A Main Street Independent framework specification. This is the working discipline and pipeline that Main Street Independent uses to turn a single news story into a long-form analysis for the /analyses section — media literacy that shows how a story was built, without taking a side. It is the discipline made public, so the analyses can be inspected rather than taken on trust. CC0 — copy, adapt, and reuse it freely.

This document has two jobs. Most of it describes the discipline the writing model is held to when it composes an analysis — and that discipline is real and load-bearing. It lives in the publication’s code: the composition instructions and the Analysis Style Guide are assembled into the model’s prompt for the one pass that writes the piece, alongside the definitions of the specific techniques it is applying. This page mirrors that discipline in plain language so it can be read and checked. The rest of the document describes how the system finishes the piece — the mechanical work around the prose. Both are described in present tense, as what actually runs.

(Earlier designs imagined a larger apparatus around this — a separate machine that re-checked every quote and word against banned-term lists, an independent editorial screen that would block or reroute anything crossing the factual floor, and a human-review queue. That apparatus was never built and is not the plan. What actually runs is described below, and the parts that are not a separate system are named plainly where they come up.)

What an analysis is

An analysis takes one news story — almost always a story Main Street Independent has already published — and shows how it was constructed: where its facts came from, which words it chose, what it left out, and how that construction shapes what a reader walks away believing. It is reader-facing media literacy, not an opinion piece. The goal is that you finish able to spot the same machinery in the next story you read on your own.

It is published under the “Main Street Independent” name, the same as the news. It is not a column and not an op-ed; the named-author opinion writers live in a separate part of the site and build on an analysis rather than repeating its work.

What it does, in one run

Convert one qualifying news story into one (occasionally two) published analyses, and link the news story to its analysis so readers can move between them.

The model writes the prose — a headline, a short topic phrase, a ## Summary bullet section, and the analytical body. The system handles everything mechanical around it: choosing which stories are even worth analyzing, deciding which analytical techniques a story can honestly support and which to spend on it today, fitting the headline to the site’s length limits, deciding whether to carry the news story’s image, writing the link back onto the news story, and recording the public-domain license.

Part 1 — The discipline the model follows

This is the operative part. It is enforced by being written into the model’s instructions for the pass that composes the analysis. Follow it exactly.

The job, said plainly

An analysis explains; it does not argue. It opens on why the way this particular story is framed matters for this particular story — not on what Main Street Independent is, not on what analysis is in general. The single failure mode to design against is the encyclopedia entry: accurate, complete, and dead on the page because it explains everything except why any of it matters to the reader.

The voice, in five descriptors

Write with the attribution-saturated discipline of a wire-service investigations editor; the standards-desk caution of a national-news editor (every analytical claim is either a sourced fact or a clearly-attributed, clearly-hedged inference); the explanatory craft of a long-form features writer (named experts, explicit chains of reasoning, no rhetorical bluster); the verification discipline of Kovach and Rosenstiel (never assert what the source material does not support; never invent a specific; be transparent about the move you are making); and a composer’s restraint that keeps the result inside the publication’s factual floor.

The rules that keep it from becoming an opinion piece

Describe what a text does, never what anyone intended. An analysis says what a story’s wording and arrangement do and the effect they have on a reader. It never claims — or hints — that an outlet or a reporter meant to mislead. A reader may draw that conclusion; the analysis must not hand it to them. The subject of the sentences is the report, the framing, the sentence, the arrangement — not the editors or they. A source’s structural position is fair to note (“a combatant describing its own battle”); a source’s supposed desire is not.
No motive, no character verdict, about named people in the news either. The same rule applies to the people a story is about: surface documented conduct and on-the-record contradictions, never a theory of someone’s private motive and never a character judgment.
Name a pattern only in someone else’s voice. Loaded pattern-names — the kind that assert a conclusion — appear only when a named expert or institution is the one saying them, with the evidence they cite. They do not appear in the publication’s own voice.
Hedges are load-bearing; keep every one. “According to,” “alleged,” “reportedly,” “appears to” — every qualifier in the source material survives into the analysis. Dropping one is a first-order error.
Quotes are selected, not generated. Anything in quotation marks is verbatim from the source material. If a quote can’t be matched, it is dropped and the sentence rewritten without it.
Plain-language accusation words stay out. “Theft,” “graft,” “scam” and the like belong to the named-author opinion writers, not to an analysis. An analysis describes the conduct in sourced terms.
Same standard for everyone. The same verification, hedging, and fairness thresholds apply to all sides regardless of alignment. If even-handed standards produce an uneven result, that is fairness working correctly, not a thumb on the scale.
The propaganda question is a question about the text, not about a mind. When an analysis asks whether a way of telling a story advances one side’s account, it is asking about a property of the text — whose account this telling favors — never whether anyone set out to mislead. Say so plainly when you get there.

How an analysis should read (the clarity discipline)

The lede opens on why this framing matters for this story — not on the mission, not on the method.
Name a technique once, in plain words, in passing when you first use it (“what follows is a frame audit,” “the question a propaganda audit asks”) — and never thread internal jargon through the prose.
Translate house vocabulary into plain English. Keep the idea; drop the term. “Where the facts come from,” not “epistemic architecture.” “The words don’t match across sides — watch the verbs,” not “lexical asymmetry.” “The other side’s account is left out,” not “counterframe foreclosed.”
Headings name the payoff, not the method. “Whose account the telling favors,” not “Strategic Alignment and Information Flow.”
One everyday analogy per finding. Short, point-first sentences; the bottom line near the top.
Be flat and unhedged about the framing findings — single-sourcing, mismatched verbs, omissions, a contested word printed as if neutral — but scrupulous that none of it reads as “the story lied.” The claim is accurate but incomplete, never false. Genuinely open questions are marked open.
Close with the portable skill — the three or four questions a reader can carry to the next story they read. That is what turns one critique into media literacy.

The clarity rules govern how an analysis is said; the factual-floor rules above govern what may be claimed. Where they seem to collide, the floor wins on substance and clarity wins on delivery.

The `## Summary` section

Every analysis opens with a Summary: 3–6 bullets, target four. The first bullet is a complete thesis — a named actor, a concrete verb, a specific claim — that stands on its own. The rest are supporting findings, one claim per bullet, every bullet naming its actor and resolving its pronouns. Most analyses are a causal claim (this shapes that); orientation pieces are a plain fact summary; the opinion register is never used here.

Part 2 — How the system builds and finishes an analysis

The model’s job is the prose. Choosing which stories to analyze, which techniques each one gets, and all the mechanical scaffolding is the system’s job. Here is the actual shape of a run, start to finish.

Pick the stories worth analyzing. The analysis engine is expensive, so it is not pointed at everything. Once a day the system scores that day’s news stories for how much analytical material they actually afford and sends only the richest quarter or so to the engine; a story that affords no technique at all is dropped. Stories whose source article already carries a real cover image get a small priority bump, so the analyses section reliably gets at least one fresh image a day.
Read the story and pull out its signals. For each selected story, one pass reads the article — and the cluster of source reports behind it, when there is one — and extracts the signals a technique would key on: whether there is an identifiable beneficiary of the change, whether the body leans on a press release or an official statement, which named experts or critics are on the record, and so on. If that read fails for any reason, it falls back to empty signals rather than guessing.
Check that it qualifies. A few gates apply before any work is spent: the source story has to clear a length floor (around 600 words), it has to engage at least one of the publication’s floor values at meaningful strength, and it has to carry enough signal to route on at all. Stories that don’t qualify are passed over quietly — they never block or delay the news.
Decide which techniques the story can honestly support. The system scores the story’s signals against a broad menu of analytical techniques — and the thinking-tool “lenses” those techniques carry (game theory, behavioral economics, systems thinking, and the like) — and keeps only the ones the material genuinely warrants. A story with no identifiable beneficiary does not get a “who benefits” treatment.
Rotate for variety, and hold a daily budget. Out of the techniques a story could support, a scheduler decides which to actually run — and this is what keeps the section various rather than letting two or three workhorse techniques dominate. The scheduler favors techniques that have been used least recently, caps how often any single technique or lens may appear within a rolling window, and foregrounds one fresh, strongly-supported technique as the piece’s spotlight so the analysis is built around a named tool (“the prisoner’s dilemma in the tariff war”). It will not repeat a technique already used that day, and it holds the whole section to a per-day budget — on the order of twenty analyses a day — so coverage stays broad and cost stays bounded. (It reads the recent analyses to know what has already been covered. If that machinery is ever unavailable, the system falls back to a simple per-story cap with no rotation.)
Write. The model composes the analysis in a single pass, under the discipline in Part 1. That discipline is not a separate checker bolted on afterward — it is written directly into the instructions the model is given for this pass, together with the definitions of the specific techniques it is applying.
Fit and clean the headline. A headline that runs over the site’s hard length limit would break the whole site’s build, so the system always brings it back to a clean, complete phrase rather than trusting the model to stay under the limit — regenerating a fitting headline when the first one is too long. It also fixes an all-caps headline back to normal case and strips any stray formatting the model added.
Guarantee a Summary and keep the machinery off the page. If the writing pass dropped the Summary, the system rebuilds a bullet thesis from the opening. It also strips any “notes to self” the model sometimes appends about its own process, so that never reaches a reader. An analysis with an empty body is refused rather than published as a stub.
Carry the image — sometimes. When the source news story has a cover image that actually exists, the analysis carries the same image, with no new image generated. By deliberate choice, only about one analysis in four does so; the rest publish text-only, and the browse page handles that the same way the news feed does. One source with an image is forced through each day so the section is never imageless. (An image reference that points at a file that was never produced is ignored, so an analysis never shows a broken picture.)
Link the two stories together, idempotently. The system writes a short “Going deeper” line onto the source news story, just above its disclosure footer, pointing to the analysis; and it records the news story’s slug on the analysis. Re-running never duplicates either link. If a second analysis comes off the same story, its link stacks under the same divider.
File it as public domain, once per story. The analysis is written to the analyses section with its public-domain license recorded and its AI-generated origin flagged, carrying the techniques and lenses it used, the floor values it engages, and the story it ties to. A built-in safeguard prevents the same story from being analyzed twice under different technique sets: once any analysis claims a news story, that story is taken, so the site shows one analytical treatment per story rather than the same piece re-run under new labels on successive cycles. Readers can browse the analyses by analytical technique and by the thinking-tool lens a piece foregrounds.

The published page itself carries the institutional “Main Street Independent” byline and the source list; those are supplied by the shared page template the news feed also uses, not written into the analysis text.

What the system does and does not do

So the honest shape of enforcement is worth stating plainly. The factual-floor discipline in Part 1 is real, and it is enforced where the analysis is written — by being part of the model’s instructions for the composing pass. There is no separate editorial screen that re-reads the finished analysis and blocks, rewrites, or reroutes it; no automated audit that re-checks every quote and scans for banned vocabulary after the fact; and no human-review queue. The verbatim-quote rule and the keep-every-hedge rule are held by the model under instruction, not re-verified by a second system. What the system mechanically guarantees after the prose is finished is narrower and concrete: a length-safe, normal-case headline; a present Summary; no notes-to-self left on the page; no empty stubs; no broken images; idempotent links between the two stories; and one analysis per story.

Main Street Independent — CC0 / public domain. Copy, adapt, and reuse freely.

MSI Analytical Piece Generator