Starting From a Published Design¶
The Potato Showcase, and what to change before reusing a design. It ships with the potato-tasks skill:
/plugin marketplace add davidjurgens/potato-skill
/plugin install potato-tasks@potato
Three sources of prior art, and they cover different things. Reaching for the wrong one wastes an hour.
| Source | What it gives | What it does not |
|---|---|---|
examples/ in Potato, 214 of them |
One scheme or display type demonstrated cleanly | Question wording worth copying; a whole study |
| potato-showcase, 440 designs | Real label sets, real question wording, written instructions, the paper each came from | Any workflow at all |
worked-example.md |
The whole study: consent, instructions, training, checks, survey | Anything about your construct |
The usual route on a new task is showcase for the questions, worked example for
the workflow, examples/ when one scheme is behaving oddly.
The showcase¶
https://github.com/davidjurgens/potato-showcase — 440 annotation task designs,
most built from a published paper or dataset. Each is a folder with
config.yaml, metadata.json (title, paper, dataset, tags, complexity) and
sample-data.json.
Two things make it worth checking before writing a scheme from a field list:
- The wording is not invented. A label set a paper used and reported agreement on has survived a pilot. Yours has not.
- 281 of the 440 carry written
annotation_instructions— the part of a study this pack otherwise leaves you to write from nothing. Seewriting-guidelines.mdfor what to do with them.
And the limit, which is sharp: no design in the showcase has a phases block,
a consent page, a training round or a post-study survey. Zero. They are task
designs, not studies. Take the questions from there and the study around them
from worked-example.md.
Searching it¶
python .claude/skills/potato-tasks/scripts/find_design.py --type span --category text
python .claude/skills/potato-tasks/scripts/find_design.py --query "dialogue safety" --with-instructions
python .claude/skills/potato-tasks/scripts/find_design.py --show text/education/mathdial-tutoring-dialogue
Filters: --type (repeatable, all must be present), --display-type,
--category, --complexity, --with-instructions, --with-paper, --query.
--json for the raw records. --show <id> prints the metadata and the whole
config.
The script reads showcase.manifest.json, which is built by
scripts/generate_manifest.py in the showcase repo. It looks for a clone in the
obvious places and in $POTATO_SHOWCASE, and falls back to downloading the
manifest from GitHub. Pass --showcase <dir|path|url> to be explicit.
If you have a clone but no manifest, build it once:
cd path/to/potato-showcase && python3 scripts/generate_manifest.py
--show prints the config only from a local clone; without one it prints the
GitHub URL.
Taking a design without inheriting its mistakes¶
Copy the folder, then go through this:
- Re-read every label against your data. The taxonomy was built for their corpus. A category with nothing in yours is worse than no category — it attracts arbitrary answers.
- Check for a residual option. Published schemes often assume trained annotators and a discussion round, so some have no "unclear".
- Instructions are a starting draft, not the researcher's guidelines. They
are what the showcase design's author wrote, sometimes condensed from the paper. Name
the source in
DESIGN.md, and put the real definitions past the researcher. - Cite it.
metadata.jsoncarriespaperReference,paperUrland a BibTeXcitation. If the scheme came from a paper, the study using it says so. - The config is the truth, not the metadata.
metadata.annotationTypesis hand-maintained and six designs disagree with their own config. The manifest records both;annotation_types_in_configis the one that runs. - Add the workflow. No consent, no instructions phase, no training, no checks. That is all yours.