Causal Powers
Superpowers for data analytics, causal inference, and econometrics.
A Claude Code skill family that adapts the discipline of
superpowers, whose name it borrows in
homage, to the failure modes specific to data work. In software the dangerous bug
is loud: a stack trace points near its cause. In empirical work it is silent —
the code runs clean and returns a confident, wrong answer. These skills make that
failure visible before it reaches a result.
A number you computed but never validated is a guess wearing a lab coat.
What makes it different
- Economic framing of mature, proven skills. It adapts the disciplines proven in software engineering — contract-checked transforms, systematic debugging, specification before code, independent review — to the silent failures and judgment calls of empirical microeconomics.
- It grows into your data's domain. Each iteration records what went wrong in a given project and surfaces it the next time the same step recurs, so the discipline grows more attuned to that dataset over time. The shared skills stay general; what accumulates is the project's own record of past mistakes.
- Built for day-to-day research, not one-shot answers. Empirical projects run for weeks across many sessions. A living, phased
analysis-plan.md keeps the state of the work on disk, so it survives /clear, automatic compaction, and interruption: a new session resumes where the last left off, with prior decisions and their rationale intact.
Motivation
The most carefully designed agent skills today are built for software engineering, yet their organizing ideas are not specific to software:
- Goal-driven execution — state the success criterion, then loop until it is met; evidence before assertion (superpowers; Karpathy's notes).
- Human-in-the-loop gates — surface a consequential decision for the user rather than settle it silently (superpowers).
- Evolving — skills that sharpen each time they run, recording what failed and folding it back in (ECC).
- Planning — a written plan that persists across sessions and resumes cleanly, treating the disk as working memory (planning-with-files).
Each transfers naturally to empirical microeconomics, where the consequential failures are silent and the work divides into two pathways with distinct purposes: reduced-form analysis, which measures an effect present in the data, and structural estimation, which recovers the primitives needed to simulate a counterfactual the data does not contain.
Causal Powers therefore introduces no new methodology. It reorganizes these well-developed practices and refocuses them on microeconomic analysis, then adds the discipline the domain demands: identification before estimation, recovery before trust, and economic, not merely statistical, judgment.
What you get
- Catches the bugs that don't throw. A join fans out and revenue triples; one
NA poisons a mean; train/test overlap fakes a model metric; a control is post-treatment — every one a clean run. These skills make them loud before they reach a stakeholder.
- A senior economist's instincts, not an RA's checklist. Forms a prior on sign, magnitude, and mechanism before the data; reads every estimate in interpretable units and judges economic, not just statistical, significance; refuses a causal claim without a named design ("what's your experiment?").
- Never changes the goal behind your back. Dropping data, swapping a spec, "upgrading" the design mid-debug — each one stops and asks. You stay in control of the estimand, the sample, and the design; the agent loops autonomously toward the agreed goal, never past it.
- Plans at two altitudes. It pins the estimand before code (study altitude) and the small-step roadmap before a merge or a debug (task altitude) — the roadmap you approve first, not a dive.
- Measured, not asserted. Ships a trigger CI and a planted-silent-failure benchmark. Most skill libraries can't tell you whether they actually help; this one is instrumented (
scripts/eval-triggers.py, scripts/run-behavioral-eval.py).
The flow
