paradigma-skill

Agent skill that maps how a field's constraints changed over the years,to find where a new product, startup or line of research has just become possible.

English | Español | 简体中文

Quick start · How it works · Install · Usage · Limits · FAQ

---

Quick start

npx skills add qaemu/paradigma-skill

Or add it to Claude Code by hand, from the plugin marketplace:

/plugin marketplace add qaemu/paradigma-skill
/plugin install paradigma@paradigma

[!NOTE] PARADIGMA gives you a map, not a prediction. It shows where constraints moved and who has noticed. Deciding when to act, and on what, is still your job.

Why constraints

Big shifts are rarely ideas nobody had before. Digital cash was attempted for twenty years before Bitcoin. Neural networks existed for decades before cheap GPUs and large datasets made them work. What usually changes is underneath: something gets 10× cheaper, a patent expires, a law takes effect, a skill stops being scarce, a norm shifts. An assumption the whole industry treats as fixed stops being true.

Those changes are usually public, so tracking them alone gives you no edge. The edge comes from comparing them against who is actually building on them. PARADIGMA does that comparison, for any field you name.

It is	It is not
A sourced panorama of what changed in a field	An idea generator
Robust trends computed in code, with confidence grades	Numbers estimated in prose
A map of who is building on each change, and how far they got	A list of hyped startups
Explicit about gaps in its own data	A substitute for your judgment

How it works

Step	What happens
1. Map the field	Breaks the field into 12–25 constraints: components, manufacturing, software and AI, data, labor, distribution, compliance, trust. Then a PESTEL sweep catches the political, economic, social, technological, environmental and legal ones.
2. Price curves	Helper agents collect historical data points in parallel, each with a link. The engine takes the median for each year, removes inflation and fits a robust trend that one bad value can't move.
3. Non-cost sweep	Official gazettes, standards bodies, patents expiring now, wage and shortage statistics, adoption surveys, export controls. Each change is dated and marked loosening or tightening.
4. Activity scan	Counts companies, funding, patents, papers and launches per constraint, with the same queries for every constraint. Each company is marked announced, deployed or paying, with evidence.
5. Gap matrix	Plots how far each constraint fell against how many are building on it.
6. Why-not check	Every gap is checked for failed attempts and remaining blockers before it's called an opportunity.
7. Independent check	A fresh agent that did none of the research re-opens the sources and checks every value. The report states the pass rate.
8. Report	A one-page panorama, one card per constraint, a timeline of dated changes, a company map, and a workbook with every data point and its link.

Ground rules

1. Every number has a link. If a number can't be found, the report says it's missing. It never guesses.
2. The math is done in code (paradigma.py), never estimated by the model.
3. Money raised is not proof of success. Each company is marked announced, deployed or paying, with the evidence.
4. Nothing is called an open gap until the why-not check has run.
5. A second agent, one that didn't do the research, checks the work.

What you get

Quadrant	Meaning	What the report does with it
Open gap	The constraint dropped a lot (cheaper, allowed, easier), but few companies use the change yet	Runs the why-not check, then sorts each item into untried, tried too early (with the failed companies and dates) or blocked
Crowded	The constraint dropped a lot, and many companies already use it	Lists who they are and how far each got. This is what most people in the field already know
Hype risk	The constraint barely moved, yet many companies are building on it anyway	Marks it as a warning: lots of activity with no real change underneath, so treat growth claims with caution
Dormant	Little change and little activity	Kept for reference

Deliverables

• Report. The panorama first: biggest drops, tightening constraints, open gaps, crowded zones, hype risks, and dated changes in the next 1–5 years such as laws taking effect and patents expiring. Then one card per constraint with its curve, yearly change and margin of error, fold change, trend shape, grade, thresholds, companies and why-not label. It ends with method, confidence and sources.
• Workbook (.xlsx). Opens on Start here, which explains every term in plain words, then Key findings. After that come the gap matrix, why-not check, constraints, price curves, yearly medians, data points, laws and timeline, companies, funding rounds, activity, failures, searches with no result, and verification. Every row has its source link.
• Charts. One log-scale curve per constraint, plus the gap matrix.

Example

A full run on predictive maintenance for industrial machinery (2006–2025) is in examples/predictive-maintenance. It has the report, the gap matrix, an Excel workbook with a Start here sheet that explains every term, and every data point with its source. The run is honest about what it couldn't find: the why-not check withdrew both open gaps the engine flagged on the first pass.

Install

PARADIGMA follows the open Agent Skills format: a SKILL.md plus a Python engine. Any agent that can browse the web, run Python and, ideally, start helper agents can use it.

Any agent (skills CLI)

npx skills add qaemu/paradigma-skill                    # current project
npx skills add qaemu/paradigma-skill -g                 # all projects
npx skills add qaemu/paradigma-skill -a claude-code -a opencode

Uses the skills CLI, which supports Claude Code, OpenCode, Codex, Cursor and many more.

Claude Code

As a plugin, which gives you updates through /plugin:

/plugin marketplace add qaemu/paradigma-skill
/plugin install paradigma@paradigma

Plugin skills are namespaced, so you run it as /paradigma:paradigma <field>.

Or as a personal skill, which you run as /paradigma <field>:

git clone https://github.com/qaemu/paradigma-skill.git
mkdir -p ~/.claude/skills && cp -r paradigma-skill/skills/paradigma ~/.claude/skills/

Claude apps (web, desktop, mobile)

1. Download dist/paradigma.zip. It contains the paradigma folder with SKILL.md and the engine.
2. Turn on Code execution and file creation in Settings → Capabilities.
3. Go to Customize → Skills, click +, then Create skill → Upload a skill, and choose the ZIP.

See Anthropic's guide: Use skills in Claude.

OpenCode

git clone https://github.com/qaemu/paradigma-skill.git
mkdir -p ~/.config/opencode/skills && cp -r paradigma-skill/skills/paradigma ~/.config/opencode/skills/

OpenCode also reads ~/.claude/skills/, so a Claude Code install works there too. Check that web search works with your provider first. Without it, PARADIGMA can't collect data.

Requirements

Needs	Why
Web search and page fetching	Every number comes from a source
Python 3.10+ with pandas, numpy, matplotlib, openpyxl (scipy optional)	The analysis engine, its charts and the Excel workbook
Helper agents (optional)	Parallel research and the independent check. Without them it runs one step at a time and marks the check as self-checked

Usage

/paradigma <field> [region] [years] [standard|deep]

Argument	Default	Notes
field	required	If it's very broad, like hardware or health, PARADIGMA suggests 3–4 narrower subfields first. Narrower fields give much better results
region	global	Notes where the data is US- or EU-centric
years	10	How many years back to look, ending at the last full year. Write 10, 10y or years=10 (5 to 50), or a range like 2010-2025. The default covers recent shifts with enough points for good grades. Use 5 for a quick look at fast fields, 20 for hardware whose costs fall slowly, 30+ for slow infrastructure. Short windows give fewer data points per curve, so grades are lower, and anything that fell earlier looks unchanged
depth	standard	standard covers 12–25 constraints. deep covers up to 40, with denser curves and why-not checks on every candidate

/paradigma agricultural machinery
/paradigma small-batch manufacturing Latin America
/paradigma lab instruments 2010-2025 deep
/paradigma dental aligners EU 10

[!IMPORTANT] A standard run takes about an hour, with several agents working in parallel, and uses a substantial part of a plan's usage. deep takes longer. Pick the field before you start.

Which model to run it on

What matters for this task is long-horizon reliability and discipline about sources. A fast model that fills a table with plausible numbers is the failure mode to avoid.

Role	Recommended	Effort
Main run (planning, gap matrix, why-not, report)	Claude Opus 5.5	high
Data-collection helpers	Claude Sonnet 5.5	medium–high
Independent check	Opus 5.5, or Claude Fable 5.1 for deep runs	high

Avoid small and "flash" models for the main run. To check any model, including free ones, run it on 3 constraints with well-published data, such as lithium-ion pack prices and solar module prices, then open every link it returns. If it invents even one link or number, it fails.

The math

1. Normalize. Values are expressed per unit of performance ($/kWh, $ per sensor channel, $/kg payload). USD values are converted to constant dollars with US CPI from FRED. Non-USD values need a cited exchange rate, or they're rejected. Valid points dated outside the window are not analysed but are kept in out_of_window.csv, so a longer-horizon forecast can reuse them.
2. Bin. Data is grouped by year, or by quarter when it's dense enough: at least 3 points in at least 60% of quarters, and at least 8 quarters. Each period keeps its median, its interquartile range and its count, and periods with fewer than 3 points are flagged.
3. Fit. A Theil–Sen line is fitted on log₁₀ of the period medians, so each period gets one vote and outliers barely move it. The slope gives the yearly % change with a 95% confidence interval. The engine never reports an endpoint-to-endpoint percentage as a trend.
4. Shape. The last 5 years are compared with the whole window to classify the trend as accelerating, steady or decelerating.
5. Classify. Each curve is improved ≥10×, improving, flat / unclear, or worsening (constraint tightening). The direction depends on whether lower or higher is better for that metric.
6. Grade.

Grade	Requires
A	≥ 8 periods, ≥ 3 independent sources, ≤ 20% thin periods, tight confidence interval
B	≥ 5 periods, ≥ 2 sources
C	usable but weak. Can't support a headline claim
D	fewer than 3 periods. No trend computed

• Drop score. The percentile rank of each constraint's improvement per decade (log₁₀ of the fold change). Non-cost constraints are scored 0–3 from the evidence, with a written justification. Crossing a viability threshold adds a bonus.
• Activity score. The mean percentile rank of the log-scaled companies, funding, patents, launches and job postings that are available. Missing signals are skipped, not counted as zero.
• Research ahead of companies. Papers in the last 3 years divided by (companies + 1). When research is far ahead of companies, the gap score goes up.
• Gap score = drop score × (1 − activity score), weighted by research-ahead. Quadrants split at the median.

These scores are relative. They rank the constraints of one field against each other, not against an absolute scale.

cd skills/paradigma/scripts
python3 paradigma.py cpi                  # US CPI annual averages from FRED
python3 paradigma.py curves --start 2016 --end 2025 --plots   # points.csv -> curves.csv, periods.csv, plots/
python3 paradigma.py papers --start 2020 --end 2025   # OpenAlex paper counts
python3 paradigma.py gap --plots          # curves + activity -> gap_matrix.csv, plots/gap_matrix.png
python3 paradigma.py workbook --field "…" --start 2016 --end 2025   # every table -> readable .xlsx

The CSV schemas are in SKILL.md, Appendix B.

Data sources

Every source PARADIGMA uses is free. The report says where paid databases would add more.

Need	Used by PARADIGMA (free)	What professionals add (paid)
Cost curves	Our World in Data, Santa Fe Performance Curve Database, BLS / FRED, USGS, NREL, IEA, price lists + Wayback Machine	BloombergNEF, teardown firms
Companies and funding	Press releases, free Crunchbase / Dealroom pages, startup directories	PitchBook, CB Insights, Crunchbase Pro
Patents	Google Patents, Lens.org	PatSnap, Derwent
Research papers	OpenAlex	Scopus, Web of Science
Regulation	Official gazettes (Federal Register, EUR-Lex, national gazettes), ISO / IEC / SAE	Legal trackers

Limits

• Funding data is the weakest signal. Free sources miss undisclosed rounds and many non-US companies.
• Statistics lean US and EU. The report flags regional bias.
• Being too early looks the same as being wrong. Drops less than 3 years old are flagged as early.
• Activity driven by speculation is not adoption. The report separates hype from users with a need.
• Status: v0.1, experimental. The engine passes an offline test suite on synthetic data with known answers. One full real run is published in examples/predictive-maintenance; it also exposed and fixed an engine bug.

FAQ

Does it touch my files or accounts? No. It reads public websites and writes only to its own working folder. It doesn't sign in anywhere.

Why not ask a model for a market report directly? Because it will give you one, numbers included, and you won't be able to tell which numbers it invented. PARADIGMA requires a source for every number, does the math in code, and has a separate agent check the work.

Can I use it on a free model? Yes, if the model passes the test in Which model to run it on. Check each provider's data policy first, since some free models train on your prompts.

How is this different from a consultant's market report? A market report sizes a market. PARADIGMA asks which assumptions behind a market stopped being true, and who has noticed.

Repository

paradigma-skill/
├── .claude-plugin/          Claude Code marketplace + plugin manifests
├── skills/paradigma/
│   ├── SKILL.md             instructions the agent follows
│   └── scripts/paradigma.py analysis engine (curves, gap matrix, CPI, papers)
├── tests/test_engine.py     offline test with known answers
├── examples/               a full real run (predictive maintenance)
├── dist/paradigma.zip       ready-to-upload skill for Claude apps (tools/package.sh)
├── tools/                   make_assets.py (README artwork), package.sh (rebuilds the ZIP)
├── README/                  translations of this page
└── assets/                  banner, pipeline and gap-matrix art

python3 tests/test_engine.py   # expect 15 PASS lines

Contributing

Issues and pull requests are welcome, especially:

• real runs you're willing to share in examples/
• better free data sources for a specific field or region
• cases where a gap PARADIGMA reported turned out to be wrong, and why

Keep the rules intact: every number gets a source, and all math stays in the engine.

License

MIT © 2026 qaemu