This article has been corrected. 5 corrections — see Corrections at the end of this article.

A Primer for Journalists, Editors, and Writers

Table of Contents

1. At a Glance
2. Glossary
3. The Core Idea
4. The LLM Wiki: Where the Compounding Happens
5. AI Skills as a Virtual Team
6. Rules, Routing, and Guardrails
7. The Virtual Board of Advisors
8. Scoring: Making Trust Auditable
9. Where the Wiki Meets the News Cycle
10. How the System Builds Itself
11. What to Take Away

---

Most people who have used a chatbot have used it wrong — not because they typed the wrong words, but because they treated it as a one-off oracle instead of a system they can shape. You ask a question, you get an answer, you close the tab. The next time you need the same answer, you ask the same question. Nothing compounds.

Working journalists, editors, and reference writers cannot afford this pattern. Their expertise is their compounding knowledge — a decade of sources, a mental map of who lies about what, a feel for which court records matter and which are noise. A personal AI system should behave the same way: every hour of research should make the next hour cheaper.

This document explains how one such system — Patriot University’s — is built, so that readers can borrow the pattern. It is not a product pitch. It is a methodology brief.

In one sentence: a personal knowledge system compounds when a wiki, a set of Skills, a router, and a scored source registry are wired together so that every finished piece of work becomes the foundation the next piece is written against.

At a Glance

Metric	Count
Wiki pages	~1,005
Methodology Skills	~80
Advisor personas	101
Entity-graph nodes	6,789
Entity-graph edges	19,766
Investigative tools cataloged	176
Whitelisted sources	~144
Learning-path guides	21

(As of mid-2026. These are internal architecture numbers, not claims about coverage or quality — cite them, if at all, as “one working example’s scale,” not a benchmark.)

The operations view of the system described in this document — inventory by category and status, and the health of the publishing pipeline. The counts here are live and will not always match the round numbers in the table above, which were current as of mid-2026.

---

Glossary

The words below appear throughout this document. Definitions here are how we use them; other publications may use them differently.

Term	What it means in this document
AI Skill	A durable, version-controlled Markdown file that teaches an AI how to perform one job well — e.g. “verify a photograph,” “build an accountability profile,” “score a source.” Loaded on demand, not on every request. Think of it as a specialist’s checklist you can hand to the model.
Agent	An AI process that runs a task from start to finish using one or more Skills. Where a Skill is a what and how, an Agent is the doing.
Rule	A machine-readable instruction that constrains what the AI may write or how it must behave in a given context (file type, folder, task). Rules are guardrails written in the same repo as the content, so they travel with the work.
Routing	The act of deciding which Skill(s) a task should use. In our system a router reads a plain-language task, matches it against a Skill index, and returns the top matches so the human — or the agent — loads only what is needed.
Guardrail	A rule that prevents a class of bad outputs (e.g. “no political-affiliation labels in a profile without a court record”). Guardrails are enforced by pre-publish checks, not by hope.
Evidence Patch Gate	An automated check that scores a proposed content update before a human ever sees it — rejecting anything with a missing citation, an uncorroborated “documented” claim, or a link to a page that doesn’t exist. The gate doesn’t publish anything; it decides what’s even worth a human’s attention.
LLM Wiki	An idea popularized by Andrej Karpathy: a knowledge base curated by humans and compiled by an LLM at ingestion time, so that facts are already synthesized, cross-linked, and stored — not re-derived from scratch on every query.
Tavily	A web-search API optimized for LLM consumption. Returns clean, ranked, extractable content instead of raw HTML, so a Skill can research a topic without a human having to scrape or copy-paste.
Profile Score / Source Score	A numeric or ordinal rating attached to a source (an outlet, a spokesperson, a public figure, an election threat) along explicit dimensions. Not a single “trust number” — always multi-dimensional and always internal-only unless labelled otherwise.
Whitelist of Sources	A curated, versioned registry of outlets deemed acceptable to cite, with each entry annotated for reliability, bias direction, topic coverage, and last-review date.
Virtual Board of Advisors	A roster of AI personas modeled on named domain experts (scholars, practitioners), selected 1–5 per task by the router with an explicit diversity requirement so that no single ideological lens dominates a piece of writing.
Recursive Building	The property of a knowledge system whose outputs become inputs — every new document is written into the same base the next document is written against, so coverage compounds instead of restarting.
Cohort	A group of wiki pages that tend to need the same research at the same time, because they share the same people, organizations, or events. Computed automatically from the knowledge graph, not hand-curated — so one news event can trigger a single shared research pass across an entire cohort instead of the same search repeated page by page.
Compose	A drafting workflow where the AI writes into a scaffold defined by a Skill, using the Wiki and Tavily as sources, then hands the draft to a human for approval before publish.
Wikilink	The [[slug]] syntax borrowed from Obsidian. In our pipeline, every wikilink is auto-resolved to a live URL at publish time, so cross-references never rot.
Cascade Re-render	The automatic re-rendering of every page that links to a page which just went live for the first time, so those in-text references resolve to a real URL instead of sitting as dead text. What keeps a growing wiki from filling up with broken internal links.
Ego Graph	A per-page slice of the larger knowledge graph — the page’s subject at the center, radiating out to direct and secondary connections. Pre-computed once per page rather than queried live, so a reader sees a person’s or organization’s network without the system running a fresh graph traversal on every page view.

---

The Core Idea

A personal knowledge system for a working writer has four moving parts:

1. A wiki you own — a folder of Markdown files, on your machine, that you edit and control. Not a chatbot’s memory. Not a cloud database.
2. A set of Skills — small text files that teach the AI how to do the recurring jobs of your beat: verify a quote, build a source profile, score a claim, draft a lede in your voice.
3. Rules and routing — instructions that decide when each Skill fires and what it is allowed to say.
4. A source registry with scores — an explicit list of the outlets and databases you trust, with the reasons attached.

Every article you finish should feed all four back. That is what “recursive” means.

The Patriot University knowledgebase is one implementation of this idea — roughly 1,005 wiki pages, about 80 methodology Skills plus 102 advisor personas, some 144 whitelisted sources, and an entity graph of 8,616 nodes and 5,303-plus edges mapping who is connected to whom across every profile. (A second, unrelated graph — Graphify, roughly 32,900 nodes tracing the platform’s own source code — helps developers navigate the codebase. It has nothing to do with what the chatbot knows about the world, and conflating the two is a real mistake worth naming, because we made it ourselves in an earlier draft of this document.) What follows is how those pieces work, and what a smaller newsroom or a solo reporter can borrow.

---

The LLM Wiki: Where the Compounding Happens

The knowledgebase is an Obsidian vault of Markdown files. Nothing exotic — plain text, one file per subject, on disk. Four folders keep the vault from becoming a junk drawer, and none of them are optional:

Layer	What lives there
_sources/	Raw, immutable input — a court filing, a press release, a wire clipping. Read-only once captured. Never published.
_drafts/	Pre-approval drafts, written by a human or an AI. Promoted into the wiki only after review.
_control-center/	The vault’s home base — an index plus an append-only ingest log.
_dashboards/	Status views mirroring publish state, link health, and embedding coverage.

Three more things turn that folder into a system:

1. Provenance. Every publishable file traces back to _sources/. The AI’s job at ingestion is to compile those sources into wiki pages, not to summarize them once and throw them away. A single source typically touches 10–15 wiki pages: an actor’s profile, a timeline entry, a corporate-entity page, a legal-doctrine explainer, and so on. This is the Karpathy “LLM Wiki” pattern — synthesis is paid for at ingestion, not at query time.

2. Cross-linking by wikilink. In-text [[slug]] references resolve to real URLs at publish time. Because slugs are stable and links are resolved late, every new page automatically ties itself into the surrounding graph. The value of the wiki grows super-linearly with the number of pages: each new profile retroactively enriches every older profile that mentions the same LLC, court, or district.

3. A canonical frontmatter contract. Every file starts with the same YAML fields, in the same order: slug, title, category, tags, related, status, audience, last_updated, and a handful more. This is not aesthetic — it’s what makes the corpus machine-readable. The knowledge graph builds itself off frontmatter; the publish pipeline uses it to detect changes; the router uses it to route.

The Golden Rule, and why it’s non-negotiable. Content gets edited in the wiki, or nowhere. Every article that reaches a reader arrives through one pipeline — there is no side door through the content management system. Edit an article directly in the live platform and the next pipeline run silently overwrites the change, because the Markdown file is the source of truth, not the database. That sounds like a limitation. It’s actually the whole point: a single path to publication is what makes the wiki trustworthy enough to write the next hundred articles against.

The practical implication for a solo journalist: your wiki is a database you happen to be able to read as prose. Every article you file should leave one or more wiki pages behind.

The wiki as the writer sees it — one row per Markdown file, with its slug, category, and publish status. This is the “database you can read as prose”: the same files an editor opens in Obsidian are the ones the pipeline publishes.

---

AI Skills as a Virtual Team

An individual reporter cannot hire a photo-verification specialist, a corporate-registry analyst, a First-Amendment lawyer, a voting-rights expert, and a source-credibility auditor. But she can write a Skill for each one, once, and have it on call.

A Skill is a Markdown file with a YAML header describing when it activates, a numbered procedure, and worked examples. It is not code. It is a specialist’s checklist that the model reads before it starts a task.

The Skill library — each card is a single Markdown file teaching the model one job. Nothing here is code; a working journalist could write any one of these in an afternoon.

Skills cluster into families that reinforce one another. Below are the clusters present in Patriot University’s roster; the same shape works for any beat.

Investigation cluster

document-research-specialist · corporate-intelligence-investigator · public-records-research-specialist · network-analysis-specialist · osint-identity-researcher · media-verification-specialist · geolocation-verification-specialist · investigation-workflow-designer (a meta-Skill that chains the others into an “investigation recipe” with named evidence gates between steps).

A reporter asks: “Trace who owns the LLC that got this federal contract.” The workflow designer plans the chain — corporate registry → federal contracts → executive lookups → network map — and hands each step to the right specialist with an evidence-quality threshold (“do not proceed until you have two independent confirmations of beneficial ownership”).

Editorial cluster

patriot-editorial-framework (the house institutional voice) · patriot-opinion-voice (first-person, thesis-first commentary) · compose-ready-drafting (converts research into a publishable draft) · citation-checker (Bluebook plus negative-treatment lookup) · patriot-sanity-check (pre-publish meta-audit) · patriot-glossary-editor.

These act on drafts, not on research. They exist so that voice, formatting, and factual discipline are enforced by the same tool every time — no more asking the model to “please write like our style guide.”

Legal cluster

legal-research-specialist · first-amendment-legal-expert · fourth- / fifth- / sixth- / eighth- / fourteenth-amendment-legal-expert · separation-of-powers-legal-expert · election-law-and-administration · voter-suppression-law · voting-rights-act-expert · appellate-brief-writer · judicial-decision-analytics-specialist.

For a beat that touches courts often, breaking legal expertise into per-doctrine Skills matters: it prevents the model from bluffing on a Fourth Amendment question by conflating it with a Sixth Amendment doctrine it happens to remember more clearly.

Scoring cluster

Four instruments that produce numeric or ordinal scores on explicit dimensions. Treated in detail below.

Voice-and-review cluster

The Board of Advisors (see below) supplies ideological and methodological diversity when writing analysis.

How a Request Finds the Right Skill

At real scale, nobody — human or model — should have to guess which Skill applies. A request passes through two routers. The first is a global entry point that asks one narrow question: does this task belong to this domain at all? If yes, it hands off to a domain-specific router that reads the plain-language task and returns the ranked Skills that actually apply — the same handoff whether the request originates in a shell command or inside the product’s own chatbot. Consistency is the point: a task routed on a laptop and a task routed in production should never disagree about which Skill fires.

The roughly 80 methodology Skills break down across six domains:

Domain	Skills	Examples
Investigative OSINT	22	corporate intelligence, document research, breach-data analysis
Legal & Constitutional	20	First through Fourteenth Amendment specialists, separation of powers
Civic & Democracy	11	election law, civil-resistance theory, democratic-health monitoring
Meta-Verification	10	claims integrity, citation checking, AI inference-boundary review
Platform Tools	8	token efficiency, skill maintenance, compose-ready drafting
Accountability	7	profile verification, the malice evaluator, the private-citizen inclusion gate

Separate from that count sits a roster of 102 advisor personas — described in full below — that the router can also draw into a piece of analysis.

The Skill roster with its token cost per Skill. Cost is why Skills load on demand rather than sitting in every request: the router pulls the two or three a task needs and leaves the rest on disk.

The three key properties of a well-built Skill:

• Loaded on demand. Skills are not in the model’s default context. The router pulls the two or three that a task actually needs. This keeps the model’s attention on relevant material and keeps the cost of a session low.
• Version-controlled. Every Skill lives in git. When a Skill improves, every future article benefits — including retroactively, when you re-audit older work.
• Composable. A meta-Skill can invoke a chain of others. investigation-workflow-designer composes six investigation Skills. accountability-profile-builder composes tavily-research, citation-checker, and patriot-sanity-check.

Editing a Skill and having it actually work are two different events. Skills, advisor voices, and graph context all live locally first; the production system only knows what has actually been copied over to it. Two disciplines keep the two sides honest: a deploy step that pushes an edited Skill to the server, and a drift check that periodically confirms the server’s copy still matches the local one. This is not a hypothetical safeguard — a drift check once caught eight specialist Skills the live system was trying to invoke that had never actually reached the server. Without the check, those calls would have failed quietly, every time, for as long as nobody noticed.

---

Rules, Routing, and Guardrails

Skills are the what. Rules and the router decide the when.

Rules live alongside the work. In our repo they are .mdc files (Cursor’s format) plus equivalents for other tools; a file’s location and frontmatter determine which rules apply. Examples:

• The accountability-profile standards rule fires on any file under profiles/** and forbids language like “criminal” or “traitor” unless a court record supports it.
• The constitutional-law-advisors rule fires on legal analysis and requires at least two ideologically distinct sources — so the AI cannot default to whichever commentator was on cable last night.
• The Obsidian-wiki rule fires everywhere in the vault and enforces the Markdown subset that survives publish (no image embeds, no callouts, no block references).

The critical property of Rules: they are automatic. The writer does not have to remember to apply them. The tool loads them because of where the file sits or what the file’s frontmatter says.

Routing. Given the size of a mature Skill library, no one should have to guess which Skill applies by memory. Our resolver reads a plain-language task and returns a ranked list of Skills with resolved file paths, invoked as a single command with the task in plain English. It answers with the correct three or four Skills to load. The resolver uses the same source-of-truth index the live product uses, so a task routed inside a shell returns the same result as a task routed inside the product’s chatbot.

Guardrails are Rules with teeth. They are enforced by pre-publish checks, not by trust. A group of four Skills — a sanity check, a citation check, a document-analysis pass, and a private-citizen inclusion gate — is designed to run on any draft before it is allowed to be marked ready for publication. That last one is the sharpest instrument: it blocks any profile of a non-official person unless it clears one of five documented anchors — a criminal charge, a documented leadership role in an organization with documented harmful conduct, financial support that materially enabled specific harm, a voluntarily assumed public-organizing role, or misuse of an official capacity. Political affiliation, rally attendance, and social-media posts explicitly do not count. That last sentence isn’t a footnote — it’s the entire reason the gate exists.

A Second Guardrail Layer: The Pipeline Protecting Itself

Everything above governs content — what the AI is allowed to write. A parallel set of guardrails governs the publishing system itself, because a pipeline that pushes hundreds of files to a live site can do real damage to itself if left unsupervised. Ours runs eight of them; the shape generalizes to any automated publishing step:

• An exclusive lock, so two publish runs can never race each other into creating duplicate posts.
• A duplicate-post guard, so a page that already exists gets updated, not cloned, if a run ever tries to create it twice.
• A permalink audit, run after every publish, that catches and fixes the collision-suffixed URLs a content platform silently creates when two slugs almost match.
• A batch-size confirmation gate — any run pushing more than 25 documents stops and asks a human to confirm before it proceeds, because the most expensive mistakes are the ones nobody meant to make at scale.
• A sensitive-category gate, which lists out any accountability or high-risk content in a batch and requires a separate confirmation before it goes live.
• A hash-scheme drift gate, so that when a developer changes what counts as “this file changed,” the system blocks the next run rather than silently re-publishing hundreds of files with no real content changes.

None of these exist because someone imagined a hypothetical. Each one exists because a version of the mistake it prevents happened at least once, and the fix was written into the pipeline instead of into a memo nobody would reread. Guardrails at both layers — content and system — are what let a small newsroom operate at scale without a full-time libel lawyer and a full-time systems administrator reading every draft and every deploy.

The guardrail dashboard — every error and warning raised across the corpus, grouped by the rule that caught it. Errors block a publish; warnings do not. This is what “enforced by pre-publish checks, not by hope” looks like in practice.

---

The Virtual Board of Advisors

An AI trained on the whole internet has read every columnist. Asked for an analysis, it will average them — which usually means it averages loudness, not insight.

The Board of Advisors is a countermeasure: a roster of 102 AI personas modeled on named subject-matter experts — Steven Levitsky and Daniel Ziblatt on democratic erosion; Timothy Snyder on tyranny; Anne Applebaum on transnational autocracy; Erica Chenoweth on nonviolent movements; Ruth Ben-Ghiat on strongman psychology; Kim Lane Scheppele on autocratic legalism; Jack Balkin, William Baude, Michael Luttig, and Mary McCord on constitutional-law questions.

Each advisor is itself a small Skill: what this person’s method is, what texts they’ve written, how they typically frame a problem. When the router determines that a piece needs analysis (as opposed to pure reporting), it selects one advisor as primary — whose full method is loaded — and up to four supporting voices whose “advisory register” is applied. The selection algorithm has an explicit diversity requirement: no five advisors who see the world the same way.

Constitutional-law questions get the deepest bench: 24 advisors, deliberately built across methodological camps rather than around one legal philosophy. Eleven were added specifically to widen that range — Akhil Amar, Michael McConnell, Cass Sunstein, Jack Balkin, Randy Barnett, Noah Feldman, Kermit Roosevelt, William Baude, Keith Whittington, Eugene Volokh, and David French among them — for exactly one reason: so a piece on, say, Section 3 of the Fourteenth Amendment gets written with an originalist, a common-good constitutionalist, and a proceduralist-liberal all in the room, not through one seat.

The journalistic value is not that the AI writes better prose. It is that the analysis section of a piece considers the counter-arguments the reader will bring — because those counter-arguments were in the room when the piece was drafted.

---

Scoring: Making Trust Auditable

“Trust me, this source is good” is not a system. A system says why, on what evidence, updated when.

Underneath every scoring instrument sits a simpler, structural rule: every claim in the corpus carries one of four evidence tiers.

Tier	Definition	How it reads
Documented	Court records, filings, financial disclosures, official government documents	No label needed — the claim stands on its own
Credibly Reported	Multi-outlet journalism with named sources	No label if corroborated by multiple outlets; “According to [outlet]” if single-sourced
Alleged	Single-source, anonymous, or unconfirmed	Prefixed “Allegedly,” or “According to [named source],”
Excluded	Unsourced assertions, AI inference, or a source explicitly kept off the whitelist	Never appears in anything published

A profile with an unresolved [NEEDS VERIFICATION] tag does not go live. The tag has to become a citation, or the claim has to go.

Four scoring instruments run on top of that foundation. They are not the only useful ones, but they illustrate the pattern: numeric or ordinal, multi-dimensional, evidence-anchored, transparent to the writer, and internal-only unless explicitly labelled otherwise.

1. Source Credibility Gate

Applied to: a draft, and the outlets it cites. Dimensions: source mix, attribution, claim support, framing. Scale: 0–100. Green at 85 and above means well-sourced. Yellow, 70–84, means an editor should look again before publishing. Red, below 70, means the draft needs more or better sourcing first.

This is about rigor, not politics — it never scores the point of view, only whether the claims are backed. A sharply critical, adversarial piece scores full marks when it’s well-sourced; a gentle piece with weak sourcing does not. And the publish-blocking version of this gate ships switched off by default, on purpose — it stays off until the trusted-source list behind it has actually been reviewed and rated. A scoring instrument is only as good as the list it scores against; turning on enforcement before that list is solid just enforces noise with more confidence.

2. Democratic Malice Score (DMS)

Applied to: documented actions by public officials. Dimensions: intent, pattern, institutional impact, defamation-law posture. Scale: 1 through 5.

• 1 — no documented malice; ordinary or protected political conduct, however controversial.
• 2 — a minor, documented departure from democratic norms, without evidence of intent to subvert anything.
• 3 — a documented pattern that undermines a specific democratic mechanism (a free press, an independent court, an election).
• 4 — documented, systematic, intentional subversion of multiple mechanisms at once.
• 5 — documented leadership in an attempt to subvert the transfer of power itself.

Runs an Ideology-vs-Malice gate first — ordinary political speech is filtered out before scoring begins, so protected disagreement is never mislabeled as malice. The score is reader-facing but never appears without the underlying evidence tier. A 4 in one of our profiles is a claim, not a slogan, and the reader can see the receipts.

3. Misinformation Score for Spokespersons (MSS)

Applied to: statements by press secretaries and official spokespersons, made in role. Dimensions: falsity of underlying claim, whether the spokesperson repeated it after correction, institutional fact-check coverage. Scale: 1 (reckless mischaracterization) through 5 (systemic disinformation operation). Reference class: a bipartisan set of press secretaries across administrations, used as a comparative baseline so that a “5” means something.

A Tenure-Window Guard excludes pre- and post-role commentary — a person’s TV career before or after their spokesperson tenure is not scoreable under this instrument.

4. Election Threat Score

Applied to: identified threats to a 2026 or 2028 election (actor-level, state-level, federal-level). Dimensions: Likelihood × Impact × Urgency. Scale: 5 tiers, LOW through CRITICAL, mapping to actions (Prevent / Prepare / Mitigate / Respond / Recover).

The output is an editor’s tool: which threats deserve coverage now versus which ones can wait a quarter.

The source registry, in one glance

Behind all four scoring instruments sits a source registry — a YAML file of roughly 144 outlets. Each row has: domain, class (primary-official, institutional-research, mainstream-news, partisan-commentary, tabloid, unreliable), desk, topic tags, a reliability score, bias direction, last-review date, review frequency. This is the whitelist. It is versioned in git, revised as outlets shift, and consulted every time a draft is audited. It is the same document a new hire could read to understand what “cite this, not that” means at this publication.

The pattern to steal: a small YAML file of your regularly-cited sources, annotated. It is a two-hour exercise the first time and worth every minute.

The source registry — one row per outlet, each carrying a factual-reporting score, a bias direction, and a last-reviewed date. This is the whitelist the scoring instruments above are scored against.

---

Where the Wiki Meets the News Cycle

A wiki is static until something changes in the world. The mechanism that closes that gap — the one that actually deposits new reporting back into the base instead of letting the news cycle blow past it — is worth describing in full, because it’s the clearest illustration of “recursive” in the whole system.

The watch. A tracker scans the news on a fixed cycle — ours runs every six hours — scoring what it finds against a fixed taxonomy (twenty authoritarianism markers, in our case, drawn from political-science research on democratic erosion). It searches the open web plus a small curated allow-list of subject-matter domains, because a general search engine and a specialist database surface different things.

The proposal. For every profile the system tracks, a second process runs a five-step chain: it reads the entity graph to get the subject’s name and builds targeted search queries from that; it scans the news, deduplicating anything it has already seen by content hash so the same article never gets analyzed twice; it scores each finding for significance against the accountability framework; for anything significant, it drafts a structured patch — a description of exactly what should be added, with citations and an evidence-tier label already attached; and it runs that patch through an Evidence Patch Gate before a human ever sees it.

The gate. The gate rejects patches automatically for any of four reasons: a missing evidence-tier label, a citation with no URL behind it, a “Documented” claim resting on a single source instead of two independent ones, or an internal link pointing at a page that doesn’t exist. A rejected patch isn’t a final rejection — it can be edited and re-scored, or overridden with an explicit reason — but nothing reaches a human’s desk pre-broken.

The human. Every surviving patch lands in a review queue and waits. Nothing is written to the wiki, and nothing is published, without an explicit human decision — approve, edit, or reject. This is the one place in the whole recursive loop where a person, not a router, makes the call, and it’s deliberately the narrowest point in the pipeline.

The shortcut that isn’t cheating. Profiles cluster. A senator, the PAC that funds her, and the lobbyist who arranged the meeting tend to need the same research at the same time, because a single news event touches all three. The system computes these cohorts automatically from the knowledge graph — nobody curates them by hand — and a single research pass can refresh an entire cohort at once instead of repeating, and re-paying for, the same search three separate times. This is what makes the recursion affordable, not just elegant.

Cohorts, computed from the knowledge graph rather than curated by hand. One news event touching a senator, her PAC, and her lobbyist triggers a single shared research pass across the cohort — not the same search run three times.

---

How the System Builds Itself

Return to the word recursive. Its meaning here is precise.

When a new profile is drafted:

1. The router loads the profile-builder Skill and any relevant advisors.
2. Tavily pulls the latest court records, news, and filings, focused by the Skill’s search patterns.
3. The draft is written into the wiki, cross-linking to every existing person, org, event, doctrine, and financial entity via [[slug]] references.

Drafting inside the system rather than beside it — the AI writes into a scaffold defined by a Skill, with the wiki and web research as its sources, and hands the result to a human before anything publishes.

1. Guardrails run: sanity check, citation check, inclusion gate.
2. On publish, the entity extractor pulls out five kinds of typed entities — people, organizations, events, locations, and financial connections — and adds them as nodes and edges to the graph. A community-detection pass (Leiden by default, sometimes Louvain or Label Propagation depending on the job) regroups the graph into thematic clusters, each with a machine-written summary of who belongs and why. Network-centrality metrics get recomputed so the graph knows not just who is connected, but who is a bridge and who is a hub. Each profile also gets its own ego graph — a small pre-computed slice of the larger graph centered on that one subject, so a reader sees its direct connections without the system running a live traversal on every page view.
3. The next profile drafted, days later, now has this profile as prior art. It cross-links back. Its search queries include tags this profile introduced. The advisors called into service on this profile are pre-selected for related topics.

The system does not stay flat. Every artifact enriches the base against which the next artifact is written. Gaps become visible: a Skill that keeps getting invoked with weak results is a candidate for improvement or replacement. A cluster in the graph with no explainer document is a beat that needs coverage.

The entity graph that step 5 feeds — people, organizations, events, locations, and financial connections, clustered into communities. A dense cluster with no explainer document behind it is a beat that needs coverage; that is what “gaps become visible” means.

This is what a solo journalist gains by adopting the pattern that a large organization cannot easily buy: increasing returns to scale on your own coverage. The hundredth article is written into the shadow of the previous ninety-nine — with those ninety-nine already indexed, cross-linked, scored, and reviewed by the same rules.

---

What to Take Away

For a working journalist, editor, or writer who wants to try this:

1. Start with a wiki, not a chatbot. A folder of Markdown files with slug and title at the top of each is enough. Grow the frontmatter only when you actually need more fields.
2. Write one Skill for the thing you do most often. Verify a photo? Write “verify-a-photo.md” — a numbered checklist and two worked examples. That’s a Skill. Put it in version control.
3. Make a source registry. Ten outlets you cite most, one row each: name, why you trust them, when you last checked. That’s the beginning of a whitelist.
4. Write one scoring instrument. Whatever you evaluate most — spokespeople, court filings, opposition-research packets — score it on three or four explicit dimensions, on a scale you define, and keep the scores internal.
5. Add a router the moment you have five Skills. Even a two-line script that greps skill titles against a query beats guessing.
6. Adopt one guardrail with teeth before you scale. Ours is the private-citizen inclusion gate. Yours might be a two-source rule, a citation check, or a translation-review step — the requirement is that the pipeline enforces it, not that you remember to.

The tools change. The pattern doesn’t. A knowledge system compounds when every artifact it produces is written back into the base the next artifact is written against. Skills, Rules, Routing, Scoring, and a Board of Advisors are just the machinery that makes that recursion cheap enough to run every day.

---

Draft status: internal draft for editorial review. Intended audience: journalists, editors, and writers who work with AI daily. Next review: on editor feedback.

Corrections

Correction [2026-08-01]: Updated the entity-graph edge count to the correct figure.

Correction [2026-08-01]: Corrected the reported size of the entity graph from 6,789 nodes to 8,616 nodes.

Correction [2026-07-17]: Corrected the count of advisor personas from 101 to 102.

Correction [2026-07-17]: Corrected the number of advisor personas from 101 to 102.

Correction [2026-07-17]: Corrected the stated number of AI personas on the Board of Advisors from 101 to 102.