how i built this

how structured research notes became a deployed site, mostly through ai tools

download design-guide.md

tldr

initial theory

can a portfolio site be built mostly through ai tools, sourced from structured research notes?

answer

yes, but only because the wiki was already structured. the schema does the work.

status

shipped you are looking at the output

build loop

wiki

→

extract

→

generate

→

review

→

deploy

→

feedback

→

iterate

the self-update agent

every sunday at 6pm, a scheduled agent reads the wiki, diffs it against the live site, and drafts one proposal. it has zero write access. i review in 10 minutes and publish or reject.


        ┌─── wiki (55 files) ───┐
        │                       ▼
   rejections            agent diffs wiki
   tighten rules         vs. live site
        ▲                       │
        │                       ▼
   human reviews  ◄───  agent drafts
   accepts/rejects       one proposal

privacy: deny-by-default

content encountered

on the allow list?yes, publish

no:

proper name? denyrevenue figure? denyprivate conversation? denyinternal strategy? denyeverything else? deny (default)

design principles

no paragraphs. tables, cards, timelines, diagrams.
honest self-assessment on everything. green/yellow/red, no spin.
propose, don't merge. the ai agent has zero write access.

phase	what happened	tool
context	loaded 55-file wiki into claude code	claude code
research	identified public-safe content across all wiki files	claude code + manual
design	created design-guide.md: palette, components, voice rules	claude code
extraction	pulled data into TLDR tables, metrics, flow diagrams	claude code
generation	generated the full site from extracted wiki data	claude code + openai codex
review	manual review for accuracy, privacy, and tone	manual
meta docs	wrote design-guide.md, portfolio-full.md, qa-facts.md	claude code

how it maintains itself

schedule

every sunday at 6pm local

runtime

claude agent sdk on opus 4.6

deliverable

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

autonomy

propose only. no git, no vercel, no source edits.

capability	status
write proposal markdown	allowed
read wiki index + log	allowed
read wiki raw/ folder	prohibited
git commit / vercel deploy	not allowed
edit source files	not allowed

#	allow rule	description
1	already public	same fact already on the site, just needs refresh
2	public source	wiki cites a published article, filing, or dataset
3	historical / frozen	HOP (shipped), curinos (former employer), public bio
4	public research pattern	methodology and frameworks, not who he talked to

deny rule	always stripped
megaeth internal strategy	internal plans, treasury, roadmap
active deals	named ongoing negotiations
non-public pricing	unit economics from private conversations
curinos client IDs	former employer confidentiality
personal contact info	emails, phone numbers

case study	default	special rules
HOP	allow-by-default	frozen and shipped, no active data
Cloud Brain	allow (technical)	flags megaeth integration numbers
Maritime	allow (public competitive)	strips interviewed-executive quotes

tool	role
claude code	primary generation, review, and maintenance agent
openai codex	second-opinion generation and cross-model verification
55-file karpathy wiki	source of truth with CLAUDE.md schema
vercel + next.js	deployment and framework
cowork agent sdk	weekly scheduled automation (cron)

start with the wiki, not the site. the research notes are the product.
write a CLAUDE.md schema. tell agents how to read your wiki.
include failures. paused and killed projects are more credible than a highlight reel.
automate the boring parts. the agent handles diffing and filtering. you handle the 10-minute review.
ship fast, iterate in public. the first version was one session. everything since is incremental.

verdict

ai tools plus structured research notes produced this site. the wiki is the product. the site is a view layer.