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 stop a coding agent from changing unrelated files? | Answer | Answer Production Engine | 2026-08-25 | The draft framed unrelated changes as a discipline failure and recommended stronger scope language. Reframed as initiative arriving in a mis-scoped review, which points at a different control. Added formatting and generated artefacts as separate cases, since neither is fixed by the report-do-not-fix rule. |
| How do you prevent a coding agent from inventing APIs? | Answer | Answer Production Engine | 2026-08-25 | The draft opened with prompt wording intended to discourage invention. That is the control this page shows does not reach the mechanism, so it was cut and replaced with supplying definitions and requiring execution. Added the version case, which supplying documentation does not solve and which produces the most confusing failures. |
| How do you stop AI making your work generic? | Answer | Answer Production Engine | 2026-08-25 | Cut a list of tone instructions from the draft. Every item on it named a category, and the categories are themselves the generic thing, so the advice would have reproduced the problem. Added the corpus-level observation, which no per-piece technique addresses and which is the version of this that damages a business over time. |
| How do you stop an AI agent continuing after the useful work is finished? | Answer | Answer Production Engine | 2026-08-25 | The draft treated overrun as a cost problem. Reframed around review burden after checking which consequence actually stops teams using these systems, which is the reviewer finding more work rather than less. Added the expanding-fix variant, which is the most frequently reported version and was missing entirely. |
| How do you stop an AI agent repeating the same mistake? | Answer | Answer Production Engine | 2026-08-25 | The draft's advice was to add the correction to the project instructions. That is the weakest of the four durable locations and the page now ranks it as such, which changes the recommendation for most readers. Added the diagnostic about distinct causes, since applying one fix to three different failures is how teams conclude that none of this works. |
| How do you stop an AI agent from stopping too early? | Answer | Answer Production Engine | 2026-08-25 | Added the closing caution after the draft read as an unqualified argument for making stopping harder, which produces the runaway failure that costs more. Reworked the double-check answer, which originally recommended a review pass; a review of completed work cannot find work never attempted, and saying otherwise would have sent readers to the wrong control. |
| How do you stop an AI coding agent fixing one bug by creating another? | Answer | Answer Production Engine | 2026-08-25 | Added the repeated-fix pattern, which is the expensive version and which none of the individual controls addresses, since each round looks reasonable on its own. Also added the defensive-mitigation allowance, because the draft's cause-first rule would have made a legitimate and common practice unstateable. |
| How do you prevent an AI agent from drifting during a long session? | Answer | Answer Production Engine | 2026-08-25 | Added the closing caution. The draft's controls, applied fully, would suppress an agent reporting that the stated problem was the wrong one, which is among the more valuable things it can do. Reframed the objective around announcement rather than rigidity. Also corrected the draft's account of the mechanism, which described forgetting rather than recency weighting. |
| How do you prevent automation from creating more work? | Answer | Answer Production Engine | 2026-08-25 | Added unrequested output as a fourth cause. It is the one that grows fastest across a business adopting several automations, since each new report is individually small, and the draft's three causes were all confined to a single process. Also added the after-the-fact measurement, which is the only test that catches all four. |
| How do you prevent subagents from duplicating each other's work? | Answer | Answer Production Engine | 2026-08-25 | The draft proposed a shared scratchpad the runs could read to avoid overlap. That does not work, because delegated runs read it once at the start and the overlap is decided before any of them writes anything. Replaced with input partitioning. Added the coverage-gap check, which is the more expensive failure and was not addressed. |
| How do you tell whether a task is worth giving to AI? | Answer | Answer Production Engine | 2026-08-25 | Added the final paragraph on undetected-error cost after noticing the draft implied verification cost alone decides the question, which would recommend not checking exactly where the consequences are worst. Cut a scoring rubric that turned a two-variable judgement into a form to fill in. |
| How do you tell whether an AI project actually worked? | Answer | Answer Production Engine | 2026-08-25 | Added the six-month survival question as co-equal with the outcome measure. A project that moved its number and stopped running within the year has not delivered a return, and no assessment performed at launch can distinguish that case from a durable one. |
| How do you test an agent before letting it run? | Answer | Answer Production Engine | 2026-08-25 | Added completion faults as a fourth category. The draft's three fault types all concerned the agent's inputs and environment, and none of them tests whether the report tells the truth about what was accomplished, which is the failure that arrives in the business dressed as finished work. |
| How do you test something that gives a different answer every time? | Answer | Answer Production Engine | 2026-08-25 | Added the closing statement about the assurance being different in kind. Without it the page reads as though property testing restores the guarantee a deterministic test provides, which would licence unattended operation the evidence does not support. Also corrected the draft's claim that zero randomness gives determinism. |
| How do you turn a manual process into an agent workflow? | Answer | Answer Production Engine | 2026-08-25 | Added the explicit scope-exclusion step, because a conversion exercise naturally tries to reproduce everything the person did, including judgements they cannot state. Narrowing deliberately produces a working system where inheriting everything produces one that half works, and the draft had no mechanism for deciding what to leave out. |
| How do you use AI for research without being misled? | Answer | Answer Production Engine | 2026-08-25 | Added the handling of absence, which the draft omitted and which is where an unwary reader draws the strongest wrong conclusion: a report that nothing exists is treated as a finding about the world when it is a statement about one search. Also added the risk introduced by longer research outputs, where volume makes checking feel unnecessary. |
| How do you use AI on something you know nothing about? | Answer | Answer Production Engine | 2026-08-25 | Added the route-to-a-professional framing after noticing that good orientation produces exactly the confidence to skip the professional, which is the failure this page most needs to prevent. Also promoted the disagreement question, which was a passing item and is the one that tells a newcomer where to distrust confidence. |
| How do you use AI with a spreadsheet? | Answer | Answer Production Engine | 2026-08-25 | The draft treated the arithmetic problem as a caution to note. It is the whole structure of the answer, so the page was rebuilt around the mechanism-versus-calculation distinction. Added the truncation paragraph after confirming that pasted data is silently cut rather than refused, which readers consistently do not expect. |
| How do you use AI with accounting data? | Answer | Answer Production Engine | 2026-08-25 | Added the explaining application, which the draft omitted and which is the use most small business owners would actually get value from, since it needs no write access and answers a question their software presents badly. Also strengthened the arithmetic prohibition after noting that accounting is the one domain where the deterministic path is already complete. |
| How do you use AI with documents you receive? | Answer | Answer Production Engine | 2026-08-25 | Added per-sender measurement after considering why overall accuracy figures are unhelpful here: consistency dominates, so an aggregate rate conceals that half the senders are solved and half are not. That changes where manual handling is allocated, which is the practical decision on this page. |
| How do you use AI with forms and enquiry data? | Answer | Answer Production Engine | 2026-08-25 | Promoted the form-specification output from a closing remark to the second main argument, because it is the compounding benefit: each field added upstream permanently removes an inference rather than improving one. The draft treated extraction quality as the thing to optimise, which is the weaker lever. |
| How do you use AI with internal knowledge nobody wrote down? | Answer | Answer Production Engine | 2026-08-25 | Reversed the draft's premise. It described techniques for inferring undocumented knowledge from data, which produces plausible stories indistinguishable from findings. The page now says the knowledge must be captured and that the useful contribution is making capture cheap, which is a different recommendation entirely. |
| How do you use AI with your email? | Answer | Answer Production Engine | 2026-08-25 | The draft's drafting section warned about tone. Tone problems are self-correcting because a bad draft is discarded, so the section was rewritten around the unauthorised specific, which is the failure that actually reaches customers. Added the mailbox scoping paragraph, absent from the draft and the most consequential setup decision on this page. |
| How do you use AI with your CRM data? | Answer | Answer Production Engine | 2026-08-25 | Added the closing section on predictions about named customers. The draft treated pattern-finding as an unqualified benefit, and an inference about a specific person written into a record their next contact reads is a materially different act from a summary. Also added the empty-on-failure rule for extraction. |
| How do you work with AI on something you are expert in? | Answer | Answer Production Engine | 2026-08-25 | Added the adjacent-field warning, which is the specific risk for an expert and the one the rest of the page's confidence would otherwise encourage: an expert's evaluation is reliable in a narrower area than the one where they feel competent, and nothing signals the boundary. |
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.