AGENTIC · THIS SITE
How edgardo-gutierrez.com is built: a static, evidence-first site maintained with AI agents
A bilingual profile generated by one script and kept in line by written rules, two hooks and a test suite. And by a person who decides what ships.
Drafted with Claude Code from this site's own repository and published in a release approved by Edgardo. How this note was written.
In brief
- edgardo-gutierrez.com, the professional profile of Edgardo Gutiérrez Rodríguez, is a static bilingual website generated by a single Node.js script with no npm dependencies and served from Cloudflare Workers Static Assets.
- The site began as a ChatGPT Site, moved to a private GitHub repository on 30 September 2026 and is now maintained with Claude Code under written rules, automated guardrails and its owner's direction.
- Every build regenerates the pages and runs structural and consent checks whose report is published at /evidence/verification.json, while results the site cannot control, such as search rankings or AI citations, stay marked as pending until they are observed.
| Site | Professional profile in English and Latin-American Spanish, an evidence checklist and these notes: 12 HTML pages. Evidence: sitemap, verification report. |
|---|---|
| First published on its own domain | 30 September 2026, release 1.2.0. Source: repository history (private). |
| Build | One Node.js generator, no npm dependencies, Node 20 or later. Source: repository (private); excerpt below. |
| Checks on every build | Consent-controller checks, then structural checks; a failed check fails the build. Evidence: consent checks, verification report. |
| Check counts | At release 1.5.9 (24 October 2026): 56 of 56 structural and 22 of 22 consent checks passed. Current counts are in the two reports above. |
| Evidence checklist | 25 criteria: 15 implemented, 8 partial, 2 pending evidence. Evidence: checklist, checklist data. |
| Hosting and deploy | Cloudflare Workers Static Assets; each change that reaches the main branch is built and deployed by Workers Builds. Evidence: a recorded deployment check in the production review. |
| Measurement | Google Analytics 4 through Tag Manager, and Microsoft Clarity, loaded only after consent; advertising storage always denied. Evidence: measurement scope. |
| Lab performance | Measured once, on release 1.4.2 (Oct 15, 2026): PageSpeed performance 99 on mobile and 100 on desktop, on the two profile URLs (/ and /es/). Not re-measured for later releases or for this page. Evidence: audit record. |
| AI tooling | Started as a ChatGPT Site; maintained with Claude Code, whose configuration has been in the repository since 24 October 2026. Source: owner; repository (private). |
| AI at runtime | None. The published site is static HTML, CSS and a small amount of first-party JavaScript; Tag Manager, Clarity and YouTube load only after consent or a click. |
What is this site, technically?
It is a set of static HTML files. One Node.js script, scripts/generate.mjs, reads my profile data and the wording for each language from JSON files and writes every page, the sitemap, robots.txt, the hosting headers and a folder of JSON evidence. There is no framework and no npm dependency; the build needs Node 20 or later and nothing else.
The content lives apart from the code. My career data is in one file, the English and Spanish wording in another, and the credentials, the talks and the checklist criteria in their own files. The generator only assembles them. That separation is what makes it safe to let an agent edit: changing a job title touches a data file, not a template.
The pages do not need JavaScript to be read. Scripts handle the theme switch, the mobile menu, the click-to-play video, the checklist filter and the consent controls. The text, the metadata and the structured data are already in the HTML the server sends. The CV PDFs are produced by a separate script from the same data and committed as files, so the deployment build needs no document tooling.
At release 1.5.9 (24 October 2026) the generator was a single file of about 39 KB and the English profile was roughly 53 KB of HTML. No model runs on the site: generative AI is used to build and maintain it, not to serve it.
"build": "node scripts/generate.mjs && node scripts/verify-analytics.mjs && node scripts/verify.mjs",Who decides, and what do the agents do?
I decide what the site says and what ships. The agents edit, build and check, inside limits I set.
What I do:
- I supply the facts: roles, dates, projects and credentials. The agents may not invent them.
- I make the decisions, and they are recorded: the positioning, what stays private, which tools may measure visits.
- I do, or approve beforehand, anything that happens in an account: Cloudflare settings, Search Console, Tag Manager.
- I ask for the changes. A request to change the site includes publishing it; a request to plan does not.
What the agents do:
- They read the rules before touching anything.
- They edit the sources in both languages, run the build and read the results.
- They commit sources and generated output together, and push.
- They report what changed, how it was checked and what was left unchecked.
- Two narrow sub-agents only review: one audits claims against the evidence rules, the other runs the build and reports. Neither edits or publishes.
The release that added this page is an example. Before anything was changed, six read-only specialist agents (technical SEO, answer-engine visibility, positioning, content and UX, architecture, and evidence compliance) each wrote a recommendation or a review. The maintaining agent consolidated them and I approved the plan. Only then did the editing start.
- OWNERAsks for a change, or approves a plan.
- AGENTReads
AGENTS.mdandproject-state.mdand finds the source file that owns the change. - AGENTEdits
src/in English and Spanish. An attempt to edit a generated file indist/is blocked by the hook. - BUILD
npm run buildgenerates the pages, then runs the consent checks, then the structural checks. If one fails, the build fails. - HOOKBefore the agent stops: do the release numbers match, and was
dist/rebuilt? - AGENTCommits sources and output together, and pushes.
- CLOUDFLAREWorkers Builds runs the same build and deploys
dist/frommain. - OWNER AND AGENTCheck the deployment of that commit and record external evidence with its date and scope.
The rules file every agent reads first
The rules are a plain Markdown file, AGENTS.md, in the repository since release 1.2.1 (30 September 2026). An agent starts every task by reading it together with the decision log, project-state.md. For Claude Code there is also a short CLAUDE.md that points to both and repeats only the hard rules.
It does three jobs:
- It says where each fact lives. A table maps every kind of change to the one source file to edit, and states that generated files are never fixed by hand.
- It sets the evidence rules. Do not invent metrics, clients, testimonials or mentions. Keep apart what I supplied, what a public record shows and what someone independent verified. Do not promise rankings, AI citations or a Knowledge Panel. Keep every lab score tied to the release and date it was measured on.
- It fixes the editorial constraints. Both languages state the same facts in their own words, and engineering and agile leadership get equal weight.
The rules are written in Spanish. One of them, in the original:
No prometer rankings, snippets, citas o recomendaciones de IA, indexación ni Knowledge Panel. Los estados externos permanecen pendientes hasta observar y registrar el resultado.
Do not promise rankings, snippets, AI citations or recommendations, indexing or a Knowledge Panel. External states remain pending until the result is observed and recorded.
When a decision changes, the log changes with it. On 24 October 2026, for example, the earlier decision not to add agent-specific configuration was reversed, and the reversal was recorded with its reason.
`editar src/` → `npm run build` → revisar lo afectado → commit (fuente +
`dist/` juntos) → push a la rama indicada. Cloudflare publica desde `main`.main. CLAUDE.mdGuardrails: a hook that blocks edits and a check before stopping
Two shell scripts, run by the tool rather than by the model, turn two rules into mechanisms.
Before an edit. The first script runs before every file edit or write. If the target is a generated file (any HTML under dist/, the evidence JSON, the sitemap, robots.txt, the headers file or the measurement scripts), it exits with code 2. Claude Code then cancels the edit and shows the agent the script's message, which says where the change belongs. Files in dist/ that are maintained by hand (styles, browser scripts, images and the CV PDFs) stay editable.
Blocked: $rel is generated by 'npm run build'.
Edit the source instead (src/*.json, src/checklist.mjs, src/static/, src/agentic/
or scripts/generate.mjs)
and run 'npm run build' to regenerate. Do not hand-edit generated dist/ output.Before stopping. When the agent is about to finish, a second script checks two things without writing anything: that the release number is identical in the three files that carry it, and that sources were not changed while dist/ was left untouched. If either fails, it exits with code 2 and the agent has to go back and fix it. It deliberately does not run the build: the build rewrites the timestamps in the evidence files, and a check that dirties the repository every time an agent finishes would create noise of its own.
if printf '%s\n' "$changed" | grep -qE '^.{1,2} (src/|qa/|scripts/generate\.mjs|scripts/agentic-content\.mjs)'; then
if ! printf '%s\n' "$changed" | grep -qE '^.{1,2} dist/'; thenThe same configuration file lists the few commands an agent may run without asking: the build, the checks, a local preview server and read-only Git commands.
The hooks are modest on purpose. They do not judge content; they only stop mechanical slips. Judging content is the job of the rules, the evidence-auditing sub-agent and me.
Evidence first: a checklist that keeps unproven results pending
Everything the site claims about its own visibility sits in a checklist of 25 criteria, each with a status, its evidence and its next step. A result nobody has observed stays pending.
The original brief was a list of search and AI-visibility criteria in five groups (SEO, AEO, GEO, AIO and SXO). Instead of declaring them done, the site publishes them at /check/ with one of three statuses. Implemented means there is something verifiable. Partial means a base exists and work or checks remain. Pending evidence means there is none yet. At this release: 15 implemented, 8 partial and 2 pending.
Three habits hold it together:
- Sources are labelled. What I supplied about my career, what a public record shows (a credential record, a conference archive) and what was independently verified are kept apart. A post on a channel I co-own is not counted as an independent mention.
- Numbers keep their origin. The PageSpeed lab scores on the checklist were measured on release 1.4.2 on Oct 15, 2026. They are not copied onto later releases as if they had been measured again.
- What depends on third parties is observed, not assumed. Whether a search engine shows a snippet or an AI product cites the site is recorded only with the query, the date and the scope of the test. Negative observations are kept as well.
Behind the page there is a folder of JSON files under /evidence/: the structured identity data, the credential records, the verification report, the consent checks and dated reviews. They are excluded from search indexing but anyone can read them.
It also means accepting that a criterion can stay partial or pending for a long time. At release 1.5.9 the two pending criteria were both outcomes that only Google decides to show: a featured snippet and an AI Overview.
What does every build verify?
Every build regenerates the site and then runs two suites of checks. If one assertion fails, the build fails.
- The consent suite loads the real consent controller into a sandbox, with a stand-in for the browser's document, and checks what it would do. Nothing is requested before a choice. A refusal is remembered. Each purpose loads only its own tool. Withdrawal clears the cookies the site can reach and reloads the page. Vendor code is never executed. It had 22 checks at release 1.5.9.
- The structural suite reads the generated files. It checks one H1 and one main landmark per page; canonical addresses and reciprocal language alternates; the indexing policy of each page; that every local link and anchor resolves; that heading levels do not skip; that the sharing image is a real file with the declared dimensions; that the structured data names one person across both languages; that expired certifications are marked and left out of the active credentials; and that engineering and agile leadership keep the same number of capabilities and cases. It had 56 checks at release 1.5.9.
The suite has grown with the site: 32 checks at the first published release (1.2.0, 30 September 2026), 56 at release 1.5.9 (24 October 2026). The rule for adding one is in the rules file: add a check when it would catch a failure that matters, not to restate the implementation.
assert(/^\d{4}-\d{2}-\d{2}T\d{2}:\d{2}:\d{2}Z$/.test(value),'Profile updatedAt must include time and UTC timezone; date-only values are invalid');The report is published with each release at /evidence/verification.json. It lists every check by name, a SHA-256 hash and size for every published file and, just as important, what the checks do not verify. Its own scope line reads: «Local structural verification of generated site files; not a live search or browser audit».
Cloudflare runs the same command when it deploys, so the checks run twice: once where the agent works and once in the deployment build. A failing check exits with a non-zero code, which is meant to stop that deployment. I have not tested this with a deliberately broken commit.
Measurement that waits for consent
Google and Microsoft measurement stays out of the page until the visitor allows it, and the two purposes are chosen separately.
The HTML carries only a small configuration block and a first-party script. On a first visit to the profile or to these notes, that script shows the choices and requests nothing from either provider. Visit statistics (Google Analytics 4, installed through Tag Manager) and navigation experience (Microsoft Clarity) are separate switches; allowing one does not allow the other. Advertising storage is denied in every case.
const defaults = {ad_storage:'denied', ad_user_data:'denied', ad_personalization:'denied', analytics_storage:'denied', personalization_storage:'denied'};What is sent is reduced on purpose: the page address without its query string or fragment, the referrer cut down to its origin, and 5 fixed event names for clicks on the CV, contact links, professional profiles, the talk reference and the video. No e-mail address, link text or full outbound address is included. When a visit arrives from a known AI-assistant link with no referrer, only a fixed source label is added, as the privacy notice explains. Clarity is skipped altogether when the address carries a query string, and the page text is masked.
The choice is stored in the browser for 180 days and can be changed from the footer of any page. Withdrawing turns collection off, clears the measurement cookies the site can reach and reloads the page. It cannot erase what a provider has already received. The checklist pages never load these tools, and neither does any host other than the production domain.
Two limits. First, the checks prove how my code behaves; that data actually reached the providers was confirmed separately, on 16 October 2026, in a production review, and a click is not evidence of a message sent. Second, Cloudflare adds its own performance beacon from the hosting side. It is not governed by these choices, and the privacy notice says so.
From commit to production
A change that reaches the main branch is the deployment.
The repository is private and lives on GitHub. Cloudflare Workers Builds watches the main branch, runs npm run build and publishes the dist/ folder as static assets on an existing Worker. None of my code runs in the request path. An unknown address gets a real 404 page.
Three details matter more than they look:
- Generated files are committed with their sources. The diff of a change shows the exact HTML it produced.
- A local build proves nothing about production. After a change reaches the main branch, the Cloudflare build check on that specific commit is read, and then the public responses. Only then is a change reported as published.
- Asset addresses carry the release number. A new release is not served from an old cache.
The move itself was the first deployment. The site was first generated as a ChatGPT Site, then migrated to this repository and published on its own domain on 30 September 2026 as release 1.2.0. Between then and 24 October 2026 it went through 17 releases. Before this release, one pull request, merged on 24 October 2026, carried release 1.5.9 and the agent tooling; the earlier releases were committed directly to the main branch. This release is the second to go through a pull request.
What went wrong, and what changed
Three dated examples. Each ended with a recorded cause, a bounded fix and a check that now runs on every build.
16 October 2026: a script was refused with 403/1010 (release 1.5.7)
- What was seen. At 20:34 UTC a request to the home page from the default HTTP client of Python's standard library returned HTTP 403 with Cloudflare error 1010. Browsers were unaffected; a production review in Chrome ran the same day.
- Cause. Cloudflare's security events showed that Browser Integrity Check had blocked that request, and an equivalent one to the Spanish page, because of the client's signature. The block was reproduced seven minutes later with the same kind of client.
- Change. I approved one narrow exception before it was switched on. Browser Integrity Check is turned off only for GET and HEAD requests to the site's listed public paths, its assets and its evidence files. The global setting, the firewall, DDoS protection and TLS were left as they were.
- Result. Three minutes later the same client received 200 on both profiles, the sitemap,
robots.txt, both checklists, the portrait and an evidence file. A request to a path outside the rule still returned 403/1010, which shows the exception is as narrow as intended. - Not established. Whether the same block explains earlier fetch problems in AI answer products. No Cloudflare event was attributed to them, and no block of Googlebot is inferred.
- Guard. A structural check now fails the build if the record loses the before-and-after pair for the same client, or the negative control.
- Consequence for this page. The exception lists exact paths, and at publication these notes are outside it. The same kind of scripted client may therefore still be refused on this address. Extending the rule is a decision about my Cloudflare account, and it is tracked in the checklist.
Evidence: hosting access record.
16 October 2026: a date without a time (release 1.5.1)
Search Console reported a non-critical warning: the profile's dateModified was a date with no time. It now carries a full UTC timestamp taken from the last substantial change to the profile data, not from each build, and a check rejects date-only, impossible or future values. As of release 1.5.9, Google's re-validation of its historical report was still tracked as pending.
22 October 2026: six duplicate addresses that were really two PDFs (release 1.5.8)
A Search Console export listed six duplicate addresses. All six were the two CV PDFs reached with version parameters, one of them over HTTP. Each PDF now declares its clean HTTPS address as canonical through an HTTP Link header, the versioned download links were kept, and HTTP was redirected to HTTPS at Cloudflare the same day. In the export I supplied, Google's validation was still pending. Evidence: indexing review.
The pattern is the same each time: record what was observed, change the smallest thing that fixes it, add a check, and write down what is still unknown.
What this does not prove
A passing build proves that the files are consistent with the rules. It does not prove outcomes.
- It is not a ranking, indexing, citation, recommendation or Knowledge Panel claim. Nothing here promises that a search engine or an AI product will index, show, cite or recommend this site, or generate a Knowledge Panel.
- The checks are structural and local. They are not a browser test of this release, a full accessibility audit, a test on physical devices or field performance data. The published report lists these gaps itself.
- The last full browser review on record is of version 1.1. It is kept as a historical baseline, not as a review of the current release.
- Lab scores belong to release 1.4.2. The notes pages have not been lab-measured.
- My career facts are owner-supplied. The site structures and labels them; it does not independently verify employment, project outcomes or the 90% deployment-time figure in my profile.
- It says nothing about recruiting systems. It does not show compatibility with any applicant tracking system (ATS).
- This is not a productivity claim. I have not measured how much time the agents saved, and I will not publish a number I cannot support.
- Agents still make mistakes. The hooks catch mechanical slips; they do not catch a wrong sentence. That is what review is for.
- One site, one person. This is how a small static site is maintained. It says nothing about client or employer work, and none is described here.
What comes next
This is the first note. More may follow on other projects of mine, in the same format: who did what, the guardrails, what went wrong and what the result does not prove. I am not setting dates.
For this site, the open items are the ones the checklist already lists: a manual accessibility and physical-device review, field performance data when there is enough traffic, and the external results that only time and third parties can produce.
If you want to talk about this way of working, get in touch.
How this note was written
This note was drafted with Claude Code from the repository's own files: the rules, the scripts, the evidence records and the commit history. Each factual statement was traced to its source before publishing, and the note was published through a pull request that I merged. The quoted excerpts are checked against the repository on every build, so a quote cannot quietly drift away from the code. No model version is named, because versions change and the method does not depend on one.