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 should AI read your database without breaking anything? | Answer | Answer Production Engine | 2026-08-25 | Reoriented the page from write protection to load, after establishing that read-only access addresses the risk people fear and not the one that occurs. Added the point about plausible figures from real data, which is the output-side failure and is more consequential than any query that errors. |
| How should AI help with scheduling and dispatch? | Answer | Answer Production Engine | 2026-08-25 | Added the two systematically optimistic inputs. The draft treated the constraints as given, and a scheduler built on nominal travel times and booked durations produces a day that cannot be worked regardless of how good the allocation logic is, which is the usual reason these systems get abandoned. |
| How should AI systems fail safely? | Answer | Answer Production Engine | 2026-08-25 | Added the silent fallback section, which the draft treated as an implementation detail. It is the pattern that produces the worst outcomes because it removes the signal at the exact moment it is needed. Also added the destination requirement, without which everything else on the page produces a correct refusal nobody sees. |
| How should AI-written code be reviewed? | Answer | Answer Production Engine | 2026-08-25 | The draft proposed a review checklist. Replaced with the ordering argument, because the items on such a checklist are the ordinary ones and the useful claim is that they should be applied in a different sequence. Added the volume paragraph, which is the constraint teams actually hit and the one a checklist silently makes worse. |
| How should AI systems handle conflicting sources? | Answer | Answer Production Engine | 2026-08-25 | Added the averaging failure, which the draft did not mention and which is the version of this problem that produces the most damage in numeric workflows, because the resulting figure has no source anyone can trace. Also added conflict logging as the durable output, since handling a recurring conflict repeatedly is worse than fixing it once. |
| How should an agent decide what to do next? | Answer | Answer Production Engine | 2026-08-25 | Added the failed-attempt record, which the draft omitted and which is what distinguishes a next-step decision from the previous one after something has gone wrong. Without it the same approach is chosen repeatedly because it still looks correct, which is the retry loop the cost controls exist to bound. |
| How should AI be connected to a messaging channel like WhatsApp? | Answer | Answer Production Engine | 2026-08-25 | Added the timing rule and the record-writing requirement, neither of which was in the draft. The first is specific to this channel and is the most common source of relationship damage; the second determines whether the channel is integrated or merely connected, which is the difference between solving the fragmentation problem and adding to it. |
| How should an agent know when a task is finished? | Answer | Answer Production Engine | 2026-08-25 | Added the partial-completion section after noticing that everything else on the page checks the quality of work performed and none of it detects work not attempted. That is the failure operators actually report, so a page omitting it would have been confidently incomplete about incompleteness. Cut a paragraph recommending stronger completion prompts. |
| How should an agent handle a task it has never seen? | Answer | Answer Production Engine | 2026-08-25 | Added the specification-gap distinction, which changes the response entirely: a case that is routine for the business does not need novel-case handling, it needs to be in the definition. Without that distinction the page recommends building exception machinery for work that should have been in scope from the start. |
| How should an agent recover context after a session ends? | Answer | Answer Production Engine | 2026-08-25 | The draft treated recovery as supplying the handover. Splitting it into read and verify changed the recommendation materially, because the most expensive resumption failures come from state that was accurate when written. Added the statement-of-understanding turn, which is the only mechanism that catches an incomplete handover before work is built on it. |
| How should an agent use previous work without blindly trusting it? | Answer | Answer Production Engine | 2026-08-25 | Added the closing counter-case after the draft read as an argument for re-verifying everything, which makes multi-session work more expensive than single-session work and would be quietly abandoned. Reworked the human comparison, which originally implied machine output is less reliable in general; the difference is track record and recourse rather than accuracy. |
| How should an AI agent be given a job? | Answer | Answer Production Engine | 2026-08-25 | The draft included a brief template. A template is a shape to copy and the page argues that the shape is not the problem, so it was replaced with the four components and the reasoning for each. Added the note on brief length after confirming that adherence degrades rather than improves with additional instruction. |
| How should an AI agent handle a production incident? | Answer | Answer Production Engine | 2026-08-25 | The draft allowed remediation with approval. Removed after considering what approval actually consists of during an incident, which is not the same activity as approval on an ordinary afternoon. Kept the rollback exception because it is genuinely different and refusing it would have made the page unusable for teams that already do it. |
| How should an AI agent handle uncertainty? | Answer | Answer Production Engine | 2026-08-25 | The draft recommended calibrated confidence reporting. Removed after considering how such a score would be used: attached to an assertion it invites arithmetic its calibration will not support. Replaced with named assumptions. Added the failure-versus-caveat distinction, which decides how the run is treated operationally and was missing. |
| How should an AI automation be designed? | Answer | Answer Production Engine | 2026-08-25 | Added idempotency and processed-item state, which the draft omitted entirely and which produce failures that get attributed to the model. Also added the exception path as a design element rather than an afterthought, because an undefined one is how a successful-looking automation accumulates a hidden backlog. |
| How should an AI system react when evidence is missing? | Answer | Answer Production Engine | 2026-08-25 | Added the downstream requirement after noticing the draft stopped at reporting. A reported absence that the workflow continues past is the same silent gap one step later, which makes the reporting decorative. Also added absence counting, which is where the durable value of this discipline sits. |
| How should an AI workflow handle exceptions? | Answer | Answer Production Engine | 2026-08-25 | Added the acceptable-rate threshold, which the draft omitted and which is the only thing preventing the slow failure where an automation covers less and less while its description stays the same. Also promoted counting by type, which was a closing remark and is where the compounding value is. |
| How should coding-agent sessions be divided? | Answer | Answer Production Engine | 2026-08-25 | The draft divided sessions by duration and context usage, which are properties of the tool rather than of the work. Replaced with the reviewable change, which is a unit the process already defines. Added the investigative session as an explicit exception, since a strict change-per-session rule has nowhere to put understanding work. |
| How should context be compressed without losing important information? | Answer | Answer Production Engine | 2026-08-25 | The draft recommended better summarisation prompts. Replaced with naming survivors in advance, because the failure is not summary quality but that nobody chose what to lose. Added the requirement that a compressed context declare itself, which changes how the next reader treats it and was absent. |
| How should evidence be separated from interpretation? | Answer | Answer Production Engine | 2026-08-25 | Added assumptions as a third category, which the two-way split leaves nowhere to put and which is where most of the actionable content sits. Also added the readability caveat, since the draft's structure taken literally produces a document nobody reads, which defeats the checking it was designed for. |
| How should long-running agent work be checkpointed? | Answer | Answer Production Engine | 2026-08-25 | Added the case of a checkpoint that faithfully records a wrong path, which the draft's design does not survive: every recommendation assumed checkpoints are correct, and the recovery mechanism when they are not is retaining the earlier ones. That changes the storage requirement from a current-state file to an appended history. |
| How should long-running agent tasks be structured? | Answer | Answer Production Engine | 2026-08-25 | The draft recommended periodic summarisation of the session. That is the mechanism this page identifies as lossy, so it was replaced with writing durable stage outputs and reading them back. Added the ordering point about checkable stages first, which changes how a plan is built rather than only how it is recorded. |
| How should parallel coding-agent work be isolated? | Answer | Answer Production Engine | 2026-08-25 | Added the closing case against parallelism. The draft assumed parallel work was the goal and described how to do it safely, which omits that two sequential runs are simpler and avoid every problem listed. Also added shared build artefacts, which cause the failures people misattribute to the agents themselves. |
| How should persistent work state be stored? | Answer | Answer Production Engine | 2026-08-25 | The draft weighed storage options on throughput and durability, which are the wrong axes for this workload. Rewritten around legibility during failure, which reverses the ranking. Kept the concurrency exception explicitly so the page does not read as an argument that files are always sufficient. |
| How should subagents return results? | Answer | Answer Production Engine | 2026-08-25 | Added the observed-versus-inferred split and the marking requirement, neither of which the draft had. Both address the same failure from different ends: a delegated result acquiring more authority than it earned, which is how a single weak conclusion propagates through everything the main run does afterwards. |
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.