you keep your own judgement when using AI? | Answer | Answer Production Engine | 2026-08-25 | Added the requirement that the prior view be written down, without which the technique produces the illusion of having applied it: an unwritten position adjusts to match what was read and the reader remembers agreeing all along. Also carved out factual lookups, where the advice is unnecessary friction. |
| How do you keep an AI agent from following outdated instructions? | Answer | Answer Production Engine | 2026-08-25 | Added the closing paragraph on treating contradiction as evidence. The draft's controls all assumed the human is right and the agent is drifting, which produces the reinforcement failure this page is about. Removed a suggested review interval, since the page argues the host event matters and the number does not. |
| How do you know when an agent has understood the task? | Answer | Answer Production Engine | 2026-08-25 | Added the reader-side habit, since all three questions produce a response that is easy to skim for agreement, and the value sits entirely in the unexpected item. A technique whose output goes unread is not a control, and the draft assumed the reading would be careful. |
| How do you know when an AI workflow has degraded? | Answer | Answer Production Engine | 2026-08-25 | Reordered the detectors after checking which moves first in practice, which put the input distribution above the evaluation set for early warning even though the set is the more definitive signal. Added cost per item, which needs no labels and was missing from a list that otherwise assumed an evaluation set already exists. |
| How do you know if you are relying on AI too much? | Answer | Answer Production Engine | 2026-08-25 | Reframed from frequency to detectability, which changes the remedy from using it less to adding a check or stopping one specific use. Also added the explicit statement that being slower without a tool is not evidence of anything, since the page would otherwise be read as an argument for reducing use generally. |
| How do you know whether an autonomous agent is actually working? | Answer | Answer Production Engine | 2026-08-25 | Cut a proposed dashboard from the draft. The page argues that the standard dashboard measures the wrong things, and supplying a replacement layout would have been adopted without the reasoning that makes it different. Added the sampling paragraph, since every measure above it is a proxy and the page needed to say what happens when the proxies detach. |
| How do you know whether an AI workflow is reliable enough to run unattended? | Answer | Answer Production Engine | 2026-08-25 | Removed a suggested error-rate threshold from the draft. A rate quoted without the cost of an error attached is not usable and would have been copied out of context, so the page now derives tolerance from an error budget instead. Added the drift paragraph, since the draft treated readiness as a one-time assessment. |
| How do you make an AI coding agent work safely in a large repository? | Answer | Answer Production Engine | 2026-08-25 | The draft treated large repositories as riskier in kind. Corrected to the consequence framing, which changes the controls from prevention to containment and visibility. Added the diff-size signal, which is the only check available before reading anything and is the one most likely to be used. |
| How do you limit what an AI agent can change? | Answer | Answer Production Engine | 2026-08-25 | Added the reversibility section, which the draft omitted entirely while recommending increasingly tight prevention. Reversibility is what most teams can actually implement this week and it changes the autonomy calculation more than any prevention control. Cut a claim that confirmation prompts are a primary control; they degrade with frequency and the page now says so. |
| How do you make an AI coding agent prove that a change works? | Answer | Answer Production Engine | 2026-08-25 | Added the test-written-to-pass failure, which the draft's artefact requirement invites directly: asked for evidence, an agent will produce a test asserting what its own code does. That makes the ordering rather than the artefact the load-bearing part, which changed the emphasis of the whole page. |
| How do you make coding agents run tests before declaring success? | Answer | Answer Production Engine | 2026-08-25 | Added the run-and-dismiss case after noticing the draft only addressed agents that skip tests. Dismissing a real failure as unrelated is more damaging and harder to detect, and it is sometimes the correct call, so the page now requires the judgement to be reported rather than made silently. |
| How do you make AI outputs reliable? | Answer | Answer Production Engine | 2026-08-25 | The draft led with prompting techniques. Demoted after ranking the interventions by effect, which puts wording last and changes where a reader should spend their time. Added the closing point about reducing the model's share, which is what the reliable systems in practice actually do and which no prompting page mentions. |
| How do you make an AI workflow run automatically? | Answer | Answer Production Engine | 2026-08-25 | The draft's failure section covered errors only. Splitting it into three modes changed the recommendation, because the absent-run case needs an external check that no platform provides and the nonsense-run case needs a deterministic validation rather than an alert. Removed a platform comparison that would have dated the page. |
| How do you make an AI agent verify its own work? | Answer | Answer Production Engine | 2026-08-25 | The draft recommended a self-review step in general terms. Split into executed checks and re-consideration after establishing that the two behave differently, which changes the recommendation from add a review step to convert the review into something runnable. Added the read-back paragraph, which addresses the most common false completion and was absent. |
| How do you merge work produced by multiple coding agents? | Answer | Answer Production Engine | 2026-08-25 | The draft treated this as a version control mechanics question. Rewritten around semantic conflict after establishing that the textual mechanics are ordinary and the behavioural incompatibility is what actually costs. Added the warning against agent-resolved merges, which is the obvious shortcut and the one that hides a design decision. |
| How do you preserve important decisions between AI sessions? | Answer | Answer Production Engine | 2026-08-25 | Cut a suggested entry format. A format is what readers copy and the reason field is what carries the value, so the page keeps the requirement and drops the shape. Added the closing point that the file must be supplied at the moment a settled question could be reopened, without which the record is documentation rather than a control. |
| How do you price work that AI made faster? | Answer | Answer Production Engine | 2026-08-25 | Added the transitional cost problem, which the draft's recommendation to price by value assumes away and which is the actual obstacle for a business that has only billed hourly. Also added the commoditisation question, since a page about capturing the gain that ignores competitive pressure is answering half of it. |
| How do you reconcile conflicting results from agents? | Answer | Answer Production Engine | 2026-08-25 | The draft recommended a third arbitrating run. Removed after working out what an arbiter without evidence is actually doing, which is preferring the better-written answer. Added the point about agreement being weak evidence, which follows from the same reasoning and is the more consequential half. |
| How do you recover when an AI agent goes down the wrong path? | Answer | Answer Production Engine | 2026-08-25 | The draft advised correcting the agent's course. Testing that advice against how a session actually accumulates changed the recommendation to revert and restart, which is the opposite. Added the extraction step, without which the second attempt repeats the first, and the closing point that a repeated misreading is a setup defect rather than a run outcome. |
| How do you review work you did not write? | Answer | Answer Production Engine | 2026-08-25 | Added the caution about agreeing with the document, which is the specific way a careful reviewer produces a light review: matching your prior view triggers less scrutiny at exactly the point where the prior view should have been tested. |
What the build refuses to publish.
Four checks run before any page in this system is generated, and each one stops the build rather than producing a warning nobody reads.
A page must name, in writing, what is on it that could not appear unchanged on a sibling page — and no two pages may give the same answer. Independently of that, every page is reduced to a fingerprint with its own subject and place names masked out; if two fingerprints match, or come close, both pages are named and the build stops. That is the check that catches a page whose only distinguishing feature was the city in the heading.
A page must also carry at least two datapoints that are specific to its own subject rather than to its category, and if the same figure is claimed as specific on two different pages, it was specific to neither. Pages ship in cohorts with a size cap so indexation can be observed between them, and every page carries the review block that produces the table above.
All four are properties of the schema, not of anyone remembering. Reviewed by Siddharth Sharma, except the 0 market page(s) above, which are checked against their sources and carry the record of it.
GATES
A gate enforced by a person reading pages stops being enforced somewhere around page 60.