Inside the Towns
One codebase, one editorial model, twenty-two Colorado town guides covering twenty-six communities. Each guide reads as if it were built for that town alone; the whole network is run by one person.
In short
- The problem. Growing Colorado towns, from the plains to the mountains, have no single place to find what's on, where to eat or what living there is like, and one person had to run every guide.
- What I built. Twenty-two town guides and a hub from one system, each with its own look and local research, and a build that will not publish a listing without a source and a date.
- What it means for you. Several locations or brands can run from one site, and a weekly update is a quick edit, not a rebuild.
- Owner
- Lerner Works. I publish and edit every site in the network myself. None of them is a town government, a chamber of commerce or a campaign.
- Sector
- Community and visitor information, Colorado's Front Range and the mountain towns above it
- Scope
- Network strategy, content model, shared design system, the Astro build, per-town research and editorial, the accuracy system (source registry, feed ingest, review), deployment, the weekly maintenance tooling
- Year
- 2026
- Status
- Twenty-two town guides and the hub live at insidethetowns.com; since October 12 two further domains, Severance and Fort Lupton, redirect into the guides that now cover them. Figures below read from the repository on October 13, 2026, at its commit of that day; every domain was last checked answering on October 19, 2026
At a glance
- twenty-three sites — twenty-two town guides and the hub — built and deployed from one repository
- 1 codebase
- place and event records across the twenty-two guides, each one a file in the repo
- 743 + 979
- people in the guides’ own towns, by the 2020 census, across ten counties
- 535,339
- the longest a mountain listing goes unchecked before the build hides it; 90 on the Front Range
- 30 days
The situation
A growing Front Range town has a strange information gap. Ask where to eat on a Tuesday, what is on this weekend, which trail starts where, or what it is actually like to move there, and there is no single place to look. The official town site, where there is one, is built for permits and council agendas — that is its job, and it does not pretend otherwise. A chamber directory lists members. Facebook groups know the answer but will not tell you twice. The rest is a scatter of aggregator pages that were last true in 2019.
I had already built one answer to this for a single town. TownofNiwot.com was a nine-page independent guide to Niwot: a directory where every row carried a source and a date, a calendar with no scraper, and neutral coverage of a live incorporation election. It worked. Then the obvious question arrived, and it was not "what else should this site do." It was: Niwot is not the only town like this.
Lyons, Berthoud, Erie, Johnstown, Timnath, Elizabeth. Same gap, different town. Between the 2010 and 2020 censuses Timnath grew 938 percent, Berthoud 102, Johnstown 75, Erie 66 — while Lyons grew 9 and Niwot 7. Two very different problems sit under one format: towns arriving faster than their own information, and towns whose information nobody has bothered to gather.
Those were the first seven. By October the list had reached the cities those towns sit beside — Fort Collins, Longmont, Loveland — and the mountain towns above them, from Golden to Leadville, where the gap has a third shape: a season, a road and a trailhead that are open in July and shut in January.
The problem in one sentence: every one of these towns needs a site that feels like it was made for it, and one person has to be able to run all of them.
Constraints
One repository, no forks. The moment a town's difference lives in code rather than configuration, there are two dozen codebases wearing a trench coat and a fix has to be applied two dozen times. Town differences live in a config file and a content folder; nothing in a component may name a town, a color or a domain.
The sites must not look like a template. A reader in Lyons should not be able to tell that the same code serves Timnath. Shared structure, shared type, shared rhythm — distinct color and distinct photography, per town.
A new town under an hour. If adding the next town is a project, the network stops where it is. Scaffolding has to be a command.
A weekly content update must be a content edit. Not a deploy someone has to think about, and not a code change.
No invented local knowledge. Same rule the Niwot guide ran on: every event cites the organizer's own page, every image is logged with its license, and anything that cannot be checked is left out or marked rather than guessed. It started as an editorial rule. It is now one the build enforces.
What I built
Parent hub
Inside the Towns
Every town's weekend in one place, a comparison for people deciding where to move, one search across every guide, and the front door for the next town added.
Boulder County
Northern Colorado
Carbon Valley & I-25
Mountains & Foothills
South Metro
Three colors per town, each with one job. Twenty-three palettes need a rule, or they stop reading as one network. Under the network's Color System 2.0, applied to every site on October 10, each town has a primary for identity and action — the town's name, links, buttons, the dark bands, always with white on it — an accent that is decoration only and never carries a word, and one of two shared papers, warm or cool. Body text, supporting text and white cards are the same on every site, and those shared neutrals are most of what makes twenty-three palettes read as one family. The network's amber is kept for meaning rather than decoration: a dated notice, an "Our pick", the hub's name in the strip under every footer.
Nothing publishes without a source and a date. Each town has a source registry — the town calendar, the library, the chamber, the venues its calendar leans on. Confirmed feeds are read into a staging folder, never onto the site, and a change to an event already published, a moved time or a cancellation, is flagged on that listing and waits for a person. A weekly review walks the queue, cancellations first, and approves, edits or rejects each item. The build then refuses anything without a source and a verified date, hides an open listing once its last check is older than its window — 90 days on the Front Range, 30 in the mountains, where seasons turn faster — and keeps a closed place's page up with a notice rather than deleting it. Every item links to a corrections form, and every site keeps a corrections log.
Mountain towns are a variant, not a fork. Ten of the guides are in the mountains, and a mountain listing carries things a Front Range one does not: a season with the months it opens and closes, a trailhead with its parking, permit and land manager's conditions page, and the state's road-conditions map for the drive up. They are optional fields on the same schema, switched on by one setting in the town's config — "Open May–Oct", "Closed for the season; reopens May" — and a season turning puts the listing back on the re-check list. The Front Range guides' This Week pages end with "Up the hill this weekend": two or three of the mountain guides' weekend listings, each linking to its page on that guide's own domain.
Forms built dark, then switched on. The contact and submit-an-event forms on every site are built behind a provider id and ship a fallback until it is set. On the hub and the first nine guides they are live, one endpoint per site rather than one shared, so a town that is ever handed to someone else takes its inbox with it. The fifteen newer guides still ship the fallback, and it was made worth using: an email to that town's own address with the form's questions already in the body, so the first message can be checked as it stands. The build checks every host a form posts to against the sites' content-security policy, so a form cannot point somewhere the page would refuse.
How it's built
One chassis, selected at build time. The whole network is a single Astro project. A TOWN environment variable picks which site is being built — TOWN=lyons npm run build — and the config loader fails loudly if it is unset or names a town that does not exist, because a build with the wrong town is worse than no build. Twenty-six Vercel projects point at the same repository, one per domain — the twenty-two guides, the hub, one town not yet launched, and the two folded towns, whose projects keep building so that every old URL redirects — each with its own TOWN. A push rebuilds only the sites it touches, and a change to any town's config rebuilds them all, because every site shows something from every config: the hub's list, each footer's sister guides, the Nearby block.
Town differences live in exactly two places. A config file in src/config/towns/ carries the domain, the tagline, the region, the county, the census figures, the school district, the commute notes, the official links to link out to, the neighbors, whether it is a Front Range or a mountain guide, and the three colors. A content folder in content/<town>/ carries the events, places, articles, pages, images and the source registry. Thirty-eight components and two layouts are shared by all twenty-five sites, and not one of them knows a town's name.
The roles are in the names. The config's three colors are primary, accent and paper, and the utility classes follow — text-primary, bg-primary, border-accent — so the next change cannot use one where it means the other. Buttons darken 12 percent on hover, mixed in plain sRGB, so the page carries an exact hex and the checker can prove white text on it.
A hue per category, so a list reads as a list of kinds. Each event category and place type carries two values: a bright dot for marks, and a dark ink for when the category is set as words. Place types borrow the event hues by what the place is for, rather than inventing more. Twenty upcoming events read as twenty kinds of thing rather than twenty gray rows.
The color system is proved, not asserted. npm run check-colors reads the real town configs and the real palette module — so it cannot drift from what renders — and checks each primary on white and on its paper, white on each primary and on its hover shade, both inks on both papers, every category color, and that no two guides share a primary. CI runs it on every push; on October 11 it passed all twenty-six configs. It is what made replacing every palette in the network in one day safe: Color System 2.0 arrived with final values for every site and was applied as written, without a hex tuned, and proved the same day.
The week starts with npm run ingest and npm run review: the confirmed feeds' new items and changes land in staging, and the review walks them under a summary of what is waiting, which listings go stale in the next fortnight and which sources have stopped answering. npm run weekly prints the longer report per town — events about to end, towns short of upcoming events, places with no photograph, that week's share of the listings to re-check, and the seasonal ones whose season is turning. Hand-gathered events go into a CSV, and npm run import-events validates every row against the schema, with a --dry-run that is meant to be read. Then npm run validate, commit, push.
Finding things, without a server. Every site has a search page: Pagefind indexes the built output at the end of each build and the page loads that static index, so search works on a host that runs nothing. The hub's index also holds every live guide's places, articles and upcoming events, each linking to the guide's own domain. The directory lists every place by kind, with a name search and an "open now" pill that is decided in the reader's browser from the Denver clock, because a build is a snapshot and "open" has to stay true after it. The events page asks when before what — today, this weekend, the next seven days, free — and carries the chosen filters in the URL, so "free things this weekend" can be bookmarked, shared and linked from the email.
Three layers that keep "upcoming" true between deploys. A static build is correct on deploy day and quietly wrong a week later. So: a scheduled rebuild fires a deploy hook for every live project each morning, around five o'clock in Denver; every event carries the day it falls on and anything past is dropped in the reader's browser, so a stale build shows fewer events rather than wrong ones; and the hub's weekend page bakes three weeks and picks the weekend from the reader's clock. The layers exist because the first one has specific failure modes worth designing around: GitHub starts scheduled runs late and drops some altogether, and it disables a scheduled workflow by itself after sixty days without a push, silently, in exactly the quiet winter when nobody would notice.
Decisions and tradeoffs
Config and content, never a fork. The rule that nothing in a component may hard-code a town name, color or domain is the single decision the whole network rests on. It costs something real: every town-specific idea has to be generalized into a field on the config before it can ship anywhere, which is slower than special-casing one site. What it buys is that a fix written once is a fix everywhere, and that the next town inherits every earlier town's work for free. October tested it twice. A guide covering three towns at once became a field, and ten mountain towns became a setting with a few optional fields. Neither needed a component that knows a town's name.
Dark mode, built and then removed the same day. It worked. I took it out anyway: I do not want these sites turning dark on a reader's phone without being asked. Two things from it were worth keeping and stayed — the text color split out from the band fill as its own token, because those two pull in opposite directions the moment anything changes; and the derived tints mixing in plain sRGB rather than oklab, so the contrast checker can reproduce them exactly. The exercise paid for itself on the way through, catching a rule at 1.45:1, which is not a rule, it is nothing.
231 comparison pages were not generated. Twenty-two guides make 231 pairs, and 231 pages built from one table with no argument in them is thin content — a search engine that decides that about one page tends to decide it about the set. A pair gets a page when somebody has actually written it. The list is a file, and three are written so far: Erie against Johnstown, Timnath against Windsor, and Windsor against Fort Collins.
No price column, and no price claims. The hub's comparison was built to draw a median-home-price column only when every live town had a price, a date and a named source — all three or none, because a gap invites the reader to read it as "cheap", and an undated price goes on looking authoritative for years after it stops being true. At twenty-two guides the answer is none: the table says there is no price column and why, and each town's Moving Here page links to what is actually on the market today. The same pass took out every remaining price comparison the guides could not back up — "Boulder prices", "short of Fort Collins" — and gave each a fact its own guide supports instead, such as a drive time.
The newsletter shipped dark, then switched on. The signup block rendered nothing at all until a provider was configured on the hub — not a disabled form, not a "coming soon" input, nothing — because a form that takes an address and drops it is worse than no form. It is live now, through Buttondown, and it is one form on every page that asks: one email a week on Thursdays, with a checkbox per guide, so the Erie issue goes to the people who asked for Erie. Each issue is drafted from the listings by a script and read by a person before it goes out, and the archive page fills as issues are sent.
Live under the line, and saying so. Each guide has a launch threshold: enough upcoming events and checked listings that the site is worth a visit. The six newest mountain guides went live on October 10, the night they were scaffolded, on my call and under their thresholds. Their content went in under the accuracy rule rather than the coverage one — a listing waits in staging until its source has been read, and a section with nothing verified stays empty — and the validator and the weekly report say what each guide is short of. Castle Rock went the other way: it has a domain, a project and thirty places, and it stays a holding page until it clears its threshold of twenty-five upcoming events.
Research delegated, and recorded as delegated. The research and first review for the newer towns were done by AI agents working under my direction, one per town, each re-reading every staged item's source page on the day. Every approval they made says so in the file — verifiedBy: "Claude Code, delegated by James Lerner" — so the record shows who checked a listing and on whose authority. The rules did not loosen for it: nothing is approved without a source, and anything whose site refused a reader waits in staging with a line saying what it still needs.
Growth is census-to-census, and says so. The brief wanted five-year growth. The state demography office publishes annual estimates, which would be the right source, but serves them from interactive lookups rather than a file — so there was no way to get a consistent figure for every town. The decennial census is authoritative and identical in basis across all of them. It is a weaker number, honestly labeled, instead of a stronger one nobody could check.
How the network scales
Adding a town is not a build. One command sets up the new site's files. What remains is the part that cannot be automated and should not be: the research. Reading the town, finding the places, confirming the events against organizers' own pages, and sourcing photographs with a license attached.
That is the real shape of the cost curve, and October tested it. The first town cost a site. Seven were live when this study was first written, on September 18. Windsor and Fort Collins went live on October 1; Longmont and Loveland were scaffolded on October 4 and live the next day; Carbon Valley was scaffolded, reviewed and live on October 5; six more went live on October 7; and six mountain guides were scaffolded and live on October 10. None of them needed a change to a component that names a town. Each needed its research, and a palette from the network's color system rather than a color chosen on the day.
Carbon Valley answered the question this study left open. The September version named carbonvalleyguide.com as the next site, and the first that would be three towns rather than one — Frederick, Firestone and Dacono — which the chassis had never been asked to hold. The answer was a field. A guide with subTowns requires a town on every event and place, in the published folders and in staging; its events and places gain a "Where" filter that stays in the URL; and one injected route gives each of the three towns its own page, built from the same collections. No other guide carries the field, so no other guide builds the route.
The next one is held rather than planned. Castle Rock has everything but its events and its colors: Color System 2.0 left it out, so it keeps a placeholder palette, flagged in its config, until it has one. It is not described as live anywhere on any of the sites, and will not be until it is.
How it's built
npm run new-town elizabeth "Elizabeth" writes the config file, appends the registry entry, creates the content folder, generates a favicon, drops in a placeholder hero, and prints what to do next. Until the config says status: 'live' the site builds one noindex holding page, and the hub does not name it; the day it flips, the hub, the network bar, the neighbors and the nightly rebuild all pick it up from the registry.
Result
The figures below are counted from the repository on October 13, 2026, at its commit of that day. They are things that can be counted, not claims about reach. The study was first written on September 18 for seven towns, and every count has moved since because the network more than tripled in three weeks.
The one reach figure the network does publish is on the hub's own home page: the twenty-two live guides’ own towns hold 535,339 people across ten counties — Boulder, Larimer, Weld, Elbert, Jefferson, Gilpin, Clear Creek, Grand, El Paso and Lake — summed from the 2020 census. It is a census figure rather than an audience figure, and the hub labels it as one. The sum fell on October 12 when Severance and Fort Lupton stopped being guides of their own: the Windsor and Carbon Valley guides now cover them, but the hub sums only the towns its live configs name.
- Live sites
- 23 — twenty-two town guides and the hub, from 1 codebase; two more domains redirect
- Shared components
- 38 components, 2 layouts
- Tests
- 44 test files; the last clean run of
npm test, on October 11, reported 305 - Place records
- 743 across twenty-two guides
- Event records
- 979 across twenty-two guides
- Page routes
- 16 per town guide, 9 on the hub, 9 shared
- Palettes checked
- 26, every config live or held, against WCAG AA, in CI
- Per-town code
- 1 config file, 1 content folder
- Feeds per town
- 2 RSS, 3 iCalendar: the calendar, one per event, one per show of a production
Not yet measured: traffic and newsletter signups. They'll be added with a date range and source after the first full month.
Where it stands
Twenty-two town guides and the hub are live, one Vercel project per domain, all deploying from the same repository; on October 11 every one of the then twenty-five domains answered. On October 12 two guides were folded into their neighbours — Severance into Windsor, Fort Lupton into Carbon Valley — rather than kept thin: their content moved into those guides, each gets a section page there, and insideseverance.com and insidefortlupton.com redirect path for path. The hub capture above is from October 11 and still shows twenty-four. The network runs on a weekly editorial session rather than a release cycle: read the feeds, review the queue, re-check that week's share of the listings, push.
TownofNiwot.com, where this started, now redirects to Inside Niwot. Every old section URL maps to its new page and everything else goes to the home page, so nothing that was ever linked is lost. The old case study stays published because the work in it is the foundation this network is built on — the sourcing rules, the dated records, the refusal to guess — and because a redirect is worth explaining rather than quietly performing.
The parts that were dark when the study was first written are on: the newsletter signup, a hello@ address at every domain that reaches a real inbox, and the forms on the hub and the first nine guides.
What is open is the honest part: no newsletter issue has gone out yet — the first was drafted for October 8 and not sent — so the archive is empty; 148 of the 743 places carry a photograph and the rest are being asked for one, business by business; the six newest mountain guides are live under their thresholds, with listings waiting in staging and Moving Here pages still to write for five of them; the fifteen newer guides' forms are still the email fallback; and Castle Rock is a holding page until it clears its threshold.
What this means for your project
You probably do not need twenty-two websites. The decisions that made twenty-two maintainable by one person are the same ones that make one site cheap to keep current.
- Content kept apart from design, so a weekly update is an edit rather than a rebuild, and one you can make yourself.
- Several locations or brands from one build. If you run three branches or two businesses, one codebase and one set of rules is the difference between a site that stays current and one that quietly stops.
- Accuracy and contrast proved by the build, not judged by eye: a listing without a source does not ship, and every page clears accessibility before it goes live and after every change.