Mode d’emploi
- Copiez la commande et examinez le paquet avant de l’installer dans votre environnement agent.
- Consultez le document original pour les outils, identifiants et environnements requis.
- Décrivez les entrées et le livrable souhaité ; comparez la sortie aux sources.
npx skills add https://github.com/probabl-ai/skore-skills --skill research-ml-practiceAvant de commencer
- L’installation n’inclut ni abonnement au modèle, ni quota API, ni accès aux fichiers.
- Étoiles et installations reflètent l’activité, pas un test indépendant de qualité ou de sécurité.
Rendu depuis le dépôt source en conservant titres, exemples, code, tableaux, liens et images.
Research ML Practice
Worker skill. Callers own the stage turn and the user-facing summary. Do not git end-turn.
Human-facing prose
Details: setup-workspace references/human_facing_prose.md. Chat stays path + one or two sentences of the finding. Do not narrate the skills framework or the wrapper CLI.
Sequence
- Intake. Infer modality from dtypes / JOURNAL when
obvious; use a stated domain and stage (data_analysis | model) if the caller named them. Then pick a mode (references/search.md):
- Caller passed a named concern (not the canned EDA
extra-analysis question) → depth. Do not re-ask. Do not run the canned survey first.
- Caller passed the canned extra-analysis survey or
“survey extra analyses” → survey.
- Neither → if JOURNAL and an EDA report exist, run the
canned survey; else AskUserQuestion for a named concern, stating the modality and stage inferred this turn and that the answer only drives a literature search, no code. A file link is an addition, never the context. Do not start a depth search with an empty concern.
- Abstract the problem class before any query
(references/search.md). JOURNAL and EDA are context, not the answer list and not search keywords for the table’s proper name.
- Run the matching search loop (
references/search.md).
Fetch primary pages. Follow up per promising extra if the first pass is thin or single-sourced.
- Write
scratch/research/using the matching structure in
references/search.md (survey: survey-<slug>.md; depth: <slug>.md with lanes). There is no templates/ directory in this skill — copy the markdown skeleton from that reference. Gitignored.
- Return to the caller: scratch path and a one- or two-sentence
finding. Name that candidates are laned (measure / declare / evaluate / confirm). Chat is path + those sentences only — no pasted headings, tables, or “Depth note — …” body. That is the complete user-facing deliverable even when a harness says to put the full answer in chat. If tools cannot search or write, stop there: still only path + sentences (name the intended scratch/research/ path). No hypotheses, diagnostics, planned-query bullets, or template headings in chat (that is the paste). Do not write data_analysis.md, the design note, or data/. The caller asks which extras to add.
Stop conditions
- Do not drop, impute, or remove outliers. Do not change the split.
- Do not pick the final learner or architecture.
- Do not treat a single blog as ground truth; say when sources
disagree. Thresholds need two independent sources.
- Do not
pixi add/uv add/env add. If code needs a
library, name add-python-package and return.
- Do not run
api getas a substitute for literature (symbols
still go through api get in the caller).
- Missing skill from a caller → that caller one-line skips.
- Do not copy Open questions / EDA findings onto the extras
list without a source.
- Do not search the dataset proper name,
sklearn.datasets,
a Kaggle slug, or “baseline pipeline”. Do not return learners / Pipeline steps as EDA extras.
- Never answer from memory when search ran. Do not ask the
user to go look something up.
- Do not paste the scratch markdown into chat (no “Depth
note —”, no survey body). Path + 1–2 sentences only. If tools cannot search or write, stop after that. A no-tools harness does not license pasting the note.
