iiinova — where your digital self keeps evolving

Labelling is a system: move one and five others shift too. Claude handles that knock-on tirelessly, but plenty of times I'd be faster by hand. But the main thing I learned from this is that labelling is a part of mapmaking that I love. Doing it through an AI was achievable, but quietly joyless.

The labelled map in full — the whole rotated Icefields Parkway strip, Jasper at the top to Lake Louise at the bottom. Hundreds of off-white names over the shaded-relief basemap: towns, peaks with triangles, italic lakes, glaciers and icefields, the Sunwapta and North Saskatchewan Rivers curving down their valleys, falls, with JASPER and BANFF NATIONAL PARK blended into the terrain and highway shields along the road.

241

Finally, combine them all with the basemap. Most labels are on top of everything, but the National Park labels had to blend into the basemap. Claude occasionally needed prompting to remember to blend, rather than just dump these on top.

Finished map, Bow Lake and the Wapta Icefield: lake names beside their water, glaciers and peaks, and the highway shield on the road.

Finished map, the Columbia Icefield: icefield and glacier names in italic, peaks and the Discovery Centre, all over the terrain.

Finished map, Saskatchewan River Crossing: the crossing, the river curving down the valley, surrounding peaks, and BANFF subtly blended into the hillshade.

Finished map, Jasper: the town, its lakes and the surrounding peaks labelled over the shaded-relief basemap, the park name blended into the terrain.

100

As a stress test I had Claude update the spec, then rebuild the labels from scratch: spec and data only, no peeking at the previous version(s). It worked, but exposed the spec's own shortcomings: one rule silently dropped labels for some of the main lakes visitors come to see.

A beige flow diagram titled 'Stress-testing the spec', with five cards: 1 Accepted so far, 2 Rewrite the spec, 3 Rebuild from scratch (spec + data only), 4 Gaps exposed (what the spec didn't capture), 5 Patch the spec. A loop runs from step 5 back to step 3, labelled 're-run and re-test until the spec reproduces the map'.

The same area in the final accepted labelling, after the gaps the rebuild exposed were fixed and folded back into the spec: the same names, now placed cleanly and evenly over the basemap.

100

I asked Claude to label the main highways with their official shields, and it told me, confidently, that it could create SVGs "from the official specs". Pushed on the source of those specs, it admitted it was making them up from memory. They were all crap. So I downloaded the real SVGs myself...

A comparison of highway shields: Claude's "from-memory" reconstructions, with wrong proportions and details, alongside the official downloaded SVGs for the Trans-Canada Highways 1 & 16, and Alberta Highways 11 & 93.

141

Lakes took the most rounds of iteration. They started as big italics sitting on top of the water. I wanted them beside the lakes, not floating on them. Then I had two lake sizes competing, so I dropped to one. Quiet names set off the water. Simple to describe, many iterations to actually produce.

Before: lake names like Bow Lake set large and bold on top of the water.

After: the same lakes named in a smaller italic set beside the water, one consistent style, near the Wapta Icefield.

Peaks looked wrong before I could say why. Two different sizes of summit made every cluster read as lumpy and random. Collapsing them to one weight quietly settled the whole sheet. Then the fiddly bit: an even, constant gap from each triangle, and never nearer a neighbour's peak than its own.

Before: the peaks around Mount Saskatchewan named in two different sizes with small triangles, the cluster reading lumpy and uneven.

After: the same peaks in one consistent size, each name tight to its own triangle, among the glaciers and the North Saskatchewan River.

Rivers fought me first. They curve along each channel, which is lovely when it works. But early on labels were straight, sometimes upside down. The halo on the smallest creeks was so heavy it swallowed them. Fixes: force the reading direction, thin the halo, bump the tiny rivers up a weight.

Before: the Siffleur River label follows the channel but is over-rotated past 90°, so it reads upside down on the page (each letter the right way round). Mount Weed, Bow Pass and Dolomite Pass sit around it.

After: the same Siffleur River label reading upright along a smooth curve, set just to the right of the channel (off the river and the track), with Mount Loudon, Porcupine Lake and Cirque Peak around it.

What to label? What matters? This is where AI was really useful. The map is for a visitor to the Icefields Parkway, so I just said: "label what a visitor cares about & fill gaps to maintain density". I didn't need to know the data schema & write out definition queries. I asked for what I wanted.

A near-black panel densely overprinted with every named feature in the data, in small uniform white type — hundreds of labels overlapping into a barely readable wall of text.

The same area with only the chosen names kept, in the same small uniform horizontal type: a clean, mostly legible scatter of labels, to show the filtering of what to include rather than any styling.

It started, like the basemap, with my own maps. In a separate chat I'd worked out how I want a map labelled (rules I usually keep in my head) and had Claude write it down as a spec. One typeface, one off-white ink, hierarchy by size, weight, style. Sixteen classes, before a word landed on the map.

A dark code-editor graphic titled 'label_spec_v2.0.json', styled like the earlier house-style spec image. Two columns of pretty-printed JSON showing a sample of the labelling spec: the global map settings (dpi, page size, scale, CRS, rotation), the Poppins fonts, the off-white ink values and the design principle, then two example label classes — town and lake-large — with a '… + 14 more label classes …' placeholder. A footer reads 'excerpt — full spec: 253 lines · 16 classes'.

A specimen sheet on a cream background with green sample panels, showing all sixteen label classes in one Poppins family and one off-white ink: a spaced-capital park name (sample 'BANFF') and range, a bold town and facility with dots, medium and regular peaks with triangles, italic icefields, glaciers, lakes, rivers and falls — the water and range names set along a gentle tilde curve — and the four highway shields centred in their panel. Each row is annotated on the right with its font, size, tracking and placement.

201

The basemap is geometry: run the recipe, get an answer. Labelling is judgement: What deserves a name? Where does the label sit? Is it legible? Claude could try to place a label anywhere I asked, but deciding what belonged, and whether it looked right, was up to me. So much iteration. So much.

A flow diagram on a warm beige background, titled 'Labelling: the process Claude wrote', with the sub-line 'Deterministic placement — but a human eye decides, and sends it back to loop.' Seven white rounded cards, each with a sage-green numbered circle: 1 Study my maps (distil my house style), 2 Type system (one font, one ink; weight is hierarchy), 3 What to label (visitor stops, corridor, prominence), 4 Pull the names (from CanVec 50k), 5 Place labels (no overlaps, curved, beside lakes), 6 Human review (does it read? does it look right?), 7 Labelled map (the finished sheet). A green 'happy' arrow leads from step 6 to step 7. A dashed clay-coloured loop runs back from step 6 to both step 5 and step 3, labelled 'not happy — back to step 3 or step 5, and re-run'.

Andrew Tyrrell @southarrowmaps.co.nz · 03/08/2026

I will never not be a fan of 3D maps in visitor centres. Here’s the one in the Glacier Public Service Center in Mount Baker-Snoqualmie National Forest

0120

So, can AI be a cartographer? Here, yes, as a partner: It set the pipeline, wrote the code, solved the water, and knew when to defer to me. The turning point was AI admitting it can't fake geography. Honesty plus real data is what made a real map. (Portrait below, by ChatGPT.) /end 🧑‍🎨🤖🗺️

230

The payoff: an accepted, print-ready basemap of the Lake Louise to Jasper corridor. Here is the whole thing, ready to be rotated and cropped to the long, thin road corridor. Unlike a flat AI image, every stage of this re-runs.

The full basemap again, the whole 285 by 311 km sheet, ranges and glaciers threading through forested valleys with turquoise lakes.

A close-up of the Columbia Icefield, a broad white glacier spilling between grey rock ridges, with the pale off-white Icefields Parkway curving past its toe and teal meltwater nearby.

A close-up of Bow Lake and Peyto, deep teal lakes set in forested valleys beneath snow-streaked peaks, with the highway threading north between them.

A close-up of the valleys around Jasper, converging turquoise rivers and lakes, forested slopes, and the highway and rail line meeting on the valley floor.

130

The coworker reality is the unglamorous bits. A cloud drive that would not mount. A sandbox blocked from the data hosts. 45-second compute limits leaving broken files. A proprietary font missing. Most of these Claude spotted and routed around itself, without me having to ask.

A four-card graphic titled "The coworker bits: what broke, and the fix", each card pairing a problem with its fix: cloud drive would not mount, 45-second compute cap, blocked from data portals, and no Futura font on the machine.

Infrastructure next, from the full 1:50,000 topographic data: every road, track and trail. The job was thinning it to what reads at 1:500k. Why not use 1:250k? I wanted to test AI. Roads become one ink weighted by class, rail a fine dark line, main trails dashed, and only the long tunnels dashed.

A before-and-after comparison at Jasper. The left panel, labelled Before, shows the full 1:50,000 network with dense white roads, tracks and trails cluttering the valley. The right panel, labelled After, shows the thinned result: a clean highway weighted by class, a dark rail line, and a few dashed main trails.

320

The hardest sub-problem, and still open: the region's famous lake colours. We tried proximity to the feeding glacier, distance along the drainage, a glacial index, then sampling real imagery. AI drove every attempt, but I made the call each time: not good enough yet. So it is parked.

A six-panel grid titled "Glacial lake colour: six iterations", each panel the same Bow Lake and Peyto valley coloured by a different method with a caption underneath: distance gradient, condensed ramp, along-drainage, desaturated, glacial index, and imagery sampling. Lake colours range from grey-teal to bright turquoise across the panels.

The crux was water. The 30 m land cover scattered false blue speckle across the icefields. The fix layered real vector lakes and rivers from CanVec, rebuilt permanent snow and ice as a union of vector and raster, and even used glacier retreat, old ice minus new, to keep genuine meltwater ponds.

The Columbia Icefield before the fix, terrain looking right but with false blue speckle scattered across the white glacier and its margins, noise from the 30 m land-cover raster.

The same area with clean vector lakes and rivers from CanVec dropped in, replacing the raster water so shorelines are crisp and correctly placed.

The same area with permanent snow and ice rebuilt as a union of vector and raster sources, giving a clean, continuous white icefield without holes or noise.

The finished water and snow result at the Columbia Icefield: smooth white ice, crisp teal lakes and thin rivers, and no blue speckle remaining.

Colour comes from land cover: the 2020 Land Cover of Canada, as flat earth-tone fills that the grey relief modulates, plus a whisper of texture. Snow and rock come from the same classes, with no hypsometric tint. Every pixel traces to a data class or a named parameter, never a guess.

The basemap with land-cover colour added: green forest, pale scree and khaki alpine ground, white glaciers and teal lakes, all modulated by the grey relief and carrying a subtle hand-made texture.

Then I briefed it like a colleague: the exact area and extent, the UTM projection, the 40 degree sheet rotation... I also told it how I build relief by hand in QGIS: multidirectional hillshade, sky-view factor, slope, and texture shading, blended to grey. It rebuilt the recipe.

Greyscale shaded relief of the Rockies around Bow Lake and the Wapta Icefield, generated from a 30 m elevation model, lit softly from the upper left so ridges and valleys read as sculptural form.

A greyscale slope layer of the same mountains, steep cliff faces rendered dark and gentle valley floors pale, isolating just the steepness of the terrain.

A high-detail greyscale texture-shading layer of the same terrain, picking out fine ridgelines, gullies and rock structure like a pencil relief drawing.

The finished blended relief, soft multidirectional shading with sky-view softening, cliff emphasis and texture combined into one layer with the signature South Arrow sculptural look, carrying a subtle warm/cool cast: warm cream on the lit slopes, cool blue-grey in the shadows and valley depths.

So I took the whole exchange back to Claude, basically telling tales on ChatGPT. Claude turned the lesson into an end-to-end plan: how to go from a real area to a printed sheet, with geometry and text kept deterministic. To manage the data and outputs, I moved to Cowork on desktop.

A flow diagram titled "Area to print: the process Claude wrote", showing ten numbered steps in teal-outlined boxes connected by arrows: concept, geographic frame, acquire data, generalise, shaded relief, basemap surface, water and snow, infrastructure, then greyed-out typography and layout marked as coming next.

It went further, doubting it could do a cartographer's job at all: it can't force the image model to ignore its priors, and "doesn't literally trace or preserve the pixels" I fed it. Its own suggestion was a split: it stylises the textures, I bring the real geography.

A screenshot of my ChatGPT conversation. My message, in the grey bubble at the top, points out that the model is using what it thinks it already knows about the geography rather than the reference I gave it, noting it still labels Indian Arm, a location that does not even appear on the map. ChatGPT's reply agrees this is an important distinction and explains that the image model is not faithfully tracing the screenshot: it recognises the area as Howe Sound near Vancouver, combines that with its own internal visual prior, and produces what it thinks the area looks like. It adds that this is why Indian Arm, Stanley Park and Burrard Inlet keep reappearing, and concludes it "can't force the image model to ignore its priors," since it is a generative model, not a geometric transformation tool.

I handed that style to ChatGPT with a description of the area. It returned a gorgeous image, but invented the geography. Howe Sound had its islands, yet the wrong shape and position, and it dropped in Indian Arm, a real inlet, but nowhere near its true spot up Burrard Inlet.

An AI-generated painterly terrain image of a coastal area, attractive but geographically wrong: an inlet whose islands are present but in the wrong shape and place, and a second arm of water added far from where it really sits.

First I had Claude study five of my own "hand-made" maps and write a full house-style spec: palette, geometric type, shaded relief, flat teal water, the land as the hero. All structured so another AI could build straight to it. It ran to over 200 lines of JSON.

One of my own hand-made South Arrow Maps, of Arthur's Pass National Park, New Zealand, in the same natural style the spec describes: painterly greyscale shaded relief under flat land-cover colour, pale braided river beds winding through beech-forested valleys and alpine tussock beneath the snow-streaked Southern Alps, with the highway and rail line crossing the pass and restrained off-white tracks and labels throughout.

A screenshot-style graphic of a file named south-arrow-style-spec.json, shown as three dense columns of small code on a dark background. Individual words are not meant to be readable; the impression is of a large, highly structured specification with many nested sections such as palette, hillshade, typography and correctness pipeline. A footer reads 214 lines and 18 sections.

Andrew Tyrrell @southarrowmaps.co.nz · 25/07/2026

How far can I get using AI in cartography? I drove Canada's Icefields Parkway last month, and want to make a poster map of it. Can AI help, and do it well? I started on my phone, then realised it needed more than that and moved to the desktop. AI-first: Claude, a bit of ChatGPT. A thread... 🧵🗺️

A zoomed-out view of the whole basemap, roughly 285 by 311 km of the Canadian Rockies in a painterly natural style. Snow-capped ranges and white glaciers run diagonally through dark green forested valleys, with turquoise lakes and rivers threading between them, opening onto flatter green-and-tan farmland in the north-east.

31713

Is it a personal side-project or an ICA Map Design Comm. initiative? If the latter, great! I want to learn about how to better use AI in carto, and see it align with the Comm’s Terms and the projects own aims & questions. If the former, fine, run it how you like, but probably not on the ICA site.

Screenshot from the #365DaysOfMaps kickoff post on mapdesign.icaci.org

mapdesign.icaci.org
The 365 Day Map
Challenge
Posted on December 19, 2025 by @kennethfield
Back in 2015 the Commission committed to providing a critique of a map a day for an entire year which resulted in (actually a few more than) 365 maps with some brief comments about what made them work. The intent was to create an atlas of examples of great map design to help people to find inspiration for their own work.
We're now ten years on from that daily effort and we're going to do it again... but with a twist. Rather than share examples of maps we know, love, and think evidence great design we're going to publish maps that WE make, every single day in 2016.
With due respect to Topi Tjukanov's fantastic grass-roots community led 30 Day Map Challenge we're going to dial it up to 11 and do one a day for an entire year.

Screenshot from the #365DaysOfMaps kickoff post on mapdesign.icaci.org

mapdesign.icaci.org
So... can we encourage better mapping through developing an awareness of how to use Al to make a half-decent map? and can we work towards establishing best practices through our
experimentation?

Screenshot from mapdesign.icaci.org Terms of Reference

mapdesign.icaci.org
Objectives of the ICA Map Design Commission
1. Discuss, develop, illustrate and promote the principles of map design
2. Identify examples of good cartographic practice which underpin high quality, effective map design
3. Provide appropriate mechanisms and support for the dissemination of current map design research
4. Develop high quality materials to support cartographers and map-makers in the development of their own maps (e.g., books, exemplars, tutorials)
5. Develop support for the design of specific cartographic products that are challenged by new modes of production (e.g., design for maps that use imagery as the base)
6. Intorm, enthuse and engage with fledgling map-makers (such as neo-cartographers) in a mutual exchange of methods, needs and ideas
7. Maintain a Web site that includes a bibliographical database of map design research, a community identity for those who work in the area of map design, resources to
assist map-makers in their design endeavors, and a forum for the exchange of ideas and expertise
8. Work with other ICA Commissions which incorporate a design perspective and
9. Have fun designing maps and enthuse others to approach their own work with an eye on design