Start here
No experience with Claude is needed. You will meet Claude and its models, install CascadiaTools in the Claude desktop app, learn every slash command, and run worked examples on your own laptop.
Each station has a short explanation, something to do in Claude, and a check question. Finish a station and it flips to 1 in the detection history at the top left. A browser remembers your progress on that device only; nothing is sent anywhere.
| Block | Stations | What you do |
|---|---|---|
| Meet Claude | 1 and 2 | What Claude is, and why validated tools matter |
| Get ready | 3 to 6 | Install, tour the app, choose a model, manage projects |
| Know the tools | 7 and 8 | Browse the analysis modules, learn every command |
| Analyze | 9 to 12 | Wood Thrush counts, the planted-truth lab, pileated woodpecker occupancy, other taxa |
| Finish the job | 13 to 15 | Figures, power analysis, private mode |
| Your data | 16 and 17 | Write your own prompt, keep the cheat sheet |
Package installation takes 20 to 40 minutes and cannot happen on conference Wi-Fi with everyone at once. Please finish the Install station beforehand.
Simulated data let us compare each estimate with the known truth. No result in this course describes a real population.
The full schedule, pre-work email and rehearsal list are in the instructor station. Turn on instructor notes in the left panel while you present, and off for attendees.
Station 1 · 15 minutes
Claude is an AI system made by Anthropic. You give it instructions in everyday language, and it reads, reasons, writes and, in the desktop app, can work with files and run programs on your computer.
Think of a fast, very widely read new technician. Claude has read a great deal about occupancy models, and it can write R in seconds. It also hasn't seen your study area, can misunderstand you, and can state a wrong answer as confidently as a right one. You would check a new technician's work, and the same applies here.
Where these appear depends on your version of the app. Look for labeled tabs, or for icons at the top left: a speech bubble for chat and </> for code.
| Tab | What it is for | Example |
|---|---|---|
| Chat | A conversation. Ask questions, draft text, think something through. | "Explain closure in an N-mixture model." |
| Cowork | Multi-step knowledge work across files you choose. | "Turn these field notes into a season summary." |
| Code | Claude Code. It works in a project folder and can run programs such as R. CascadiaTools runs here. | "Fit an occupancy model to these detections." |
| Good at | Be careful about |
|---|---|
| Writing and explaining code, and fixing it from error messages | Wrong statistics that run without error. CascadiaTools exists for this. |
| Summarizing and restructuring long documents | Invented details. Always look up any citation it gives you. |
| Following a written procedure step by step | Recent events. Current models have a reliable knowledge cutoff of June 2026. |
| Explaining a method at the level you ask for | Agreeing with you too easily. Ask it to argue the other side. |
| Model | The AI itself. Claude comes in several models, from faster to more capable. |
| Prompt | What you type to Claude. |
| Session | One conversation in the Code tab, with its own history. |
| Context | What Claude can hold in mind during a session. It fills up in long sessions. |
| Agent | Claude working through steps on its own: reading files, running commands, checking results. |
| Skill | A written procedure Claude follows. You can start one with a slash command. |
| Plugin | An add-on that bundles skills and tools. CascadiaTools is a plugin. |
| Marketplace | A list of plugins you can install from. |
| Usage limits | How much work your plan allows over time. Larger models and long jobs use more. |
In a Chat, send the message below. Then look up one of its citations. This takes two minutes and teaches the most important habit in this course.
Give me three peer-reviewed papers, with full citations, on how snow depth affects Pacific marten occupancy in Oregon. Tell me how confident you are that each one exists.
Whatever Claude reads or prints is sent to Anthropic to generate a response. Don't paste sensitive locations into any tab. Private mode covers this in detail, and your institution may have its own terms.
Ask who has used Claude or ChatGPT, then pitch to the newest people in the room. The citation exercise tends to change how people read everything after it. Results vary, so try it yourself first, and don't promise it will invent a paper.
Station 1
You can. It will write plausible code, and it will usually run. Ecological models fail silently, so a model that ran is weak evidence that the answer is right.
An occupancy model with too few surveys can return an occupancy of 1.00 with a tiny standard error. An N-mixture model on data that break closure can return abundance ten times too high with good-looking intervals. Neither prints an error.
You trust unmarked without reading its source because it was written once, published, tested and used by many people. New code written by an AI for each request has none of that history. CascadiaTools gives agents the same arrangement: vetted code, used the same way each time.
| Fresh code from a chat | A CascadiaTools module | |
|---|---|---|
| Statistical code | Written anew each time | Written once, tested, independently verified |
| Shown to work? | Rarely tested | Simulation recovery recorded in VALIDATION.md |
| Known limits | Unknown | Written down, with numbers |
| At your design? | Not checked | Planted-truth check at your sites, visits and detection |
| When it can't be trusted | You may never hear | Verdict of OPEN, with the reason |
You own the question, the biology and the study design. You read what the agent says it is doing, and you confirm the data description. CascadiaTools checks the statistics at your design. It cannot know that your observers changed protocol in June unless you say so.
Audience resistance to AI is the point of this station, so raise it yourself. The manuscript's argument: nobody audits every line of unmarked; we rely on validation and community checking. Show one VALIDATION.md from the repository, then the validation page.
Station 2 · about 30 minutes, once
A computer where you can install programs (university machines may need IT), and a paid Claude plan. Claude Code requires a Pro, Max, Ultra, Team or Enterprise account; the free plan does not include it. R, the compiler and CascadiaTools are free.
The app changes often. Typing works everywhere in the Code tab, one line at a time. The plugin window may open for you to finish the step.
/plugin marketplace add taaltree/CascadiaTools
/plugin install cascadia-tools@cascadia
/reload-plugins
Create an empty folder called cascadia-course in your Documents folder. In the Code tab choose that folder as the project and keep the environment on Local. Type /cascadia and look for the three commands in the menu. Then send:
/cascadia-tools:cascadia-setup
It checks R and the compiler, installs the R packages (20 to 40 minutes the first time, and you can keep working), and finishes with a short test that ends in a report path.
| You see | Do this |
|---|---|
/plugin is not recognized | You are in a Chat tab. Switch to the Code tab, and update the app. |
| No Add plugins in the + menu | Look for Plugins, then Browse plugins. Plugins are not offered in cloud sessions, so check the environment says Local. |
| R not found | Close and reopen the app after installing R. Still missing? Tell Claude where R is, for example C:\Program Files\R\R-4.5.2. |
Windows: errors mentioning bash, or Claude has no Bash tool | Install Git for Windows with the default options, and restart the app. |
| Compiler not found, or Nimble won't build | Install Rtools (Windows) or run xcode-select --install (Mac), restart the app, and run setup again. Likelihood modules work without a compiler. |
| Some packages failed | Run /cascadia-tools:cascadia-setup again. It reports what is missing and which modules that affects. |
The install page repeats these steps with more troubleshooting help. You can also paste any error into Claude and ask what it means.
Do not plan on installing in the room. Screen labels come from Anthropic's documentation (November 2026) and your own install. See the rehearsal list.
Station 3 · 10 minutes
Claude Code in the desktop app works in a folder on your computer. Seven controls cover nearly everything you need.
| Control | Where | Use it to |
|---|---|---|
| Code tab | Top of the app | Everything in this course. The Chat tab has no CascadiaTools. |
| Environment | Under the prompt box | Keep this on Local, so R runs on your computer. |
| Project folder | Beside the environment | Choose one folder per project. Claude works there. |
| + button | Next to the prompt box | Add plugins, browse slash commands, add a folder. |
| / (slash) | In the prompt box | Open the list of commands. Keep typing to filter it. |
| Model | Under the prompt box, toward the right | Choose which Claude does the work. Next station. |
| Mode | Under the prompt box | Set how often Claude asks before acting. |
| Mode | What it does | For this course |
|---|---|---|
| Manual | Asks before commands and edits. Older versions call it Ask permissions. | Start here. You see what Claude runs. |
| Accept edits | Accepts file edits and still asks about other commands. | Fine once you are comfortable. |
| Plan | Proposes a plan before changing anything. | Useful for a big, unfamiliar job. |
| Auto | Runs with fewer prompts, with safety checks in the background. Availability depends on your account and model. | Handy for long runs, after you trust the flow. |
| Bypass permissions | Removes the prompts. | Don't use it with real data. |
In your cascadia-course folder, set the mode to Manual and send the message below. When Claude asks, read the command, then allow it.
Use R to print the R version you can find on this computer, and tell me in one sentence what command you ran.
When Claude asks permission, read it the way you'd read a field assistant's plan. Commands that run core/run.R are the toolkit. If a command wants to delete files or touch folders outside your project, ask why first.
Station 5 · 10 minutes
The model is the AI doing the work. You can leave it on the default for this whole course. Change it when a job is unusually simple or unusually hard.
The model picker is toward the right end of the row under the prompt box, and it shows the current model and effort, such as "Opus 5.5 High". Click it to choose a model. Effort opens from the same menu, and More models lists others. You can change the model during a session, type /model, or press Cmd+Shift+I on a Mac.
| Model | The app's menu says | Anthropic's model docs say | Speed | API price per million tokens (in / out) | Suggested use with CascadiaTools |
|---|---|---|---|---|---|
| Fable 5.1 | For your toughest challenges | For demanding reasoning and long-horizon agentic work | Slower | $10 / $50 | A new module, or a hard multi-day problem |
| Opus 5.5 | For complex work and everyday tasks | For long-running agentic coding and knowledge work | Moderate | $4 / $20 | Everyday analyses. The default, and what we rehearsed with. |
| Sonnet 5.5 | Most efficient for simpler tasks | The best combination of speed and intelligence | Fast | $2 / $10 | Figures, write-ups, follow-up runs of a set analysis |
| Haiku 5.5 | Fastest for quick answers | For high-volume, latency-sensitive tasks such as classification, extraction and routing | Fastest | From $0.10 / $0.50 | Quick questions and simple file tidying |
Descriptions and prices are from Anthropic's models overview, November 2026. Models change, so check the current list. The last column is our suggestion, and only Opus 5.5 was rehearsed. Plan subscribers don't pay these prices directly, but larger models use up plan limits faster. Your account may list other models under More models.
The fitted numbers come from R, using module code that is the same whichever model is selected. What the model changes is how well the agent chooses a module, reads your data, follows the protocol and explains the result. Check its work on any model.
Effort sets how hard the model thinks before acting, and higher levels use more of your limits. Anthropic's advice is to use the default for most tasks. When a result disappoints, ask: did it not try hard enough, or did it not know enough? If it skipped a check, raise the effort. If it tried and was still wrong, move to a larger model. First check that your prompt was clear.
Open the model picker and look at what your plan offers. Switch to Sonnet 5.5, then send:
Which model are you, and in two sentences, what is the difference between occupancy and abundance?
Now switch back to the default.
Keep this short, and tell people to stay on the default today. The price column makes the size of the difference concrete. Fitting a model on a cheaper model is not rehearsed, so don't promise equal results.
Station 6 · 10 minutes
In the Code tab, a project is simply a folder on your computer, and a session is one conversation about it. Keep one folder per study, and start a new session for each new job.
Make a folder for the study and put your original files in data_raw. When you run an analysis, CascadiaTools adds the rest:
marten-occupancy/ the folder you open in Claude
data_raw/ your original files, never changed
projects/marten-winter-2026/
data_raw/ data/ a copy of the originals, then cleaned versions
analysis_plan.md written before fitting
config.yaml the settings that reproduce the fit
recovery/ planted-truth results
results/ report, figures, estimates
log.md what was done, in order, with commandsOn Windows, avoid folders synced by OneDrive or Dropbox for active runs, and paths with unusual characters. We haven't tested synced folders.
| To do this | Do this |
|---|---|
| Open a project | Click the project folder control under the prompt box, then Open folder… |
| Start a new session | + New in the sidebar, or Ctrl+N (Cmd+N on Mac) |
| Return to earlier work | Click the session in the sidebar. It keeps its history. |
| Rename a session | Open it, then click its title in the title bar |
| Tidy the list | Hover over a session, open its menu, and choose Archive |
| Find archived sessions | Hover over a heading in the session list, click Group and sort, then View all and the Archived tab |
| See two sessions side by side | Right-click a session, then Open in and Split view |
| Bring in another folder | + next to the prompt box, then Add folder |
| See how full a session is | Type /context. /compact summarizes the session to free space. |
| See your plan usage | Click the usage ring in the row under the prompt box |
The sidebar's own Projects item is a separate feature, described below. Labels are for the current layout. If a new session shows no environment under the prompt box, you have the earlier layout, and some controls sit elsewhere.
A new session starts fresh, so it doesn't know what you decided last week. Files carry that forward: analysis_plan.md and log.md record the analysis, and a CLAUDE.md file in the project folder is read at the start of each session. Put what stays true about the study there.
In your cascadia-course folder, send:
Create a CLAUDE.md for this folder saying: this is a practice project for learning CascadiaTools; all data here are simulated; never change anything in data_raw. Then show me the file.
Now start a + New session in the same folder and ask: "What do you know about this project?" Rename the first session "Practice", then archive it.
Archiving a session isn't a backup, and Claude can change files in the folder you open. Keep your raw data elsewhere as well, as you do now. If you know Git, a repository for each project also gives you a history of changes.
The Chat tab has its own feature called Projects, for grouping chats and reference documents. In this course, "project" means a folder in the Code tab.
Station 4 · 15 minutes
These are the statistical models. Each module has a tested simulator, a fitting engine, figures, interpretation and a worked example. Filter by your field, read what each is for, and find the one you would use first.
The chip shows the verdict on that module's worked example at its design: . OPEN means at least one quantity the example would report failed a check, and the example says which one and what to do. The rest of its results can often still be used.
OPEN because 6 of 24 quantities failed. Standardized coefficients involving the bird counts ran about 17% too high. The raw paths and rate ratios passed, so the advice is to report those and use the standardized values only to rank routes.
Your own study gets its own verdict at your design. A module that is DONE here can be OPEN for you, and the reverse.
Describe a study you have run. Claude reads the catalog and names the modules that fit.
/cascadia-tools:cascadia-start I work on [species or system]. I collect [what you record]. Which modules fit, and what data would each one need?
Prefer clicking? The module finder asks two questions and gives you a starting prompt.
Station 5 · 25 minutes
A slash command loads one set of instructions and hands it your text. All 14 CascadiaTools commands are below, along with those that manage plugins.
/ at the start of the prompt box, or click + and choose Slash commands./eco shows the protocols. Pick one, and it appears highlighted./cascadia-tools:eco-power. The short /eco-power also works unless another command uses that name.eco-power. The slash guarantees it.The one you will use most is eco-analyze, which calls the protocol commands in order. Use the others to run a single step, to repeat a step, or to ask for that step by name when you want it held to its rules.
Eight situations from a season of fieldwork. Score carries through to the end of the course.
In your course folder, run the orientation:
/cascadia-tools:cascadia-start I'm new here and I work on forest carnivores.
Then see how a request reaches the same skill without a slash. Start a new session and send:
What is your standard for calling an analysis done?
Claude should describe the acceptance gate. If its answer isn't specific, send /cascadia-tools:eco-acceptance-gate and compare the two answers.
Don't walk through all the cards aloud. Demonstrate the /eco filter, explain the two families (eco-analyze runs everything; the others run one step), then let people do the match game in pairs for 5 minutes.
Station 6 · 25 minutes, hands-on
A simulated point-count study: 45 stations, 4 visits each, some counts 5 minutes and most 10. Does the number of thrushes change with canopy cover, elevation and forest type?
glm-example folder as the project./cascadia-tools:eco-analyze I have Wood Thrush point counts in wood_thrush_counts.csv: 45 stations, each counted 4 times in May-June 2025. Most counts were 10 minutes, some were 5. I want to know whether the number of thrushes changes with canopy cover, elevation and forest type. Please analyse it and give me a report with figures I can use in a paper. I'm on a laptop, so use 4 cores.
Rehearsal: about 4 minutes on a cluster node, with no pauses for approval. Allow 15 to 25 minutes on a laptop with permission prompts. The laptop figure is an estimate.
| Effect | Rate ratio (95% CI) | Reads as |
|---|---|---|
| Canopy cover, per 10 points | 1.39 (1.18 to 1.64) | 39% higher expected count |
| Elevation, per 100 m | 0.949 (0.854 to 1.05) | Compatible with no effect |
| Conifer vs deciduous | 0.482 (0.248 to 0.935) | 52% lower expected count |
| Mixed vs deciduous | 1.15 (0.651 to 2.01) | Compatible with no effect |
The report says "is associated with" and avoids "affects". It avoids "significant" and gives size and interval. It says these counts combine abundance with detectability. Those are rules in eco-interpretation, applied every time.
This example is DONE even though two recovery rows failed: the station random-effect SD (biased 18% low) and the negative-binomial size parameter. Neither is interpreted, so neither blocks the verdict, and both are named as limits. The gate applies to the quantities you report.
If Claude wrote a new model of its own and didn't flag it, or gives you results without a verdict, ask it to follow the gate. Custom work has to be cross-checked with a second implementation (gate step 4).
Rehearsed once on Linux. Point estimates matched this page (canopy 1.39, conifer 0.48, mixed 1.15), and the verdict was DONE with limits. Run time 4 min 4 s, on up to 16 cluster cores. Claude Code reported $2.28 of usage at API-equivalent rates. It wrote some project scripts of its own (an exact-design recovery and sensitivity fits), which makes question 5 a good live discussion. It also read README_data.md, which contains the true values, and said so. For a blind demonstration, delete README_data.md, WALKTHROUGH.md and config.yaml first.
Station 7 · 20 minutes, in this page
You set the true occupancy and detection. The page simulates surveys, fits an occupancy model to each one, and applies the same gate CascadiaTools uses. Find a design that fails.
A teaching simulation in your browser: constant ψ and p, maximum likelihood, Wald intervals on the logit scale, seeded so results repeat. Real modules handle covariates and missing visits, and the thresholds are fixed in core/R/recovery.R.
| Check | Passes when | What failure means |
|---|---|---|
| Bias | |bias| ≤ the larger of 10% of truth or 3 Monte Carlo SE | The method systematically over- or undershoots |
| Coverage | 95% intervals contain the truth at least 0.95 − 3 Monte Carlo SE of the time (floor 0.80) | Intervals are too narrow, so you would be overconfident |
| Convergence | At least 90% of all attempted fits converged | The model often can't be fitted at this design |
The naive line is the negative control. Counting a site as occupied only if you detected the species ignores missed birds, so it fails unless detection is high enough that few sites are missed. A check that can't fail tells you nothing.
Then the verdict: DONE only if every step passed; otherwise OPEN, with what failed and what it means.
Here is the planted-truth step by itself, for a thin design. It should take a few minutes on a laptop.
/cascadia-tools:eco-planted-truth For the occu-single module, check recovery at 40 sites with 2 visits each and low detection. Use 4 cores and 50 replicates so it stays short. Tell me which quantities pass the gate, and whether 50 replicates is enough to rely on.
The acceptance gate asks for at least 100 replicates, or 50 if each fit takes minutes. Treat this short run as practice.
Project this. Run the pileated design (should pass), then the thin design, then ask the room to predict before each run. A useful question: "How many more visits does this need?" Then let them hunt for the smallest passing design.
Station 8 · 25 minutes, hands-on
Call-playback surveys at 120 stands, up to 4 visits each. Does forest cover or elevation affect occupancy, allowing for birds that were missed?
occu-single-example folder as a new session./cascadia-tools:eco-analyze I have pileated woodpecker call-playback surveys from 120 forest stands, up to 4 visits each in May-June 2025. The files are detections.csv (one row per visit: stand, visit number, date, day of year, observer, detected 1/0) and sites.csv (forest cover %, elevation in m, stand age). Does forest cover or elevation affect whether woodpeckers occupy a stand? Please account for birds we missed, and allow detection to change with date and observer. I need results and figures for a paper. I'm on a laptop, so use 4 cores and tell me how long the recovery check will take before you start it.
Rehearsal with this exact prompt: 4 min 39 s to the planted-truth stop on Opus 5.5. The recovery check itself took 1 min 46 s on 4 cluster cores; a laptop may be slower. Allow more time for the fit, the report and permission prompts.
This is the course's most useful surprise. The forest cover slope really is about 11% too strong at this design, just past the 10% limit, so most short recovery checks fail it, and the protocol tells Claude to stop and ask you. If it stops, send:
This is a teaching example, so please continue. Don't change the gate. Fit the data, treat the slope sizes as a stated limit, and give me the report and your verdict.
Claude may also tell you the files are a simulated example and that it read the answers in them. That's accurate. It should not change a threshold or rerun with new seeds to get a pass. If it offers to, say no.
This is the agent's first deliverable, and the one most worth your attention. It should report: 120 stands, 405 surveys, 72 stands with 4 visits, 21 with 3, and 27 with 2. Missing visits have no row and are not zeros. Woodpeckers were detected on 52 surveys at 34 stands.
| Quantity | Value | What it tells you |
|---|---|---|
| Naive occupancy | 0.28 (34/120) | Detected at least once. Ignores missed birds. |
| Stands actually occupied | 47 (37 to 58) | Applies to these 120 stands only |
| Detection across 4 visits | 0.87 (0.70 to 0.96) | About 13% of occupied stands go undetected |
| Detection per visit, by observer | AB 0.39, JK 0.13, MR 0.52 | Observer differences matter. This is detection, not habitat. |
| Goodness of fit | c-hat 1.31, p = 0.138 | No evidence of lack of fit, though the test has low power here |
These are real rows from a large planted-truth run of this example's full model: 5,000 replicates at this design.
| Quantity | Truth | Mean estimate | Rel. bias | Coverage | Gate |
|---|---|---|---|---|---|
| Occupancy, average site | 0.40 | 0.40 | +0.7% | 0.954 | PASS |
| Detection, average visit | 0.40 | 0.40 | −0.6% | 0.948 | PASS |
| Share of stands occupied | 0.42 | 0.43 | +2.4% | 0.958 | PASS |
| Forest cover slope (logit) | 1.02 | 1.13 | +10.8% | 0.968 | FAIL |
| Elevation slope (logit) | −0.515 | −0.572 | 11.1% too strong | 0.958 | FAIL |
| Run | Replicates | Forest slope bias | Monte Carlo 95% interval | Gate |
|---|---|---|---|---|
| First published run | 200 | +9.6% | roughly 5 to 14% | PASS |
| Course rehearsal | 200 | +12.7% | roughly 8 to 17% | FAIL |
| Large run, set up to settle it | 5,000 | +10.8% | 9.7 to 12.0% | FAIL |
A recovery run is itself an estimate. The first run passed by chance, because the true bias is close to the limit. When a rehearsal failed, we ran 5,000 new replicates, with the decision rule written down first, and the example's verdict changed to OPEN. The rule stayed where it was. We then did the same for eight more examples near the limits, and three more changed to OPEN. See the validation page.
The science is unchanged, and it is what you should report: the direction can be trusted, intervals are reliable (coverage 0.968), occupancy and detection pass, and slope sizes are somewhat too strong, by about 11% on average and 6% at the median. An elevation effect of this size was detected in only about half of replicates, so a null result for elevation at this design would say little.
/cascadia-tools:eco-interpretation Write the results paragraph for my paper. Then tell me which sentence a skeptical reviewer would push on, and what the recovery table says about it.
This run usually stops at planted truth, because the true bias is about 10.8%. With 200 replicates, roughly a third of runs would still land under the limit, so expect a split room, and use it: it shows the agent reporting a failure rather than hiding it. The follow-up prompt in the box was blocked from rehearsal, so try it yourself. A direct fit of this example from the downloaded zip took 18 s on 4 cores, and the JWM greyscale settings on the Figures station were run directly on that fit (PDF and a 600 dpi TIFF, 85 mm wide). The eco-figures prompt itself was not rehearsed through Claude. The module's own 5,000-replicate results are in its validation record.
Station 9 · 15 minutes, pick one
Try the example nearest your own work. Each zip has a START_HERE.txt with the prompt. Open the unzipped folder as a project and paste it in.
Bayesian and telemetry examples take longer. The times below are for one fit on the lab cluster, and the recovery check multiplies them by 100 or more.
All 29 examples, with every prompt and result, are on the examples page and in Appendix S1.
Station 10 · 15 minutes, hands-on
Every fit exports a figure suite to figures/ with a gallery.html. You change figures by asking; the model isn't refitted.
In the same session as your woodpecker run:
/cascadia-tools:eco-figures Prepare the forest cover occupancy figure for the Journal of Wildlife Management at single-column width, greyscale. Label the x axis "Forest cover (%)" and give me a PDF, a 600 dpi TIFF, and the caption. Show me the config block you used.
| You say | The config line |
|---|---|
| "Format this for JWM" | journal: wiley_jwm (also esa, besjournals, cjfas, plos, nature, slide) |
| "Make it two-column" | per_figure: {<id>: {width: double}} |
| "Black and white for print" | greyscale: true |
| "Show percent, not z-scores" | units: {forest_cover_pct: {center: [mean], scale: [SD]}} |
| "Call it Forest cover (%)" | labels: {forest_cover_pct: "Forest cover (%)"} |
| "Only these two figures" | include: [occupancy_forest_cover_pct, coefficients] |
| "PDF and TIFF at 600 dpi" | formats: [pdf, tiff] and dpi: 600 |
| "A big version for my talk" | journal: slide |
The figures live in a config file, so you can reproduce them and hand them to a co-author.
Restyling must not change what a figure means. If you ask to drop the uncertainty band, cut the axis to exaggerate an effect, or use a dual axis, the skill tells it to explain why that misleads and offer an honest alternative. Try it: "Remove the confidence band, it looks messy."
Station 11 · 15 minutes
eco-power answers "how many sites?" by simulating your design many times, using the module that will later analyze the data.
/cascadia-tools:eco-power I want to monitor a forest owl with call-playback surveys. Occupancy is about 0.4, detection per visit about 0.3, and I can afford about 600 visits total. Can I detect a logit slope of 0.5 per SD of forest cover? Compare 100 sites with 6 visits, 150 with 4 and 200 with 3. I'm on a laptop, so use 4 cores, keep it under 20 minutes, and tell me how precise the result will be with that many replicates.
200 replicates give power to about ±0.03 near 0.8. In a short workshop run you will have fewer, so read it as a rough answer.
You don't need data yet, nothing sensitive leaves your computer, and the answer can save a field season. Many first-time users should start here.
A live power run will outlast this station. Start one at the beginning of the break and return to it, or show the module's recorded power numbers. The fixed-effort comparison in the prompt usually sparks discussion among people who run surveys.
Station 12 · 10 minutes, hands-on
R runs on your computer. The AI model does not. Anything Claude reads or prints goes to the model provider.
| Normally | With private mode |
|---|---|
| Claude can open your CSVs and see rows, coordinates and IDs | Claude sees column names, dimensions, value ranges, missing-value counts, check messages and aggregated results |
| Estimates and figures go through the conversation | Summaries, column names, error messages, estimates and figures are still sent |
Use a practice folder containing a copy of an example's CSVs in a subfolder called data_raw.
/cascadia-tools:cascadia-private My sensitive files are in data_raw.
Then try to break it:
Show me the first five rows of the file in data_raw.
Claude should be refused and should say so. Always run this test on your own computer before relying on private mode.
To turn it off, ask Claude to remove private mode for the project. It deletes .cascadia/private.json and the deny rules.
Spotted owl nests, wolf dens and tribal-agreement data will come up. Be plain that this reduces exposure and doesn't guarantee anything. See the privacy page. On Windows, test the block in rehearsal: the guard hook calls python3, which may be absent.
Station 13 · 20 minutes
Your prompt needs five things: what was recorded, how the files are organized, the question, what you need back, and what the computer can do. Edit the example below.
The filled-in text above is an example. Replace it with your own study.
OPEN is useful information, and you have options. Ask:
/cascadia-tools:eco-acceptance-gate Explain this OPEN verdict. Which results can I still report, what would change the verdict (more visits, a simpler model, a different module), and how should I describe the limit in my paper?
Claude should say so rather than force a poor fit. /cascadia-tools:eco-new-module can build a project-specific module, starting from the primary literature, with its own simulator, recovery runs, negative controls and cross-check. Expect that to take hours or longer, and have a statistician look at it. Results are unverified until they pass the gate.
Bayesian recovery can take hours on a laptop. Plan to leave it running overnight, or use a lab server: the Code tab's SSH environment supports plugins, and on a SLURM cluster the toolkit's job wrapper submits the jobs. Compute time depends on the module, so ask Claude to run a short timing test first and tell you the estimate.
Station 14
| I want to | Type |
|---|
Third-party plugins don't update automatically by default. Type /plugin, open Marketplaces, select cascadia, and choose Enable auto-update or Update marketplace. A running session keeps its loaded version until /reload-plugins or a new session.
Levi, T. CascadiaTools: validated analysis modules and protocols for AI-assisted ecological analysis. Cascadia Wildlife Lab, Oregon State University. https://github.com/taaltree/CascadiaTools
Cite the statistical packages too; each report lists them. In your methods, say that an AI agent ran the analysis, and include the module, the recovery results and the verdict.
/cascadia-tools:eco-analyze with your own data, or /cascadia-tools:eco-power before your next season.Instructor station
A schedule for about 3 hours, with a 90-minute cut. The length is a suggestion, because no slot length was given; change it to fit the program.
| Station | 3 h | 90 min | Format |
|---|---|---|---|
| Welcome, Meet Claude | 15 | 8 | Talk, citation exercise |
| Why not just ask? | 10 | 5 | Talk and discussion |
| Install check | 5 | 3 | Hands up: who has a report path? |
| Code tab, models, projects | 15 | 7 | Projected demo |
| Modules | 10 | 3 | Find your module |
| Slash commands | 15 | 7 | Demo, then match game in pairs |
| Wood Thrush | 25 | 25 | Hands-on |
| Break (start a power run) | 15 | 0 | |
| Planted-truth lab | 15 | 15 | Projected, predict before each run |
| Woodpecker occupancy | 25 | 5 (show) | Hands-on (90 min: read results only) |
| Other taxa | Skip or homework | Skip | Self-paced |
| Figures | 10 | 3 | Hands-on |
| Power | 5 | 2 | Results of the break's run |
| Private mode | 10 | 5 | Demo |
| Your data, cheat sheet | 5 | 2 | Prompt builder, questions |
Send this at least a week ahead, and again 3 days before.
Subject: Before the CascadiaTools tutorial: 45 minutes of setup Hello, Please set up your laptop before the tutorial. We won't have time or bandwidth to do it in the room. You need: - A Windows or Mac laptop on which you can install software (ask IT early if yours is managed). - A paid Claude plan that includes Claude Code. The free plan doesn't include it. Plans and prices are at claude.com. Follow the Install station of the course (about 30 minutes, mostly waiting): [COURSE LINK]#install The last step installs R packages, which takes 20 to 40 minutes. You can keep working while it runs. It's finished when Claude gives you a report path. If you can't get through it, reply with the error message or come 30 minutes early. If you don't have a Claude plan, you are welcome to pair with a neighbor. Please don't bring sensitive data to the session. We'll use simulated examples.
Run on the lab's Linux cluster, using the plugin and the website's example zips: Wood Thrush through Claude (the run was accidentally sent with the command and no prompt text, so Claude took the prompt from START_HERE.txt; the laptop wording was not tested there); woodpecker through Claude with the exact prompt, to the planted-truth stop; the setup self-test (11 s); a direct woodpecker fit; the JWM figure settings; and the private-mode guard, against 23 cases that matched the previous guard, plus PowerShell commands. Both Claude rehearsals used Opus 5.5 and predate the 0.2.1 fixes. Not tested: the desktop app itself, Windows, macOS, a clean package install, a real laptop, other Claude models, the follow-up and eco-figures prompts through Claude, the prompts on the Planted-truth, Power, Meet Claude, Model and Projects stations, the private-mode skill from end to end, and most slash commands run live. Screen labels and diagrams come from Anthropic's documentation (November 2026) and your own install.
Version 0.2.1 fixes what we could find, but none of it has run on Windows. Fixed and tested on Linux by simulating Windows conditions: the run scripts now find R in its usual Windows folders; parallel recovery and power runs use socket workers, because Windows has no forking (same estimates as forked workers, to the last digit, on two modules, and a Nimble module ran without errors); and the private-mode guard runs in R, covers PowerShell, and refuses tool calls if it can't run. Still required: Git for Windows, which provides the Bash shell the run scripts need. Socket workers also start more slowly, and each one compiles its own Nimble model, so Bayesian runs will take longer on Windows.
These diagrams stand in for screenshots, because the course was built on a cluster that can't see the desktop app. Each one is replaced automatically when an image with the right name is added to tools/course/img/ on the website. No other edit is needed.
| File name | What to capture |
|---|
If you use computer use in the desktop app, this prompt starts the job. Review each image yourself before adding it.
| Problem | Plan |
|---|---|
| Someone's install isn't finished | Pair them with a neighbor. The Meet Claude, Modules and Slash commands stations and the planted-truth lab work without an install. |
| Wi-Fi is overloaded | Have the two example zips on USB sticks. Claude still needs a connection, so tether your own machine and project it. |
| Rate or usage limits | Switch people to reading the example results on the website, and continue with the lab. |
| A live run stalls | Open your rehearsal copy of the report. Keep finished results from each example on your machine. |
| A hands-on run gives different numbers | Good discussion. Same data and config should match closely; ask Claude what differed. The agent's route can vary, so check the module and config it used. |
Whatever Claude reads or prints does. R itself runs locally. Private mode narrows what is read. Terms depend on the account type, so check your institution's terms.
Today the packaged plugin targets Claude Code. The modules are ordinary R code and the protocols are written instructions, so other agents could use them, but that route hasn't been packaged or tested.
Because we tested the borderline ones harder, and reported what we found. Most OPEN verdicts name one quantity or a slightly narrow interval, and the rest of the example can still be used. Examples not rerun at large size haven't had this stricter test.
Independent verifier agents reviewed each module against recorded recovery results, and problems found and fixed are tallied on the validation page. Humans should keep checking. The code is public, and problem reports are welcome.
The engines are established packages (unmarked, glmmTMB, Nimble, secr and others), and the report gives a methods paragraph with citations. Disclose the agent's role and report the recovery results. Journal AI policies vary.
No. It handles standard models with checks. Hard design questions and new models still benefit from one, and the OPEN verdicts make a good agenda for that conversation.