Long-form, sourced, and written in plain terms before the technical ones — because skepticism is not an obstacle to understanding AI; it is a requirement of it.
These guides are researched and drafted by Urania — an AI system, disclosed and bounded — and edited, verified, and signed by Zach Rossmiller. Don’t cite Urania or these guides; cite the primary sources in each guide’s colophon.
The italic date beside each title is a revision stamp, not a publication date. When the ground shifts, the guide is revised in place and the date moves.
The instructions you wrote for last year's model are still being read by this year's.
A vendor deleted over 80% of its own system prompt and reported no measurable loss. This guide separates the two distinct costs of a bloated context, works through the six reversals behind the deletion, maps what belongs in each layer, and gives you an audit you can run by hand — including an honest account of what subtraction breaks and what the evidence does not yet settle.
A loop is a graph with one node and an edge back to itself. Everything past that is what you build once one loop stops being enough.
In one July 2026 weekend, six words from a developer's tweet became a same-day obituary for 'loop engineering,' a wave of courses, and a viral statistic that turned out not to exist. This guide separates the real design step — wiring specialized agent loops into nodes, edges, and shared state — from the rebrand, walks LangGraph, Google's ADK, and Microsoft's AutoGen side by side with a worked example, and gives the plain test for when a graph earns its complexity and when a loop is still enough.
Both runtimes are self-hosted. Only one of them lets the agent grade itself — and neither lets you keep the thinking at home.
The two dominant open-source personal-agent runtimes, read as architecture rather than a scoreboard: where the skills come from, what the gateway exposes, what the marketplace ships, and who says no when the agent declares itself finished. Ends with a hardened install you can actually run.
Five advisors, one model, one set of blind spots — agreement is not the same as confirmation.
The 'AI council' prompt asks one model to answer as five advisors who debate, then deliver a verdict. This guide maps the whole design space — from one model in five hats to genuinely independent models with a check that can say no — so you can build the right version for a task and know exactly what its agreement is worth.
NotebookLM gives you the same two hosts forever; own the pipeline and every episode can sound like its own show.
Build a podcast pipeline from two models doing two jobs — Claude drafting the personas and a tagged two-host script, ElevenLabs voicing it — so you own the cast, the script, the length, the edit, and the file, where the vendor box fixes all five. With persona-first writing, model and voice choices, per-clip synthesis and stitching for long-form episodes, the short-form Text to Dialogue call and its limits, an honest credit-cost accounting, and Podcastfy as the open-source implementation of the whole path.
Own the model, own the tools, own the stack. In images, audio, and embeddings, open runs lighter than the coding models — and stands at the frontier.
Unit three of the local-models cluster: open models for image generation, speech in and out, and the embedding layer behind local RAG — with the license traps that bite hardest in media, the data-sovereignty payoff of a private knowledge layer, and an honest account of the one premium ceiling closed still holds.
Own the model, then own the tools. A self-hosted coding agent keeps your source on your machine — and the open stack is closer to the frontier than most people think.
Unit two of the local-models cluster: the open coding models and how to read their benchmarks, the open agents that drive them, the fully-local stack where your code never leaves the box, and an honest account of where open still trails the closed frontier.
The frontier is roughly four months from your own hardware. Owning the model — not renting it — is the sovereign move.
Unit one of the local-models cluster: what 'open' actually buys you (open-weight vs open-source, the licenses that bite, the closing gap), how a model is shaped (parameters, quantization, context, temperature), and why running it locally is a question of authority, not just cost.
Own your knowledge in files you can read and ship — instead of renting it inside a vendor's catalog.
Google's Open Knowledge Format formalizes the LLM-wiki pattern into a portable bundle of Markdown files. What the spec actually requires (one field), how to build and validate a conformant bundle, the own-versus-rent argument for keeping your knowledge layer in version control, where OKF sits beside MCP and RAG, and an honest account of what its minimalism leaves unsolved.
A prompt says what you want, once. A skill says how to do it, every time — and travels to any agent that reads the open format.
A skill is a folder that teaches an agent how to do one job the right way — and since December 2025, an open format that runs unchanged across Claude, Codex, Gemini, and Copilot. Build one from an empty folder: the SKILL.md, progressive disclosure, the description that does the triggering, when to split a reference file, and the decision rule for when a skill beats a prompt or a server.
AI levels some fields and steepens others. The difference is who has to supply the judgment.
Two careful bodies of research on AI at work reach opposite conclusions — compression that lifts novices, and returns that reward expertise. The reconciliation is one question you can ask of any task: who supplies the judgment? With the jagged frontier, the verification burden, persuasion bombing, and an honest account of what we still cannot say.
The through-line is deletion: instructions written for weaker models now work against this one.
A working adaptation of Anthropic's official guide — the behavioral deltas from Opus 4.8, effort calibration, long runs and their edge cases, memory, and the drop-in blocks worth keeping verbatim.
The leverage point moved: stop prompting agents, and design the systems that prompt them.
From the viral June 2026 discourse to a buildable discipline — the lineage, the loop contract, the six building blocks, independent verification, memory, hard stops, and the three risks that sharpen as loops improve.