how this website was built

the meta case study: ai-native creation from research notes to deployed site

prerequisites process automation tools principles verdict

tldr

initial theory

can you build a useful, authentic portfolio website entirely through ai-assisted workflows, sourced from structured research notes?

hypothesis

a 55-file karpathy-style research wiki, combined with claude code and openai codex, can produce a portfolio site that is more honest and more information-dense than anything hand-written from memory

maintenance

scheduled cowork agent runs every sunday, drafts a privacy-filtered update proposal, never auto-commits

status

shipped you're looking at the output

skills

ai-native workflowsclaude codeopenai codexkarpathy wiki patternprivacy-first automationvercel deploymentinformation architecture

studio blueprint

research wiki

→

extract + structure

→

generate ui + copy

→

review + tighten

deploy + test

→

feedback from recruiters

→

iterate in public

prerequisites

structured wiki: 55-file markdown wiki following the karpathy pattern, with one canonical file per topic, changelogs, cross-references, and a CLAUDE.md schema
schema for agents: CLAUDE.md defines operations (ingest, query, lint), page conventions, folder structure, and a privacy model so agents know what they can and cannot surface
eval outputs: every venture and research thread has a structured evaluation with kill criteria, honest self-assessments, and verdict bars
honesty as a constraint: the site includes paused and deprioritized projects alongside shipped ones, not just a highlight reel

process

phase	what happened	tool
context	loaded the full wiki into claude code sessions, established the design guide and voice guidelines	claude code
research	reviewed all 55 wiki files for public-safe content, mapped case studies to visual structures	claude code + manual review
design	created design-guide.md during a cowork session: color palette, component patterns, content rules, voice guidelines	claude code cowork
extraction	pulled structured data from wiki into TLDR tables, metrics, flow diagrams, and honest self-assessments	claude code
generation	generated the full site: HTML, CSS, and all case study content from extracted wiki data	claude code + openai codex
review	manual review of every section for accuracy, privacy compliance, and tone. tightened copy, removed emdashes, lowercased headers	manual
meta docs	wrote design-guide.md, portfolio-full.md, and qa-facts.md as structured reference files for the chat panel and the self-update agent	claude code

self-update agent

how it maintains itself

schedule

every sunday at 6pm local time (cron 0 18 * * 0)

runtime

claude agent sdk on claude opus 4.6

what it does

reads wiki index + log, identifies pages changed in the last 7 days, diffs against site content, runs a privacy filter, writes one proposal file

deliverable

exactly one markdown file at proposed-updates/YYYY-MM-DD.md

autonomy

propose only, never merge. no git commits, no vercel deploys, no direct edits to source files

wiki index + log

→

diff against site

→

read changed pages

→

privacy filter

write proposal

→

manual review

→

git + vercel

section	contents
what's new in the wiki	every wiki page whose changelog was touched in the last 7 days, summarized
proposed edits	concrete bullet or table changes that passed the privacy allowlist
flagged for manual review	edits the agent is not sure about, with the uncertainty noted
audit trail	anything deliberately stripped, with the denylist rule that fired

pii methodology: deny-by-default allowlist

governing heuristic: would ayush be comfortable if the source of the insight read it on the public site? if yes, allow. if unsure, flag. uncertainty always routes to manual review.

#	allow rule	description
1	already public	the same fact is already on the site and just needs a refresh
2	public source	the wiki page cites a published article, press release, SEC filing, or named dataset
3	historical / frozen	content about HOP (shipped), curinos (former employer, frozen data), or public biographical context
4	public research pattern	technical design, methodology, or frameworks that describe how ayush thinks rather than who he is talking to

deny rule	always stripped or flagged
megaeth internal strategy	token plans, treasury, roadmap, unannounced partnerships
active deal or partner names	any named ongoing negotiation or counterparty
non-public pricing	unit economics pulled from private conversations
internal treasury figures	any treasury or financial data not publicly disclosed
curinos client IDs	client or bank identifications from former employer
personal contact info	emails, phone numbers, anything doxxable

primary research (trickiest category): business-neutral, non-confidential insights are allowed when framed as generic patterns without attribution. quoted sentiments tied to identifiable roles, reverse-identifiable anonymous attributions, and active-counterparty quotes are always stripped.

case study	default mode	special rules
HOP Network	allow-by-default	frozen and shipped, no active data
Cloud Brain	allow-by-default (technical)	flags specific megaeth integration numbers and contributor names
Maritime	allow-by-default (public competitive)	strips any quote tied to an interviewed executive or insider-access ship management detail

capability	status
write proposal markdown	allowed
read wiki index + log	allowed
read wiki raw/ folder	hard-prohibited
git commit	not allowed
vercel deploy	not allowed
edit source files	not allowed
browser control / web scraping	not allowed

consistent rejections: tighten the allowlist. a rule that lets too much through is too loose
consistent manual additions: loosen the denylist. if ayush keeps adding facts the agent did not surface, the filter is overclassifying
task prompt is editable in-place in the cowork scheduled-tasks system. already revised once to loosen the primary-research rule
revision rule: never relax a privacy guardrail without being able to articulate why the original concern was wrong

tools

tool	role	notes
claude code	primary generation + review	used for design system creation, content extraction, site generation, and ongoing maintenance via the self-update agent
openai codex	second-opinion generation	cross-model research and verification, alternative drafts for copy and structure
55-file karpathy wiki	source of truth	structured markdown knowledge base with changelogs, cross-references, and CLAUDE.md schema
vercel	deployment + hosting	single-command deploy, preview URLs for review before publish
next.js	framework	app router, server components, static generation for performance
cowork agent sdk	weekly automation	claude agent sdk on opus 4.6, scheduled via cron for the self-update agent
manual review	quality gate	every generated output reviewed for accuracy, privacy, and tone before publish

design principles

no paragraphs: diagrams, tables, and bullets only. dense visual layout. paragraphs hide information density
no emdashes:use "not" or commas instead. consistent punctuation across the site
lowercase headers:consistent voice, less formal. "the problem" not "The Problem"
status pills for status: visual parsing is faster than prose. green/yellow/red only
honest tables for assessment: every case study includes an honest self-assessment table with strong, adequate, and weak scores
themed backgrounds: each case study has a unique background treatment (meta grid, hop glow, cloud drift, ocean waves)
ai disclosure always present: honest about what ai was used for. not apologetic, just factual

if you want to build similar

start with the wiki, not the site: the research notes are the product. the website is just a view layer on top
write a CLAUDE.md schema: tell agents how to read your wiki: what operations are allowed, what the folder structure means, what the privacy model is
include failures: paused and killed projects build more credibility than a highlight reel. recruiters notice
automate the boring parts: the self-update agent handles diffing and privacy filtering. you handle the 10-minute review
ship fast, iterate in public: the first version of this site was generated in a single session. everything since has been incremental refinement

verdict

verdict: shipped. ai-native workflows can produce an honest, dense portfolio site sourced entirely from structured research notes. the weekly self-update agent keeps it current without compromising privacy.