βοΈ How it works
This guide is written with the help of AI agents; its author reviews and approves every change before publication. This page explains who does what, and what the AI is not allowed to do.
Three separate roles
| Who | What it does | What it does not do |
|---|---|---|
| A collection, no AI | Records new releases, security advisories, new models and vendor news, then drops what does not matter using fixed rules | It does not judge whether a page is wrong |
| An AI agent (vault-watch-agent) | Finds the passages that have become wrong, proposes new text and its source, checks its proposals against the whole guide, translates the changes into English | It changes no file itself and publishes nothing |
| A human, the author | Starts the agent, reviews each proposal, authorizes publication | They do not copy the watch by hand |
An ordinary program then applies the accepted corrections. It refuses any correction whose quoted text does not occur exactly once in the page, so the AI cannot rewrite a passage it misread. This is the human-in-the-loop principle.
How the projects fit together
Two projects work together: the guide holds the pages, the editorial rules and the list of sources to watch, and the agent collects the watch, then does the analysis. A third one, the publishing engine, turns the pages into a website.
flowchart TB
S["Public sources<br/>releases, advisories, models"]
subgraph G["The guide"]
W["Collection of the watch<br/>no AI"]
I["GitHub issue<br/>the list to process"]
R["Current pages<br/>and editorial rules"]
C["Automated checks<br/>and pull request"]
end
subgraph A["The vault-watch agent"]
N["Triage and sourced proposals"]
X["Check against the whole guide"]
T["Corrections applied<br/>and English translation"]
end
H1{{"Human review"}}
H2{{"Go-ahead to publish"}}
E["Publishing engine"]
Site["Live site"]
S --> W --> I --> N --> X --> H1
R -. "read on every run" .-> N
R -. "read on every run" .-> X
H1 -->|"accepted corrections"| T --> H2
H2 -->|"yes"| C --> E --> Site
classDef human fill:#fff4d6,stroke:#a65f00,color:#3d2a00
class H1,H2 human
The editorial rules stay in the guide and the agent loads them on every run: it keeps no copy that could drift.
Timeline of an update
- On Monday morning, every other week, the collection records the changes of the last 15 days. It runs on its own on GitHub. In the other weeks, it only steps in if a serious vulnerability is not covered by the guide.
- It opens a public GitHub issue listing the items to examine, and the author is notified by email.
- The author starts the agent by hand. This is a choice, to keep an eye on the budget: the interface shows an estimate of the runβs cost before it starts. This step could be automated if needed, without removing the two human approvals that follow.
- The agent sorts the items: to correct, possibly to add, or of no interest to the guide.
- It proposes corrections. Each one quotes the current text word for word, and gives the new text and a source.
- It reviews itself. Another pass compares each correction with the whole French guide and with sources on the web, then keeps, revises or rejects it.
- First pause: the review. The author accepts everything, nothing, or a selection. The agent can wait for days without losing anything.
- The accepted corrections are applied on a separate branch, and the English pages are updated.
- The checks run: metadata, diagrams, translations, external links of the changed pages.
- Second pause: the go-ahead to publish. Nothing leaves the authorβs machine before this green light.
- The pull request passes GitHubβs checks, then the author merges it.
- The site updates automatically after the merge.
Reported errors follow the same path: they are checked against primary sources, then go through the same review and the same pull request. So do two other kinds of work: trimming the pages when they grow heavy, and looking for what the guide is missing (details).
A cloud AI for a guide on self-hosting?
Yes, and it is better to say so. The agent calls Claude models through the Anthropic API, a cloud service. It only processes the content of this guide, which is meant to be public, and no confidential data, so the question of sovereignty does not arise as it would for medical or legal files. This choice is not advice for your sensitive data, which this guide helps you keep on your own premises.
What the upkeep costs
Cost of the Claude API calls at list price, measured in November 2026, with Claude Opus 5.5 for the triage (medium effort), the proposals and the cross-check (high effort), and Claude Sonnet 5.5 for the translation (medium effort). Another model or effort level would give other figures:
| Run | Cost |
|---|---|
| Watch with simple page excerpts | $0.90 |
| Full watch, checked against the whole guide | $2.95 to $4.45 |
| Fixing reported errors, 26 corrections | $6.60 |
| Looking for what is missing: the leads, then the additions for 2 to 5 leads | $2.28, then $8.70 to $12.95 |
| Trimming the pages, with Claude Fable 5.2: the report, then 30 findings handled | $6.28, then $18.87 |
The public journal lists each call, with its model and cost.
These amounts are indicative: they are what the calls are worth at the APIβs list price. The authorβs calls went through the API access of his Claude subscription: he paid his subscription and nothing more.
The limits
- Errors remain possible. Check the cited sources before relying on a figure, and report an error if you find one.
- Dates are hard. A software release has a creation date, a tag date and a publication date. When its sources do not settle a date, the agent declines.
- The agent does not write new pages. It can add to existing pages what they are missing, after a check on the web and a review. When a subject needs a page of its own, it says so; writing it remains an editorial choice.