Repro Capsule Builder
Capture your compute environment so you can reproduce results later
Researchers struggle to reproduce even their own results when hardware and software drift between runs. This recipe creates a "good enough" reproducibility capsule: dependencies, versions, run commands, random seeds, and a minimal runbook someone else can follow.
INGREDIENTS
PROMPT
Create a skill called "Repro Capsule Builder". Ask for: - Language/toolchain (Python/R/MATLAB/etc.; if unknown: unspecified) - How code is run (script/notebook/pipeline; if unknown: unspecified) - Where data lives and whether it can be shared (sensitive? yes/no/unspecified) Output: - A reproducibility capsule checklist: - dependencies + versions capture approach - run commands - seeds and configs - logging and output paths - A minimal runbook README template. Rules: - Never assume sensitive data can be uploaded; default to local-first suggestions.
How It Works
This recipe focuses on practical reproducibility — not perfection.
Record what matters, make reruns predictable, and produce a runbook
that a colleague (or future-you) can follow without a phone call.
What You Get
- An environment manifest (conda/pip/renv or equivalent)
- A run command template with seed and config management
- A logging checklist (what to capture, where outputs go)
- A minimal README runbook:
- Inputs
- Steps
- Expected outputs
- Where outputs are stored
Setup Steps
- Tell the Claw your language/toolchain (Python/R/MATLAB/etc.)
- Describe how code is run (script/notebook/pipeline)
- Clarify where data lives and whether it can be shared
- The Claw generates the capsule checklist and runbook template
Tips
- Essential before publishing code or submitting to journals requiring code availability
- Critical for lab handoffs (student graduating, postdoc leaving)
- Sensitive data? The Claw defaults to local-first suggestions — nothing gets uploaded