How shared context, codified workflows, and team contributions compound what the Lab knows.
Net Zero Industrial Policy Lab · Johns Hopkins University
2 / 27
The problem: knowledge stays local
We use Claude independently. Every session starts from zero.
When one person discovers that DataMexico's /data endpoint silently drops manufacturing sectors, that knowledge lives in their session. The next person hits the same wall.
Today
Person A discovers an API gotcha. Person B discovers the same gotcha two weeks later. Person C discovers it next month.
With shared context
Person A adds a one-liner. Person B and C never hit the problem. Their Claude sessions already know.
Same for data sources, methodology patterns, reusable commands, and design standards. Every insight that stays local is an insight the team rediscovers.
3 / 27Interactive · Zoom poll
Where are you today?
Quick read on the room. Pick the one that best describes you right now.
A
I use Claude.ai (or ChatGPT) in a browser for writing and research.
B
I have used an AI coding tool in a terminal or IDE.
C
I run pipelines, build prompts, or set up automations with Claude.
D
I have never used Claude or anything like it.
No wrong answer. The session is calibrated so that whichever letter you pick, you leave with something useful.
4 / 27PrimerFor anyone new to Claude
What Claude is (and isn't)
Level-set before we talk about how we use it as a Lab.
Claude is an AI assistant that makes it easier to talk to a computer. You describe what you want in plain English — no code to write, no command syntax to memorize, no buried menus to navigate — and Claude translates that into the underlying work: reading files, pulling data, drafting text, running an analysis. Useful framing: a very well-read generalist who also speaks every technical language the Lab uses, joined yesterday, no institutional memory, has to be briefed every time.
Good at
Turning a plain-English question into working code, a query, or an analysis
Working with data you give it — summarizing, reshaping, comparing, spotting patterns
Moving between formats (CSV ↔ JSON, table ↔ chart, prose ↔ outline, EN ↔ ES)
Explaining unfamiliar concepts, codebases, or datasets in accessible terms
Drafting a first version of almost anything, fast
Structured reasoning when you supply the structure
Limited at
Facts after its training cutoff — confirm with a live source
Exact arithmetic on anything nontrivial — have it write the calculation, don't trust the mental math
Citing specific papers without source material in front of it — will hallucinate references
Things you didn't tell it about — unpublished work, internal context, sensitive data
Replacing your judgment on contested questions
The kit fills the "didn't tell it about" gap — for the Lab.
5 / 27PrimerFor anyone new to Claude
How to talk to Claude effectively
Five moves that improve most prompts.
1
Be specific about the data and the output you want
Analyze this data.Open nzipl_mex_state_rca.json. Return the top 5 states for EV Components by RCA, as a markdown table with state name, RCA, exports (USD), and product count. Sort descending.
2
Give context before the ask
Is this pipeline right?This Python pipeline pulls DataMexico /stats/rca to get state-level manufacturing employment by SCIAN code. I'm worried the multi-month parameter isn't doing what I think. Walk me through what each request returns and where the aggregation happens.
3
Iterate — the first answer is a draft
Build me a dashboard.First pass: a single HTML page with Leaflet, Mexico states shaded by RCA for grid hardware, using the design tokens from nzipl-design. Keep it minimal. We'll iterate on layers, filters, and click behavior after I see the base map.
4
Ask it to cite its sources and show reasoning
Why is Nuevo León's RCA high?Using nzipl_mex_state_rca.json, explain Nuevo León's RCA for grid hardware. Point to the specific HS6 products driving the score and the exports that anchor them. Flag anything ambiguous or anything that might be a data artifact.
5
Tell it what good looks like
Make me a chart.Build a horizontal bar chart of the top 10 EV Components exporters by state using D3 v7. Match the chart style in nzipl-playcard-mexico-ev-components.html — dark background, green bars, Archivo font, animate on scroll using the observer pattern the other cards use.
Bad answers are usually under-prompted. Good answers are specific, contextualized, and iterated.
6 / 27Act 1 · Shared Context
Claude Team: what it solves and what it doesn't
A Team subscription gives everyone access. But access is not the bottleneck.
What a Team plan solves
Shared billing. One invoice, one admin, no personal subscriptions.
Higher usage limits. More messages, longer context, priority access.
Projects. Upload documents to a Project and Claude remembers them across conversations.
Web access. Everyone can use claude.ai for writing, research, and analysis from day one.
What it does not solve
No shared memory. One person's conversation history is invisible to another's.
No code execution. Claude can write code but can't run it. No pipelines, no file generation.
No repo awareness. Claude doesn't know what's in our codebase, data files, or deliverable architecture.
No compounding. Each session starts fresh. Team knowledge doesn't accumulate.
The Team plan is the foundation. What we build on top of it is what matters.
7 / 27Act 1 · Shared Context
What Claude shouldn't see
Privacy rules for a research Lab.
OK to share
Published data and papers
Public trade data, public policy documents
Anonymized methodology
Your own draft thinking
The kit's content (shared by design)
Never share
Unpublished findings shared in confidence
Client or partner data marked confidential
Unsigned grant agreements or proposals
Personal data about identifiable individuals
Anything covered by NDA
The Claude Team plan does not use your conversations to train the model. Anthropic has committed to zero training on Team and Enterprise inputs. Model behavior is not the same as data handling — treat anything you paste as if it could leave your screen. When in doubt, ask before pasting.
8 / 27Act 1 · Shared Context
Claude Code: the step beyond chat
A version of Claude that lives on your computer, next to the files you work with every day. Point it at a project folder and it reads what's there, writes new files, and runs code for you — no copy-paste loop between a browser and your laptop.
Reads your project folder
Claude sees the data files, documentation, code, and deliverables in the folder. It understands the whole project, not just what you paste into a chat box.
Runs code for you
Python pipelines, data processing, small scripts. Results come back in the same conversation. You don't need to run anything yourself.
Writes and edits files
Generates HTML deliverables, updates JSON configs, creates charts. The output lands directly in the project folder — ready to open, share, or commit.
Reads a briefing file on start
CLAUDE.md is a plain-text file at the top of the project folder. Think of it as the note the Lab leaves for any new analyst on their first day: data sources, vocabulary, methodology, what to watch out for. Claude reads it before you say anything, so every session starts already briefed.
Included with every Claude Team subscription. No separate license.
9 / 27Act 1 · Shared Context
How Claude Code learns from a repo
Three mechanisms. Each loaded automatically when Claude Code opens in the repo directory:
CLAUDE.md
Auto-loaded context. Data sources, API patterns, deliverable standards, glossary, methodology.
→
Skills
Invocable expertise. Design system, chart styles, analytical templates. Run with /skill-name.
→
Commands
Reusable prompts. Structured tasks anyone can run, producing consistent output. Run with /command-name.
The content of these files is the Lab's collective knowledge. When anyone opens Claude Code in this repo, they inherit everything the team has documented.
10 / 27Act 1 · Shared Context
What's in the kit
One repo. Clone it and your Claude sessions know the Lab.
File
Content (actual snippet)
CLAUDE.md
Lab context loaded on session start.OEC API requires three headers plus browser User-Agent (Cloudflare protection).
glossary.md
31 acronyms + 22 internal terms.Relatedness density: mean proximity of a play's target products to a location's existing exports. Forward-looking complement to RCA.
Running log, append-only. Primary way the kit grows (next slide).
.claude/skills/nzipl-design/
CICE design system: tokens, HTML patterns, chart styles, PPTX layouts. This deck uses it.
The structure is scaffolding. The team fills it. Not one person's work product; a shared resource that gets better as more people contribute.
11 / 27Act 1 · Shared Context
discoveries.md
One line, push, done. That is the contribution mechanism.
Not just errors. Data sources, patterns, commands, methodology insights. Anything worth sharing.
discoveries.md · actual entries
2026-04-07 | ggv | DataMexico /data drops manufacturing sectors 31-33; use /stats/rca with multi-month parameterpipeline:denue
2026-04-08 | ggv | FDI enrichment: one web search per row covers 5-7 of 8 source columns; don't search per-columnpipeline:enrichment
2026-04-21 | you | Your next discovery goes here — one line, any teammate can addpattern
The tags are optional. In practice, just write a one-liner. The cost is 30 seconds. The return is permanent: every Claude session from that point forward knows what you wrote.
One thing you rediscovered in the last month. A data source, an API quirk, a methodology pattern, a naming convention. One line is enough.
We'll read 2-3 aloud and reframe each one. Each is a candidate entry for discoveries.md. Today it lives in your head. With the kit, it becomes a line the whole Lab inherits.
No judgment. Half-formed thoughts welcome. The point is to feel the difference between "I know this" and "the Lab knows this."
13 / 27Act 1 · Shared Context
Example: /enrich-fdi
A task that started as one person's need. Now it's shared infrastructure.
The fDi Markets dataset has 698 FDI project rows. Each needs independent source URLs from press releases, trade publications, and government announcements. Manually, this takes hours.
Input
xlsx file with FDI project data Columns A-O: company, sector, city, investment, year Columns P-W: empty (need source URLs)
Output
Same xlsx, columns P-W filled Each cell: URL to confirming article Reuters, PV Magazine, Electrive, press releases
Anyone can run /enrich-fdi from the kit. The command knows the column map, search strategy, quality rules, and fallback logic. The person running it doesn't need to know any of that.
14 / 27Act 2 · Shared Platform
From deliverables to platform
The Lab already produces play cards, infrastructure maps, and cross-country comparators. The question is whether they stay bespoke or become a system the team operates.
Behind every deliverable is a stack: trade data pulled, RCA computed, infrastructure layers assembled, charts generated. That stack can be parameterized so any team member produces analytical output for any country and play.
Knowledge
The kit (Act 1). Glossary, gotchas, discoveries, design system, commands. Makes the other three layers accessible.
Visualization
Play cards, infrastructure maps, play selectors. Self-contained HTML, scrollytelling, D3, Leaflet.
These layers exist for Mexico. The question is whether they stay as one person's toolbox or become something the team operates.
15 / 27Act 2 · Shared Platform
Where we are
Mexico proved the architecture end-to-end. The rest is a roadmap, not a promise.
Mexico: three play cards, an infrastructure map with 15 layers, a play selector, a working Atlas Bot prototype. Meanwhile CICE serves cross-country comparisons, WP4 maps competitiveness across 155 countries, and the team uses Claude for research and writing. None of this starts from zero.
Phase 1 · this summer
Systematize Mexico + Brazil
Country pipeline configs. /play-card --country=brazil --play=solar as the goal command.
Phase 2 · fall
Extend to India + CICE
India data + play cards. Formal export pipeline from the analytical layer into CICE's data format.
Phase 3 · winter
Atlas vision
Subnational depth for CICE. Process chains. Temporal dimension. Policy interface. Details in appendix.
Current workflow
One person runs the pipeline, embeds the data, builds the deliverable, publishes. Single point of production.
Platform workflow
A team member runs /play-card --country brazil --play solar. Claude handles pipeline, data, template. The team member reviews, refines, adds local knowledge. Multiplied production.
16 / 27Act 3 · How the team adopts this
Levels of Claude adoption
Most of the team is at Level 0. The kit moves everyone to Level 1 on day one.
0
Chat
Claude Pro or Team, web interface. Writing assistance, research synthesis, data interpretation. Stateless: every session starts from zero.
1
Kit-informed
Clone the kit repo, open Claude Code. Sessions load Lab context automatically. Every answer uses the right vocabulary, data sources, and methodology.
2
Pipeline
Run data pipelines and generate deliverables via commands. Claude handles the code; you review and refine the output.
3
Contributor
Append discoveries, propose gotchas, improve commands. The kit gets smarter because you used it.
4
Agent
Scheduled tasks that run without prompting — data freshness, consistency audits, deliverable QA. Aspirational. Later in the year. Not the ask today.
Nobody needs to jump to Level 4. Level 1 is free, immediate, already valuable.
17 / 27Act 3 · How the team adopts this
Level 0 vs. Level 1
The difference is context. The effort is zero.
Level 0: Generic Claude
You:What is Mexico's comparative advantage in solar equipment?
Mexico has a growing solar industry... [generic overview, no HS6 codes, no subnational detail, no Lab methodology]
Level 1: Kit-informed
You:What is Mexico's RCA for solar supply chain products?
Running S0 methodology against BACI 2022. Mexico has RCA > 1 for several HS6 products in the solar supply chain. Pipeline: nzipl_playcard_pipeline.py --play solar
Same user. Same question. The only difference: the kit repo was cloned. Claude loaded CLAUDE.md on session start and answered in the Lab's analytical frame.
18 / 27Act 3 · How the team adopts thisLive demo · 2 min
Watch it happen
Same prompt, two places. One has the kit loaded. One doesn't.
The prompt, both times
What is Mexico's RCA for grid hardware?
A volunteer · Claude.ai, no context
I'll ask one person in the room to open Claude.ai in a fresh browser, paste the prompt, and read the answer aloud.
Expect: a general description of RCA, some high-level language about Mexico's industrial base, no specific numbers tied to our taxonomy.
Me · Claude Code inside the kit
I'll run the same prompt from the kit folder on my machine.
Expect: RCA 3.02, $16.8B exports, top states (Nuevo León, Sonora, Baja California), references to the underlying data file — because CLAUDE.md briefed Claude on the Lab's methodology before the question was even asked.
Same person could run either side. The only variable is whether the kit is loaded.
19 / 27Act 3 · How the team adopts this
Research use cases
Six workflow stages where a Lab-informed Claude earns its seat at the table.
Data integration
"Pull this World Bank csv and this DataMexico file. Normalize the country codes, align on year, and produce a single tidy dataset we can use to compare EV components trade."
Data analysis and sense-making
"Why might Nuevo León's RCA for grid hardware have dropped between 2019 and 2023? Check the kit's HS6 taxonomy, flag the products doing most of the work, and propose two candidate explanations."
Visualization
"Suggest three chart types for comparing subnational RCA across 32 Mexican states for three plays. Build the strongest one in D3 using the Lab's design tokens."
Rapid prototyping
"Sketch a one-page Leaflet map of Mexican industrial parks colored by net-zero readiness. Use nzipl_industrial_parks.json. Working prototype I can click through, no polish yet."
Deep research
"Scan the last 18 months of published work on barriers to EV component manufacturing in Mexico. Summarize the three main arguments, name the most cited sources, and flag where the evidence is thin."
Peer review of your own work
"Read this draft. Where would an economist from a different school of thought push back? Where are my assumptions unexamined?"
Pick one. Try it next week against something already on your plate.
20 / 27Act 3 · How the team adopts this
Level 2: pipeline users
Run the analysis. Review the output. Refine what matters.
# The command knows the column map, search strategy, quality rules.
# The user didn't write a single line of code.
A command wraps expertise into a reusable prompt. The critical skill at Level 2 is not coding; it is knowing what the output means and whether it is right. That is analytical judgment, which the Lab already has.
21 / 27Act 3 · How the team adopts this
Level 3: the norm that makes it work
The kit grows because the team uses it. One line at a time.
When you hit an API quirk
Add a line to discoveries.md. The next person's Claude session inherits it. You just saved a colleague two hours.
When you find a better data source
Append to discoveries.md with the endpoint, auth details, any quirks. Significant ones get promoted to the glossary at quarterly review.
When a task recurs
Draft a command following the /enrich-fdi pattern. Save to .claude/commands/. Everyone can now run it.
When Claude misreads the kit
That is also a discovery. Note which file confused it. The kit is wrong, not Claude. Fix the kit.
How you actually do it
Add a line to discoveries.md
discoveries.md · append at the bottom
2026-04-21 | your-initials | The thing you learned, in one sentence
Open the file in any editor, type the line, save, push. If Git feels like a hurdle, ask Claude: "add this as a new entry in discoveries.md and commit it." Claude does the rest.
Draft a command
.claude/commands/your-command.md
---
description: One-line summary of what this does
---
Plain-English instructions for Claude:
· what file(s) to read
· what steps to follow
· what to produce
· quality rules, fallbacks, edge cases
You can write this by hand, or ask Claude to draft it after you've run a task once manually: "turn what we just did into a reusable command." Save as /your-command, anyone on the team can run it.
Level 4 (scheduled agents for data freshness, consistency audits, deliverable QA) is real and running in other corners of this work. We'll explore it for the Lab later this year. Appendix has details.
22 / 27Act 3 · How the team adopts this
The compounding effect
Every discovery makes every future session smarter. Every command makes every future task faster. Every glossary entry saves a question.
Month 1
2
3
4
5
6
7
8
Chat is linear: each session is independent. Collective intelligence compounds: each session builds on everything the team has contributed before it. The gap widens every month.
The team that treats Claude as a chatbot gets a calculator. The team that builds collective intelligence gets an operating system.
23 / 27
What Claude doesn't replace
Drawing a clear line around what stays human.
Lab judgment
Which plays to prioritize, which countries to engage, which partners to work with. Contested decisions belong to humans with context Claude doesn't have.
Peer review
Methodological rigor, contested interpretations, credentialing. Claude can spot gaps; it cannot confer validity.
Fieldwork
Interviews, site visits, relationships with ministries and firms. The signal isn't in the transcript — it's in the room.
Client trust
The Lab's reputation rests on humans being in the room. Claude reduces friction; it doesn't shortcut presence.
The kit is a force multiplier, not a substitute. A Lab without judgment, peer review, fieldwork, and trust doesn't become a Lab by adding Claude.
24 / 27
Honest limits
Three questions worth answering before you ask them.
What breaks?
The kit drifts from reality if nobody updates it. The first stale entry erodes trust. Mitigation: quarterly review, named maintainer, team contributes.
What's the risk?
Wrong discoveries propagate. Mitigation: two-lane system. discoveries.md is free-for-all (fast). Promotion to gotchas.md or glossary.md requires PR review.
What's the cost?
Claude Team: ~$30/seat/month. Contribution: 30 seconds per discovery, maybe once a week. That's the full operating cost of collective intelligence.
If you have a fourth worry not covered here, Q&A is coming.
25 / 27
Implementation timeline
From "yes" to "everyone is contributing" in one quarter.
Week 1 · Admin
Team plan signed. Seats provisioned. Kit repo shared at Lab level. Admin owner: one named person.
Week 2 · Onboarding
30-min hands-on session per person. Open Claude with the kit loaded. Try one real prompt from an actual task they're working on.
Week 3 · First discoveries
First Lab-wide discoveries.md entry contributed by someone other than me.
Month 1 · First pipeline
A team member (other than me) runs an existing command like /enrich-fdi and ships the output into a live deliverable.
Month 3 · First contributed command
A team member drafts a new command from a recurring task. PR, review, merge. The kit expands under new authorship.
Quarterly · Review
Measure discoveries contributed, commands created, pipeline runs by team members. Decide what to systematize next. Promote signal entries to gotchas.md.
None of this depends on everyone becoming technical. It depends on one norm: when you learn something, add it.
26 / 27
The ask
Four things to make this work at Lab scale.
☐ Claude Team plan
A shared subscription for the Lab. Claude.ai for everyone (chat, Projects, web) plus Claude Code for pipeline and deliverable work. One invoice, one admin.
☐ Kit as Lab resource
Repo at github.com/Gilix/nzipl-claude-kit. Already public, already readable. Once the Team plan lands, everyone can use it, contribute to it, and inherit what others have added.
☐ One norm
When you learn something worth sharing, add a line to discoveries.md. 30 seconds. Permanent return across the whole team.
☐ Quarterly review
Measure discoveries, commands, pipeline runs. Review the roadmap. Decide what to build or systematize next. First review ~90 days after approval.
The gap between "I learned something" and "the whole Lab knows it" should be one line and one push. That's what a yes gets us.
27 / 27Interactive · Chat · 3 min
One question before Q&A
Help me understand what would be most useful to you.
Post in Zoom chat
What question would you most want answered by a Lab-informed Claude? One sentence.
Your answers become the first test cases for the kit — the list of questions a Lab-informed Claude should be able to answer once the team has access.
Questions →
Appendix
Reference material for the curious. Not presented in the session.
A / BAppendix A
Toward the Atlas
CICE evolves from cross-country snapshots to subnational depth.
The Clean Industrial Capabilities Explorer shows which countries have competitive advantages in clean energy technologies. The Atlas extends this in four dimensions:
Process chain products (capital goods, testing equipment) alongside supply chain products. Connects to WP4's capability cluster framework.
Temporal dimension
Competitiveness trajectories over time, not just snapshots. Which regions are building capability? Which are losing it?
Policy interface
Constraint maps, sequencing timelines, finance architecture alongside the data. Analysis that leads to action.
The platform produces the analytical layer. CICE displays it. The platform feeds the Atlas.
B / BAppendix B
Level 4: scheduled agents
Aspirational tier. Scheduled tasks that run without prompting.
Data freshness
Check whether OEC BACI data has been updated. Flag stale caches. Refresh when upstream changes.
Consistency audit
Verify all play cards use the same HS6 taxonomy version. Flag drift between country pipelines.
Deliverable QA
Scan HTML files for broken chart containers, missing data sections, design token violations. Surface issues before they reach a presentation.
Discoveries triage
Cluster new discoveries by topic. Propose which should be promoted to glossary or gotchas at the next quarterly review.
Not speculative in general — scheduled Claude agents already run in production elsewhere in this work. Weekly status reports, risk scans, compliance checks. The pattern transfers directly to the Lab's analytical pipeline. Post-approval, Q3 2026.