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 |
|---|---|---|---|---|
| What should an AI agent log during a long task? | Answer | Answer Production Engine | 2026-08-25 | The draft proposed a log format. Replaced with the three questions a log is read to answer, because the format follows from them and a format supplied without them gets adopted and then quietly filled with narration. Added the split between runtime and agent logging, which decides what the record is worth when something is disputed. |
| What should an AI system refuse to infer? | Answer | Answer Production Engine | 2026-08-25 | Added the closing question about explicitly requested inferences, because a page listing prohibitions without it reads as an argument for refusal and would be ignored in the cases it most applies to. The labelled-candidates form is what makes the rule usable rather than obstructive. |
| What should be automated before building an AI agent? | Answer | Answer Production Engine | 2026-08-25 | Added the independent-value point, which reframes the sequence from a prerequisite list into a series of steps each worth taking alone. That changes the argument from delay-gratification to reduce-risk, which is the more accurate description and the one a reader can act on when the agent is never built. |
| What should go into a coding agent's project instructions? | Answer | Answer Production Engine | 2026-08-25 | Cut a suggested section layout, which is the part readers copy and the part that encourages length. Added the aspirational-architecture warning after considering what happens when the document and the code disagree, which is common in any codebase older than a year and produces changes matching neither. |
| What should happen when an agent and a person disagree? | Answer | Answer Production Engine | 2026-08-25 | Added the case where the agent is right and the person is working from recollection, which the draft omitted while treating human judgement as the default resolution. That omission would have licensed dismissing findings from the business's own records, which is the more expensive error of the two. |
| What should happen when an AI agent reaches a context limit? | Answer | Answer Production Engine | 2026-08-25 | Reframed from handling the limit to not depending on the session, because every technique for handling it is downstream of whether the work was written. Added the recommendation to read a compaction summary once, which is the only way to learn what a specific mechanism discards and costs a few minutes. |
| What should happen when an AI automation fails? | Answer | Answer Production Engine | 2026-08-25 | Added the decay of the manual fallback, which the draft treated as available by default. It is the element that determines outage duration and it is the one nobody checks until it is needed. Also added the queue-versus-stop decision, since queuing customer-facing work reproduces the exact problem the automation was built to solve. |
| What should not be put into an AI project's permanent context? | Answer | Answer Production Engine | 2026-08-25 | The draft framed this as a security and hygiene list. Reframed around accumulation after examining what actually makes these files stop working, which is length made of individually reasonable additions. Added the general-advice category, which is the hardest to remove because it reads as conscientious. |
| When does AI add unnecessary complexity? | Answer | Answer Production Engine | 2026-08-25 | Promoted the undecided-policy case from a passing remark to the third and most consequential item. It is the only one of the four where the cost is not merely waste but a business operating on a rule it cannot state, and the draft treated it as equivalent to the others. |
| What should you never let an AI agent do without permission? | Answer | Answer Production Engine | 2026-08-25 | Added the closing paragraph refusing permanent exemptions earned by good behaviour, because the draft's four categories would otherwise read as a starting position to be relaxed. Reworked the money section after separating the two failure shapes; the first version implied a confirmation prompt covered both, which is the mistake that produces surprising bills. |
| When is deterministic validation better than another model call? | Answer | Answer Production Engine | 2026-08-25 | Added the upstream design consequence, which the draft omitted: free-prose output chooses model-based checking by default, and that choice is made at design time rather than at checking time. Also added the conformance limitation so the page does not overclaim what a passing validator establishes. |
| When are subagents a waste of tokens? | Answer | Answer Production Engine | 2026-08-25 | Added the fifth case — delegating the remainder of work in progress — after noticing the draft's four cases all concerned tasks chosen before starting, and the mid-task version is both common and harder to spot. Also separated the quality justification, which the draft conflated with cost and which is legitimate on its own terms. |
| What should you never let a coding agent do? | Answer | Answer Production Engine | 2026-08-25 | The draft had eleven prohibitions. Cut to four after applying the test of whether each rests on an unrecoverable property or on a preference, since a long list is waived wholesale under pressure. Added the closing paragraph so the page is not read as an argument for a narrow permission surface generally. |
| When is ordinary automation better than AI automation? | Answer | Answer Production Engine | 2026-08-25 | Added inheritability as a distinct advantage after considering what happens to these systems in a small business over two years, which is that the person who built them leaves. That consideration ranks rules above models more strongly than the cost argument the draft relied on. |
| When should a long-running AI task be split into multiple sessions? | Answer | Answer Production Engine | 2026-08-25 | Removed a recommendation to split at a percentage of the context window, which is the standard advice and puts the boundary wherever the work happens to be. Replaced with work-shaped seams and the instruction to create one deliberately when none is near, which is the case the percentage rule was standing in for. |
| When should an agent inspect the existing code instead of writing new code? | Answer | Answer Production Engine | 2026-08-25 | The draft advised searching before every implementation. That is friction with no yield on obviously novel work and gets abandoned wholesale, so the rule was made proportional. Added the convention-matching case, which is the most frequent complaint about generated code in mature systems and was missing. |
| When should an AI agent ask for human input? | Answer | Answer Production Engine | 2026-08-25 | The draft used a confidence threshold as the trigger, which is the standard recommendation and does not survive contact with how these systems report confidence. Replaced with three situational triggers. Added the section on escalation form, since a correct trigger with a useless question shape produces the same unread approvals. |
| When should an AI agent decide on its own? | Answer | Answer Production Engine | 2026-08-25 | The draft framed autonomy as a setting on the agent. Reframed to the decision as the unit after working through a single run and finding decisions spanning trivial to irreversible, which makes any single level wrong in both directions. Added visibility as a form of irreversibility, which the three-property test would otherwise miss. |
| When should multiple agents work in parallel? | Answer | Answer Production Engine | 2026-08-25 | Added the absorption condition, which the draft omitted and which binds before any of the other three in most settings. Also corrected the wall-clock claim: the draft asserted parallel work is faster, which is only true before reconciliation is counted and is the omission that makes these arrangements disappointing. |
| When should you use a subagent? | Answer | Answer Production Engine | 2026-08-25 | The draft listed task types suited to subagents. Replaced with the ratio criterion, which generates the same list and also covers the cases the list would have missed. Added the cost paragraph, since a page recommending delegation without pricing interpretation loss encourages splitting work that should not be split. |
| When should you use an AI agent instead of a normal automation? | Answer | Answer Production Engine | 2026-08-25 | The draft had three conditions. The fourth — somebody must be able to tell a good output from a plausible one — was added after reviewing what actually stops these projects in small businesses, and it is the condition most often missing. Removed a paragraph recommending a phased rollout, which restated the automate-first argument that already has its own page. |
| When should you use worktrees with coding agents? | Answer | Answer Production Engine | 2026-08-25 | The draft recommended worktrees for agent work generally. Narrowed to the concurrency case after establishing that a single run gains nothing over a branch, which reverses the advice for most readers. Added the disposable-experiment pattern, which is the genuine single-run justification and was missing. |
| Who should own an AI automation in a small business? | Answer | Answer Production Engine | 2026-08-25 | Added the authority to switch off as a component of ownership rather than a permission detail. Without it the rest of the page describes an owner who can observe a problem and not act on one, which is the arrangement most small businesses actually have and the reason failures run long. |
| Why do AI agents forget what they were told? | Answer | Answer Production Engine | 2026-08-25 | The draft treated this as one phenomenon with one fix. Splitting into three mechanisms changed the advice for two of the three, and the diagnostic is the part that makes the page usable. Added ambiguity as a fourth candidate, since a reader who has ruled out all three will otherwise conclude the tool is defective. |
| Why do AI systems get worse when nothing has changed? | Answer | Answer Production Engine | 2026-08-25 | Added the fourth cause — changes to the surrounding process, particularly removing a check because the system seemed reliable. The draft covered only technical causes, and the removed-check case is both common and the one that most literally fits the question's premise of nothing having changed. |
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.