Every tool in this category hands you a template. This one is 6,600 lines of refusal.

It turns a person into a portfolio that is live: interviewed, positioned, designed, written, built, and deployed to infrastructure they picked. There is nothing to install. The markdown is the product and your agent is the runtime.

The rules it will not bend

Five of the twenty-five carry a lock. They are integrity rather than quality, so they hold even when the human asks otherwise. The rest bend on request.

Every claim traces to evidence. A portfolio is a claim about a person, and a fabricated metric is a lie told on their behalf to someone making a hiring decision. No source, no ship.

Accessibility basics. Alt text, contrast, keyboard, and a designed reduced-motion state rather than animation: none on a layout that assumed movement.

Deploy is reversible. Snapshot before overwriting. The rollback command goes in the report, because the person who needs it is reading at speed at a bad moment.

Gate C. Nothing becomes publicly visible under someone's name before they have seen it. Not for time pressure, not because they said "just ship it" three loops ago.

Identity safety. Push and deploy as the subject, with their credentials, then switch back. A commit credited to the wrong person is not cosmetic.

The other twenty, at a cost stated once and written into the run:

The pipeline

Loop 1 decides the rest. The interview is the only source of the story, the voice, and the strategy, and every later loop reads what it produced. Run it thin and you get eight thin loops — no model in Loop 4 repairs a shallow Loop 1.

That is also why the cheap-early instinct is wrong here. What is genuinely mechanical at the start is the scrape, and that runs at the bottom tier in parallel while the interview happens (MODELS.md).

What it refuses, and why

Refusal	Because
No template	Two runs with the same archetype must not look alike. Every component is reinvented and the run records what it invented
No stock photography	Visuals are generated: CSS, SVG, canvas, WebGL, procedural. Generated imagery gets treated in-browser rather than dropped on the page raw
No badge service	A third-party endpoint going down leaves a broken site forever. Every asset here is a committed SVG, built by scripts/
No CDN in production	Free rein while prototyping. Vendored, committed, and self-hosted before ship
No fabricated metric	Every number appears in EVIDENCE.md with a source, or it is attributed as speech, or it is cut
No silent degradation	Skip whatever you want. The skip and its cost go in SKIPS.md and into the confidence grade
No generated likeness	The one image never synthesized is the subject's own face

The craft

Advanced visual technique is the point. 72 techniques in CRAFT.md — raymarching and SDFs, curl-noise flow fields, reaction-diffusion, GPGPU particles, post-processing, gooey SVG filters, scroll-driven timelines, View Transitions, spring physics, variable fonts driven across every axis, a designed preloader. Pull any library while prototyping. Look anything up. The list is a floor.

48 styles in STYLES.md, across seven families, and the useful part is not the list. It is the collision: liquid glass over a blueprint substrate, risograph texture on a Swiss grid, terminal chrome around a physics world. One parent carries structure, the other carries surface. Two styles held in tension is where the un-templated work lives, because a single named style is one search away from being everyone else's site.

Three rules keep it honest. Every technique gets a runnable prototype before the design depends on it. Every technique ships three states: full, designed reduced-motion, and a no-WebGL fallback. And if the technique is more memorable than the project it frames, it is the wrong technique.

Deploy

Nine targets, with a subject's own server as a first-class option rather than a fallback: GitHub Pages, Netlify, Vercel, Cloudflare Pages, VPS with nginx, VPS with Docker, cPanel, S3 with CloudFront, and handoff.

The target constrains the stack, never the reverse. Each adapter in deploy/ has the same shape — prerequisites, snapshot, deploy, rollback, DNS and TLS, forms — so reading one teaches all nine.

The doctrine

File	Governs
PRINCIPLES.md	25 binding rules, 5 of them hard
BRAND.md	Archetype, positioning, voice, the translation table
ARCHITECTURE.md	How many pages, routing, and the blog
STYLES.md	What it looks like, and how to collide two
CRAFT.md	How it renders
MODELS.md	Which model tier runs which step
SKIPPING.md	What each step costs to leave out
AGENTS.md	The universal entry point for any agent

Run it

/portfolio.me <name or handle>          # Claude Code

Anywhere else, point your agent at AGENTS.md and name the subject. Configs ship for Cursor, Codex, Gemini, opencode, Cline, and Windsurf. A platform without subagents runs the same loops sequentially.

Two scripts, standard library only, no dependencies:

python3 scripts/preflight.py <dir | file | url>   # the mechanical half of Gate C
python3 scripts/badge.py                          # regenerate the artwork

preflight.py checks what a machine can check: alt text, reduced-motion states, lang, CDN assets, hotlinked stock, trackers, the share layer, the shell budget, and whether every figure on the site appears in EVIDENCE.md. It does not judge design, prose, or truth. That stays the human's call at the gate.

Lineage

Sibling of git-a-profile, which does this for a GitHub README. Same shape, harder problem: a README can be built from an API, and a portfolio cannot. The story only exists in the subject's head, which is why this one opens with an interview and not a scrape.

Built by PIIIX. MIT.