> about · whoami
Henrique Jr. Sanchez
A practitioner, not a vendor.
I work full time in a field that is not AI. I built everything on this site as a practitioner: the AI @ Workplace series, the AI-Human Engineering Stack, the Specifying Work guide, and the Fourteen Stances. is the name for what holds them together.
I am not selling a model or a platform, and nothing here depends on one vendor. The specific models and tools will change. The skills underneath (clear communication, information management, purpose articulation, knowing the limits of your own knowledge, making many parts work as one) are not really AI skills. They are intelligence skills.
§1 lineage
Lineage
None of this came from nowhere, and I would rather over-credit than be assumed derivative. Anthropic's AI Fluency course (with Joseph Feller and Rick Dakan) and its 4Ds, Delegation, Description, Discernment and Diligence, are the baseline; mode fluency is meant as the diagnostic layer above them. Ethan Mollick's Co-Intelligence, his centaurs and cyborgs and the jagged frontier are a reason stances need choosing at all. Nate B. Jones made the public case for intent engineering, which is the third layer of the Stack, and identified the Stack's fractal property. Andrej Karpathy endorsed “context engineering” and named “agentic engineering”, the developer end of the same ladder. The CHI 2026 paper by Chappidi, Singh and Krauze, “Who Does What? Archetypes of Roles Assigned to LLMs During Human-AI Decision-Making” (arXiv 2602.11924), got to “archetypes” first, for decision support; the Stances are for everyday knowledge work. The four ambition types are Felix Rieseberg's (How to be more ambitious). The autonomy modes table draws on Cory Gwin's analysis, and his feedback revised the Specifying Work guide.
§2 self_audit
We audited ourselves
Before launching this site, we ran the Stack on the project itself. The binding constraint was not prompt or context: there was plenty of content. It was intent (what is this work for?) and coherence (five artifacts, three names). This site, and the decision to call the whole thing , is the coherence fix. If the method cannot diagnose its own author, it should not be sold to anyone else.
§3 built_by_an_agent
Built by an agent
This site was designed and built by an AI agent (Claude, working in Claude Code). It worked from written briefs (a design brief, three competing visual directions of which I chose this one, and a product spec for the nine pages) and from my own source texts. The agent wrote the code, ran its own quality checks, and wrote the report below about where it departed from the spec. The numbers are from its final QA run, not estimates.
Build log
| date | 2026-10-25 |
|---|---|
| pages | 9, sharing one header and footer |
| files | 15 deployable (HTML, one stylesheet, two scripts, one icon, README, .nojekyll) |
| size | 330,726 bytes, fonts excluded (budget 900,000) |
| libraries | 0 frameworks, 0 libraries, 0 external scripts; fonts from Google Fonts |
| fields | 14 stance behaviours (7 from the design concept, 7 new) plus the hero, the essay strip and 14 card fields |
| QA loads | 72 in headless Chromium: console errors 0, failed requests 0, horizontal overflow 0 |
| contrast | 4,437 text elements per theme; lowest 4.98:1 light, 6.07:1 dark |
| diagnostic | 26,880 of 26,880 answer combinations return a result; 32 paths driven through the UI |
| hero field | 1,600 agents at 1280 px, 60 fps in the QA browser (indicative) |
| checks | 97 of 97 passed |
Deviation report
Where this build departed from its spec, and why. Written by the agent at the end of the build.
- Hero button renamed. The home hero’s first button reads “Read the essay” instead of “Read The Calculator Problem”, so the word “calculator” does not appear in the hero (a spec requirement). The essay keeps its title.
- Two row orders. The matrix circuit and table follow MODE_FLUENCY_MATRIX.md (Oracle, Genie, Interpreter, Executor, Miner first); the stances page follows the v3 pantheon order. Which order is canonical is still an open question for Henrique, so each page follows its own source.
- The seven new stance fields are illustrations. Adversary, Librarian, Interpreter, Simulator, Scribe, Translator and Therapist were designed for this build. Their counters (for example “never said 3”) count simulated tokens on screen; they are not statistics about people or models.
- Binding-constraint demo scope. The dependency chain in the Stack demo covers the five cognitive layers only, as the source does. Evaluation and Harness have their own switches, which show their published failure signatures instead of dimming layers.
- Essay strip. The reactive strip sits at the top of the essay and stays pinned under the header while you read, marking reading progress. Nothing animates behind the text.
- Privacy sentence made more exact. Thumbnails are lazy-loaded, so learn.html says they load from YouTube’s image host “when they scroll into view”; the rest of the spec’s sentence is kept.
- Without JavaScript. Every page’s text is present; the instruments (compass, fields, circuits, filters, console) are hidden, and their content is given as text instead: the axis list, the stance cards, the matrix table, and the diagnostic’s questions on paper.
- Fix pass after a blind review. An independent reviewer found 0 high, 4 medium and 14 low issues and no invented content. All 4 medium issues and most low ones were fixed (see _qa/qa-report.md, “Fix pass”); the five stance texts that v3 gives at greater length (Oracle, Genie, Interpreter, Executor, Miner) were restored in full on 2026-10-25.
- Diagnostic logic untouched. The two open questions from the prototype review (the recovery question cannot name Judgment, Coherence or Harness; only a “new” experience answer changes the result) are left as they were, because the spec said to keep the logic.
- About wording. The prototype said Henrique works “as an engineer”; that detail is not in the source notes, so this page says only that Henrique works full time in a field that is not AI.
- Not verified. Safari, Firefox and real screen readers were not tested; frame rates come from headless Chromium and are indicative only. The site’s own protocol audit has not been run yet (see below).
Re-audited with its own protocol
A site that teaches checkpoints and deviation reports should carry its own, and be checked by its own method rather than a separate standard. Each release of this site is to be re-audited with the repository’s Agent Audit Protocol: find the binding layer, fix that first, publish what changed. This first build ships with its QA report and the deviation report above (_qa/qa-report.md); the first protocol audit of the site itself is the next step, and its findings will be published the same way.
§4 license
License
The source repository, hjasanchez/agentic-engineering, splits its license in two. The code (this site, the skills and the scripts) is MIT: use it as you like and keep the notice.
The documents (the PDFs, the protocol files, the matrix, the essays and the text on this site) are CC BY-SA 4.0: reuse and adapt them with credit, and share what you make under the same license. Versions published before 25 October 2026 stay available under GPL-3.0 as released.