Three Products, One Pipeline

The Citizen’s Daily Brief ships three products: a daily briefing, a weekly assessment, and an occasional fictional intelligence dossier. They share the same data, the same Supabase backend, and the same editorial philosophy. They’re also three genuinely different things — different artifacts, different schedules, and in one case a different idea of what “production” even means.

Two of them are written by language models on a clock. The third is written by a person, when the news earns it. Getting the relationships among the three right was one of the more interesting design problems in the project.

Daily: What Changed Today

The Daily Brief answers a simple question: what are the five to nine most significant things that happened in the last 24 hours? It runs Monday through Saturday, ingests from thirty-four RSS news feeds spanning the political spectrum, clusters and scores stories by editorial perspective diversity, and synthesizes structured assessments with trust signals.

Claude Sonnet handles all the daily analytical work — clustering, significance scoring, and synthesis. Sonnet is fast, cost-effective, and excellent at following structured output constraints. For the daily brief, those qualities matter more than raw analytical depth because the task is well-defined: take these sources, group them, score them, assess the top clusters. The prompt is precise, the output schema is strict, and the validation layer catches anything that slips through.

The daily brief is a snapshot. It tells you what the world looks like this morning. It’s designed to be read in under ten minutes, and each item stands alone. You can read item three without reading items one and two. The items are ordered by significance, but they’re not narratively connected. The main thing the items share is enough weighted importance to make the cut for a given day.

Weekly: What Does the Week Mean?

The Weekly Assessment is a different animal entirely. It runs late every Saturday for Sunday consumption, takes the entire week’s daily items as input — typically thirty to forty-two assessments from Monday through Saturday — and produces a long-form analytical document. Four to six thousand words. Fifteen to twenty minutes of reading.

And it runs on Claude Opus.

The model choice isn’t arbitrary. The weekly assessment asks the LLM to do something qualitatively different from the daily work. Instead of assessing individual stories against their sources, it’s analyzing the week as a whole — looking for narrative arcs, cross-domain connections, developing situations, shifts in confidence, and how trends and trajectories varied, interconnected, and resolved across that week’s briefs.

These are harder analytical tasks. Tracing how a trade policy story connects to a labor market story connects to a consumer confidence story requires holding multiple complex assessments in context simultaneously and reasoning about their relationships. Noticing that Monday’s “developing confidence” item resolved to “high confidence” by Thursday, and what that trajectory implies, requires a kind of meta-analytical thinking that benefits from Opus’s deeper reasoning capabilities.

The cost difference is substantial on a per-call basis — Opus is significantly more expensive than Sonnet. But the weekly assessment runs once per week. At that frequency, the cost difference amounts to a few extra dollars per month. The analytical upgrade is worth every penny.

The Relationship: Meta-Analysis, Not Re-Ingestion

Here’s the design decision that matters most: the weekly assessment doesn’t go back to the raw sources. It doesn’t re-ingest RSS feeds from the past seven days and start from scratch. It receives the already-synthesized daily brief items — complete with their headlines, assessments, trust signals, confidence levels, and source attributions — and builds on top of them. So in this way the data builds on itself, which helps a great deal on consistency and reliability.

This is meta-analysis. The weekly assessment analyzes analyses.

There are two reasons this works better than re-ingesting raw sources.

First, it preserves the daily assessments’ integrity. Each daily item went through clustering, significance scoring, synthesis, and validation. Those trust signals — confidence levels, agreement indicators, source counts — were calibrated against the sources available on that specific day. If the weekly assessment re-ingested raw sources, it would lose that daily context. Was this a high-confidence assessment on Tuesday or a developing-confidence one? That distinction matters for understanding how a story evolved.

Second, it enables pattern detection that raw sources can’t provide. The weekly assessment can see that four daily briefs across the week each included an item touching on supply chain disruption, even though each individual item was about a different specific event (port delays, chip shortages, agricultural export restrictions, shipping route changes). No single day’s sources would reveal that pattern. But the structured daily assessments, with their topic tags and significance scores, make it visible.

The weekly prompt explicitly asks Opus to do several things the daily pipeline can’t:

Week in Review — A narrative arc of how the week unfolded. Not a list of things that happened, but a story about how the week’s events connect and build on each other. Three to five paragraphs that give you the shape of the week.

Deeper Dives — Two or three stories that deserve more than the daily brief’s compressed format. Six hundred to a thousand words each, with room for alternative interpretations, historical context, and cross-story connections that a daily assessment can’t explore.

Trend Analysis — Patterns across the week’s items that were invisible in any single edition. The supply chain example above is one pattern type. Others include: escalation sequences (where each day’s story was slightly more severe than the last), convergence patterns (where stories from different domains started pointing in the same direction), and reversal patterns (where Monday’s assessment was contradicted by Friday’s developments).

Developing Situations — Ongoing stories that are building toward something but haven’t produced a decisive event yet. These are explicitly forward-looking and explicitly uncertain. The weekly assessment says “this is developing” in a way the daily brief can’t, because the daily brief is optimized for “this happened.”

Corrections and Shifts — Where the week’s confidence and agreement levels changed. If a story was “developing confidence” on Monday and “high confidence” by Thursday, what changed? If agreement shifted from “broad” to “disputed,” what new information caused the disagreement? These shifts are the most analytically valuable part of the weekly assessment because they show how understanding evolved, not just what happened. They’re my favorite piece of the weekly assessment puzzle because they show how purely focusing on one day at a time without prior or future context can lead you astray.

Dossiers: What Could Happen

The third product breaks the rhythm the first two establish. The daily brief and the weekly assessment are both automated, and based on current known information — a model takes structured input on a schedule and returns a structured artifact. A FICINT dossier works nothing like that. No model decides when one ships, what it argues, or how it reads. A person does. Usually me. And it’s not about what we currently know, it’s about what we don’t.

FICINT stands for Fictional Intelligence, and a dossier is exactly that: a fictional case file, built to look like a declassified intelligence packet. A cover page with a classification bar and a case number. A sequence of artifacts in distinct documentary voices: an incident report, a quality-review memo, a parent’s kitchen-table testimony, a hearing transcript. Each one anchors on a real pattern in the news and projects it three structural moves out, to the specific, non-obvious consequence the daily brief can name but never dramatize. It attempts to answer the question, “now what?”

Where the daily and weekly run on a clock, dossiers run on a threshold. Four themes (for now) sit on a watchlist — Anti-AI Fracture, Autonomy Gap, Legitimacy Vacuum, State Capacity Atrophy — each tied to a historical blueprint. The daily pipeline ranks every story it clusters, not just the handful it publishes, and keeps the full list. That list feeds a tracker watching for real events that cross the line a given theme is built around. When a theme trips its threshold, it becomes eligible for a dossier. Filing one discharges the threshold, and the theme goes quiet until fresh events trip it again. The discipline is the entire point: a dossier lands maybe once a month at most, often less, and only when reality has earned it. That restraint is the most editorial decision in the whole system, and it’s the one I’d never hand to a model.

The production path shares almost nothing with the other two. The automated products emit JSON against a strict schema, validated and poured into a fixed template. For those, the human is deliberately out of the loop. A dossier is hand-authored, assembled from custom artifact components — one for the incident-report register, one for testimony, one for the federal-document voice — each styled to read like the real document it imitates. The automated products are judged on schema conformance. A dossier is judged on whether a mother’s testimony earns its weight on the page. Different craft, different failure modes.

And yet it ships on the same rails. Same Supabase tables for storage and metadata. Same push infrastructure to put it in front of readers. Same web codebase to render it. The pipeline built to move a brief and an assessment from output to reader turned out to move a hand-built fiction package just as well. That is the payoff I didn’t plan for: the infrastructure turned out general enough that a third product, built a completely different way, could join without a rebuild.

The Same Data, Differently

The elegance of this architecture is that each product is a real artifact in its own right. The daily brief is an assessment of today’s events. The weekly assessment is an assessment of the week’s assessments. The dossier is a disciplined fiction about where a pattern in those assessments could lead. Three artifacts, three cognitive jobs, one body of data underneath. All three together provide a fuller picture of what’s going on in the world than any one of them in isolation.

The daily brief answers: “What should I know this morning?”

The weekly assessment answers: “What does this week mean, and what should I be watching?”

The dossier answers: “If this keeps going, what does it look like from the inside and what might it lead to?”

You can read any one without the others. Daily-only readers get a practical, finite morning briefing. Weekly-only readers get a deep analytical document that stands on its own. Readers of both get something neither product provides alone: the experience of watching a structured analytical process unfold over time, with the weekly assessment explicitly building on the daily foundation. And a dossier often runs the other way — it’s a stranger’s first contact with the project, the fiction arriving before the brief that grounds it.

Matching the Tool to the Task

There’s a fair question here: why not just use Opus for everything? The daily brief would be even better with Opus’s analytical depth, right?

Honestly, maybe marginally. But the daily brief’s quality is constrained more by format than by model capability. Each item has tight field limits — “what changed” is one to three sentences, “why it matters” is a short paragraph, “what to watch” is a sentence or two. Within those constraints, Sonnet produces assessments that are clear, accurate, and well-structured. Opus might produce slightly more nuanced assessments, but the nuance would be compressed into the same small fields, and the practical difference for readers would be minimal.

The weekly assessment has room for nuance. A thousand-word deeper dive can contain the kind of multi-perspective analysis that benefits from Opus’s reasoning. A trend analysis section can draw connections that require holding a week’s worth of complex assessments in context. The format allows the model’s capabilities to shine in a way the daily format’s constraints don’t.

Matching model capability to format requirements is more cost-effective and produces a better product than using the most powerful model everywhere. Use Sonnet where the format is tight and the task is well-defined. Use Opus where the format is expansive and the task is open-ended. The daily brief is the first. The weekly assessment is the second.

That logic reaches past the choice between models. The daily and the weekly both ask which model fits the format. The dossier asks an earlier question: should this be automated at all? For a case file whose whole value is behavioral truth — whether a fictional triage nurse sounds like a real one under oath — the honest answer is no. A dossier is commissioned, authored, and edited by a person. A model might help draft a line, but nothing about it ships on a schedule a model controls. What it argues, which historical blueprint it leans on, whether it has earned its black swan: that is editorial judgment, and the pipeline exists to let a human exercise it. Sonnet where the format is tight. Opus where the format is open. An editor where the truth is behavioral.

All that said, I constantly experiment with generation options, trying out newer models as they come and go, so this model arrangement is subject to change as things continue to evolve.

The Business Model

I wanted the CDB for myself, to satisfy my desire for a daily brief as untethered as possible from news bias. I originally had plans for it to be under a subscription model, with daily briefs free and weekly assessments behind a paywall. But the more I used it, the more I realized it shouldn’t be pay gated, so it’s entirely free: the briefs, the weekly assessments, even the hand-curated FICINT dossiers.

That doesn’t mean I can’t use help, so there’s always the option to support the CDB with voluntary contributions to help pay the robot and hosting bills. But it’s not a requirement — I’d rather more people have access to it than try to make it some kind of profit center. It’s not. It’s a utility I built for myself, and I hope you find it useful.

Three products, one pipeline, one commitment to assessed information over raw content. The same rails carry a daily brief, a weekly assessment, and an occasional dossier without tripling the engineering. Two of them choose a model to fit the format. The third sets the models aside and hands the page to an editor.

The Citizen’s Daily Brief is a free daily intelligence briefing from Stalefish Labs.