ai (愛) means love

New here? Read the manifesto first — four short points that explain the whole idea. Then read: coding, UI, data, ops — or skills for your agent.

Manifesto

Build for the reader that matters. The reader is an agent, not a person. Short modules, explicit I/O, JSON-first data, tables over prose.
Tokens are the budget. Every abstraction, adjective, and clever trick is a token tax on every future read. Plain beats clever.
Humans give verdicts, not line reviews. Working = tests pass. Desirable = screenshot looks right. Nobody needs to admire the source.
Agents come and go; guardrails stay. Switch agents mid-project and nothing changes — a new agent with zero context reads the marker, the map, and the key, then performs identically. Consistency comes from the framework, not from any single agent's memory.

The old way and the new way

Every page on this site makes one argument: the era of software written for humans to read is ending. The old way optimized for human readers and human operators. The new way — agentic ops — optimizes for agent builders and agent operators, with humans judging only outcomes.

The old way — built for humans

Code was written so tired humans could read and manage it. One big framework did everything: Django, Rails, Laravel — one way of doing things, one manual to memorize. Interfaces were hand-built, pixel by pixel. Systems shipped by hand: SSH, console clicks, Friday deploys nobody fully understood. When things broke at 3 a.m., a hero got paged with no notes. Audits were screenshots taken the week before.

The new way — agentic ops

AI writes most of the code now, so code is written for agents to read: explicit, grouped by feature, managed through skills and context. The engineer is the driver — they point, the AI builds.

Assembly over artwork. Interfaces assemble from stocked pieces; humans judge the screenshot.
Rails, not heroics. Systems ship through gated pipelines the agent runs: previewed, health-checked, reversible with one action.
Machine-readable ops. Pages carry runbooks and trigger events, so agents acknowledge and roll back without parsing human text.
Continuous proof. Evidence collects itself. Humans verify five things: tests pass, the screen looks right, the dashboard is green, the drill record exists, and a named human owns the blast radius — and nothing else.

Industry zombies shamble on: MVC, MVP, and layered architectures were fences so human teams could share code without colliding. One agent doesn't need the fences — it sees the whole flow at once. But teams of agents still need to stay out of each other's way, so the contract replaces the layer: shared models, tests at the seams, features grouped together. Write code that lets agent teams work together, not human teams.

Proof, measured

A typical task pack reads about 2.5 to 3.6 thousand tokens — 12 to 18 percent of the 21.0k-token bank — because the agent fetches matched skills only, never the whole bank. Counted by our own self-test (script/bench.py), same method as the card badges. The full with/without trial — matched pack versus whole bank versus no skills, scored on tokens, wall time, and defects — is the verify-benchmark protocol: run it, publish method plus raw numbers, or the claim is marketing.

First trial, run November 2026: three design tasks (webhook receiver, evolution-proof validation, stateful rollout) × bare versus skill-fed agents. Defects: 0 and 0 on all three — the model already knows industry doctrine, so skills didn't change correctness here. Outputs ran 136–289 tokens in both arms. What the pack arms added was completeness: freshness SLAs, external probes, deploy receipts — house specifics, not textbook recall. The trial that should discriminate is a coordination-heavy build, not doctrine recall; that run is open. Whole-bank arm not run: feeding all 102 skills to test recall nobody disputes would be cost without a question.

👉 Get the full 116-skill bank — everything above, as instructions your agent installs.