research
Classifies intent, then queries the wiki or ingests new sources, deduping and extending what's already there. The compounding core, vibe coded and then versioned.
Show Us Your Agent Skills / EP 03 / guest dossier
PAUL IUSZTIN
DECODING AI
RESEARCH SKILL
SECOND BRAIN
8 TERMINALS
PAUL IUSZTIN
Paul builds a second brain that writes with him. A personalized, LLM-curated wiki over his own sources, Obsidian notes, Readwise highlights, GitHub repos, that compounds across every research run, then feeds a pipeline that drafts pieces for Decoding AI, his AI magazine, in his voice. He keeps the argument, the order, and the final edits; the agents gather, distill, plan, and draft. Leave them too much room and the generic AI voice pops out, so he trains the system on the diff between his edits and the draft, like a loss function for his own skills.
WIKIa compounding, LLM-curated wiki over his own second brain
SCRATCHbuilt from scratch so it fits his own processes
TRUSTEDreadwise and rss feeds as high-signal sources, not the open web
OUTLINEhe still owns the argument, the order, and the point of view
DISTILLdistills the wiki to each piece, never dumps the whole thing in
READlearns coding-agent architecture by reading the systems directly
EP 03 · PAUL IUSZTIN · the research wiki, the writing pipeline, and a rogue live demo
"You shouldn't fight LLMs. You should find new ways to get into some symbiosis with them."
His answer to brittle agent behavior: adapt the workflow, don't treat the model as an adversary. The agents research, distill, plan, and draft; he keeps taste and the article's argument. 02:18:33
THE WIKI THAT COMPOUNDS
the research skill, his entry point for ingesting and querying it
Paul built a custom knowledge base modeled on Karpathy's LLM Knowledge Base, on top of his own second brain. The research skill is the single entry point: it classifies what he is asking for, then queries what he already has or ingests something new. Each run deduplicates against and extends what came before, so research compounds across sessions instead of being thrown away after one answer.
It mines sources he trusts, an Obsidian vault, Readwise highlights, NotebookLM collections, GitHub repos, and keeps a wiki layer over them: per-source pages, entities, concepts, comparisons, open questions, and contradictions, each linked back to the raw source. His running example is coding-agent architecture: he ingested Claude Code, OpenCode, and Pi to read their memory, permission, and sandbox systems side by side.
"If you leave them too much interpretation and gaps, the LLM and AI voice immediately pops out."
Coding tolerates autonomy once the architecture is pinned down; writing breaks differently, so the writing skills carry far tighter guidance. 02:11:31
FROM RESEARCH TO DRAFT
his writing pipeline, and the loop that keeps it in his voice
Distill the wiki to the pieceHe never hands the whole wiki to the writing agent. A distillation skill pulls only the ideas that match the article sketch: "I compiled it into a very distilled version of my research relevant to what I care about." 02:21:35
Sketch the outline himselfFifteen minutes dumping his own ideas in order: the argument, the teaching goal, the narrative, the problem, the solution. The outline is a set of references for later retrieval, not prose. 02:24:32
Compile a machine-readable guidelineThe article guideline create skill turns the outline plus the queryable wiki into a plan: the what, why, and who, theory-to-practice ratios, point of view, sections. "The intent of this is not necessarily to be human-readable. It is to be machine-readable." 02:26:14
Draft it in his voiceThe article create skill compiles the guideline into prose through an MCP server and his own marketplace of writing plugins. "It cares a lot on how it sounds, to follow my voice on how I would do things." 02:27:22
Update the profiles from the diffHe keeps the agent's draft, edits it, diffs the two, and asks an LLM to turn the difference into profile changes. The writing profiles are living instructions, retuned from his own edits. 02:31:16
THE SKILLS HE SHOWED
one suite around research, writing, and keeping both current. only research is packaged; the rest he runs locally
research distillation
Pulls only the ideas that match the article sketch and compiles them into a small research packet, so the writing agent never sees the whole wiki.
wiki introspection
Inspects the wiki for repeated or clashing ideas, and pulls more out of the research base when a topic is still thin.
render
Renders visualizations over the research base. Shown briefly: the capability he's explored least so far.
article guideline create
Turns his outline and the queryable wiki into a machine-readable plan: metadata, sections, point of view, theory-to-practice ratios.
article create
Compiles the guideline into a prose draft in his voice, calling his writing tools through an MCP server and plugin marketplace.
profile update
Diffs his edited article against the agent's draft and proposes updates to his writing profiles, so the system tracks his current taste.
"It's similar to a loss function: this is how I want it to look, and this is the bad version. Do a diff between them and find what signal I can use to improve my skills instead of the model's weight."
Models and harnesses keep changing, so he tunes his own skills from his edits instead of revisiting every rule by hand. 02:32:28
THE SECOND-BRAIN STACK
his rule: Obsidian plus an IDE you love; the rest is sources he trusts
Zedthe snappy editor he codes and writes in
Obsidianthe durable second brain, ingested via its cli
Readwisecapture layer: highlights and notes he trusts
NotebookLMdeep research on resources he doesn't know yet
Claude Codethe agent he runs, and one he studies
Decoding AIthe magazine the whole pipeline feeds