# Robert Evans — Specs, Tools, and Concepts

> CTO and systems builder publishing spec pages, tools, and architectural concepts at robertrevans.com. Each page is built on The Point CSS framework and optimized for both traditional SEO and AI discoverability.

## Specs
- [ActiveAI](https://robertrevans.com/specs/active-ai/): A Rails-native Ruby gem that gives LLM calls a declared, testable home — with generators, config/ai.yml conventions, a built-in agentic loop, multi-agent Workflows and LLM-routed Orchestrators, Skills as reusable system-prompt instruction blocks, a 4-stage memory pipeline, and full ActiveSupport::Notifications instrumentation (10+ events, zero configuration). Built by Robert Evans; extracted from writer-v3, a production writing application. Covers: what ActiveAI is (what-is-this), the problem it solves (LLM calls scattered without convention), how the agentic loop works (stream → tool dispatch → continue → final response), Workflow vs. Orchestrator routing, Skills definition, instrumentation events, how ActiveAI differs from langchainrb and ruby_llm across generators/config/loop/workflows/instrumentation, and three-step install instructions.
- [LLM Failure Modes](https://robertrevans.com/specs/failure-modes/): 22 LLM agent failure modes, each with a drop-in fix prompt. Covers reasoning failures (fabrication, premise acceptance, premature closure), instruction failures (context dilution, scope creep, silent assumption-filling), output integrity failures (placeholder syndrome, claimed execution, sycophancy), and agentic execution failures (prompt injection, test gaming, goal drift, error cascade). By Robert Evans, The AI Design Architect.
- [LLM Failure Modes — Full Catalog (markdown)](https://robertrevans.com/specs/failure-modes/llm-failure-modes.md): All 22 failure modes with descriptions, mechanisms, use-with tags, and copy-paste fix prompts. Machine-readable version — prefer for ingestion.
- [LLM Failure Modes — Full Catalog (web)](https://robertrevans.com/specs/failure-modes/llm-failure-modes-catalog.md): Human version with filtering by work type.
- [The Roster](https://robertrevans.com/specs/the-roster/): The actual scale of a production multi-agent system Robert built: 101 agents across 11 teams, roughly 20 disciplines, and one conductor above all of them. Covers: the teams, the disciplines, the conductor's role, and direct answers to what the system does at this scale.
- [The Depth Gauge](https://robertrevans.com/specs/the-depth-gauge/): Eleven agent teams, sounded three layers deep — the contract the conductor gets, the real headcounts and skills behind each team, and how the pieces actually move, without opening the box further than that. Covers: eleven sealed cards with four facts each, how deep the conductor is allowed to see, why the reveal deliberately stops there, and which rungs of the system are climbed, built-and-quiet, or left untouched.
- [The Handoff Graph](https://robertrevans.com/specs/the-handoff-graph/): Why this agent system isn't a pipeline — teams declare only what they consume and what they produce, and the conductor infers a run order from those declarations live, every time, rather than following a hardcoded step sequence. Covers: the pipeline model people assume versus the graph actually running, what each team needs and hands back, five real execution shapes inside eleven team-boxes, parallel execution between unrelated teams, and what adding a twelfth team actually requires.
- [The Conductor Pattern](https://robertrevans.com/specs/the-conductor-pattern/): How a real multi-agent system coordinates eleven independent teams with one conductor that never reads their internals — trust is built by refusing to look, not by inspecting. Covers: sequencing versus supervising, knowing a team by only two things, why a straight line breaks, and the steady pulse the system remembers across runs.
- [Field Report](https://robertrevans.com/specs/field-report/): What a real production multi-agent system's actual failure taught, with the specifics redacted — proof of operating depth, not a blueprint to copy. Covers: a review process that quietly stopped doing its job, building two things while refusing to build a third, why the same finding twice is a different problem than once, and what real operating experience looks like on paper.
- [The Post-Office Pattern](https://robertrevans.com/specs/agentic-post-office/): A centralized mail-routing pattern for multi-agent systems — agents drop messages in one shared outbox, a shell script (not an agent) routes them to inboxes, so scripts handle mechanics and agents handle judgment. Covers: why the outbox is the one decision that makes it work, why the postmaster is a script rather than an agent, the registry as wiring diagram, per-message audit trails, and how nothing gets silently lost.
- [Why LLMs Fail: The 5 Root Causes](https://robertrevans.com/specs/root-causes/): The 22 named LLM agent failure modes aren't 22 separate problems — they trace back to just 5 root mechanisms in how the model works, so guarding the root covers the whole group. Covers: generation over ground truth, no native calibration, attention economics, no instruction/data boundary, and trained-in pressure, plus how each root maps to its symptom cluster.
- [The Memory Ladder](https://robertrevans.com/specs/memory-ladder/): Four rungs of AI/agent memory — files, a database, RAG, or a knowledge graph — each with a climb trigger that tells you when the rung below has failed, so complexity is earned, not adopted upfront. Covers: the four rungs, the climb trigger for each, why the top of the ladder splits into RAG/Graph vs. Graph/RAG variants, and how all four rungs coexist in one real system.
- [Agent Architecture](https://robertrevans.com/specs/agent-architecture/): Every multi-agent system is secretly organized as hub-spoke, mesh, or hierarchy — each shape has a name and a predictable way it breaks, and most teams pick one by accident rather than on purpose. Covers: hub-spoke (one agent runs the room), mesh (nobody's in charge until it isn't), hierarchy (the org chart you meant to draw), and how to notice when your system's topology has stopped fitting.
- [The Anatomy of an Agent Prompt](https://robertrevans.com/specs/prompt-anatomy/): Most agent prompts read like job descriptions when they need to read like behavior specs under pressure — five parts most prompts are missing at least three of: identity, boundaries, procedure, voice, and self-report. Covers: identity built from metaphor not job title, boundaries spelled out rather than implied, procedure as real steps not a vibe, voice shown by example not described as an adjective, and structured self-report versus free-form reflection.
- [Stop Listening to the Edges](https://robertrevans.com/specs/edges-are-noise/): Both "AI does the thinking for you" and "AI does no thinking for you" are absolutist takes, and both are wrong — the real signal lives in the conversation between them, not at either edge. Covers: what each edge gets wrong, why the middle is a conversation rather than a verdict, and a one-question filter for spotting edge noise in AI discourse.
- [The Black Box of Averages](https://robertrevans.com/specs/black-box-of-averages/): A system prompt isn't a description of a job, it's a selection — a choice about which narrow region of the training distribution the model reasons from. Nine real agent definitions show what each part of that choice actually overrides. Covers: the average as a place you can name, role selecting everything versus occupation selecting one thing, why "helpful" needs edges written down, giving disagreement somewhere to go, deciding in advance what happens when a prompt runs out, and the loop underneath all six parts.

## Tools
- [RaySpec](https://robertrevans.com/plugins/rayspec/): A Claude Code plugin built by Robert Evans that takes any idea, brief, or URL and runs it through a seven-agent Ray Pattern pipeline — Discovery, Blueprint, Build, Code Review, SEO, AI SEO, and Deploy — to produce a complete, deployed HTML spec page built on The Point CSS framework. Each agent has one job and produces a reviewed artifact the next agent reads; the pipeline is not a single large prompt. Covers: what RaySpec is, how the seven-agent pipeline works, what each agent does, how RaySpec handles both traditional SEO and AI discoverability (answer density, entity clarity, AP reading layer), how The Point CSS framework integrates, and how RaySpec differs from single-prompt generators like MindStudio and Mintlify.
- [The Point](https://robertrevans.com/tools/the-point/): A CSS framework built by Robert Evans specifically for AI coding assistants — 99.7% token coverage with native cascade layers, so an AI reading a single reference file (llm.md) builds consistent UI without inventing new class names. Covers: why utility classes have no memory but the AI reading the reference does, how base classes are already design tokens that layers make safe to override, building the same component two ways, and the "build first, brand once" workflow.
- [Rails Agentic Engineering Team](https://robertrevans.com/tools/rails-agentic-engineering-team/): A pattern by Robert Evans for a pipeline of 10 specialized LLM agents, each with one job and defined inputs/outputs, that hand artifacts through fixed stages, log every struggle to a SQLite flight recorder, and improve their own rules over time. Covers: the city-planning-department analogy for why nothing advances until the prior artifact exists, the self-improving loop, the parallel review stage, and the human approval gate.
- [Simple Context Memory](https://robertrevans.com/plugins/simple-context-memory/): A Claude Code plugin built by Robert Evans that fixes session amnesia — three slash commands (/opening, /closing, /report) implement a surgical handoff system based on the SBAR-C nursing handoff framework, so no session starts cold and no decision gets re-litigated from scratch. Covers: the archaeology document, /opening modes, accumulated state across sessions, the three-hook layered defense against context loss, and 30-second zero-config install.

## Concepts
- [AutoFry Pro](https://robertrevans.com/concepts/autofry-pro/): A concept by Robert Evans for a fully closed-loop autonomous commercial fryer system — both oil tanks live outside the building, with one truck visit refilling clean oil and pumping out waste on a schedule the outdoor tank sets itself, so nothing inside needs touching except filter cartridges. Covers: continuous auto-fill, quality-triggered oil swap, 3-stage filtration, full system specs, nine revenue streams, revenue projections and unit economics, and the case for a $100M+ exit.

## Games
- [The Arcade on the Third Floor](https://robertrevans.com/games/): The personal story behind three playable arcade rebuilds — Robert's parents ran a conference center with a small arcade on the third floor, where his dad would open the Gorf, Defender, and Asteroids cabinets to give him free credits as a kid; he's now rebuilt all three as framework-free HTML5 Canvas/vanilla-JS games to share with his own kids. Covers: Asteroids' vector-line rendering, Defender's scrolling world and radar, Gorf's five mission modules under one cabinet, and the shared three-letter high-score dial-in screen. Links to all three playable games.