[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"skill:01M6NPFP1AQJ9NW21P0VRE9HSH":3,"skill-same:01M6NPFP1AQJ9NW21P0VRE9HSH":101},{"skillNo":4,"name":5,"nameZh":6,"description":7,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":16,"categories":17,"keywords":31,"publishedAt":48,"linkOnly":49,"markdownContent":50,"markdownBody":51,"frontMatter":52,"refName":61,"commitSha":62,"skillDirectory":63,"forksCount":14,"repositoryHtmlUrl":64,"repositoryHomepageUrl":65,"ownerHtmlUrl":66,"ownerBlogUrl":6,"ownerDescription":6,"updatedAt":67,"install":68,"resources":81},"01M6NPFP1AQJ9NW21P0VRE9HSH","data-science",null,"Data analysis workflow from import through modeling and communication. Use when analyzing a dataset, exploring data, building a statistical model, selecting features, or communicating findings to stakeholders.","projectious-work\u002Fprocesskit","projectious-work","https:\u002F\u002Favatars.githubusercontent.com\u002Fu\u002F263889726?v=4","01M6N97TJJ118WQ35CHRV8RD0E","projectious.work","01M6N97TPWJPF0BMKYQ4PBY5W8",0,"MIT","https:\u002F\u002Fgithub.com\u002Fprojectious-work\u002Fprocesskit\u002Fblob\u002Fmain\u002Fsrc\u002Fcontext\u002Fskills\u002Fdata-ai\u002Fdata-science\u002FSKILL.md",[18,23,26],{"categoryNo":19,"categoryName":20,"taxonomyType":21,"taxonomyTypeDesc":22},"01JBSEEDCAT000000000000013","分析与可视化","SCENARIO","使用场景",{"categoryNo":24,"categoryName":25,"taxonomyType":21,"taxonomyTypeDesc":22},"01JBSEEDCAT000000000000023","Skill与智能体开发",{"categoryNo":27,"categoryName":28,"taxonomyType":29,"taxonomyTypeDesc":30},"01JBSEEDCAT000000000000033","数据分析师","PROFESSION","适用职业",[32,33,34,35,36,5,37,38,39,40,41,42,43,44,45,46,47],"data science","data","science","model","features","import","dataset","modeling","communicating","use","work","data analysis","analysis","analyzing","communication","workflow","2026-11-10 10:12:30",false,"---\nname: data-science\ndescription: |\n  Data analysis workflow from import through modeling and communication. Use when analyzing a dataset, exploring data, building a statistical model, selecting features, or communicating findings to stakeholders.\nmetadata:\n  processkit:\n    apiVersion: processkit.projectious.work\u002Fv2\n    id: SKILL-data-science\n    version: \"1.1.0\"\n    created: 2026-04-06T00:00:00Z\n    category: data-ai\n---\n\n# Data Science\n\n## Intro\n\nA solid data analysis flows from import through cleaning, tidying,\nexploration, modeling, and communication — in that order. Tidy data,\nhonest uncertainty, and plain-language findings beat clever models on\nmessy inputs.\n\n## Overview\n\n### Import and cleaning\n\nInspect shape, dtypes, nulls, and duplicates immediately after\nloading. Print `df.info()` and `df.describe()` before any analysis.\nHandle missing data explicitly — drop, fill, or flag, but never ignore\nsilently. Parse dates at load time (`parse_dates=`) and set correct\ndtypes early. Validate assumptions: no negatives where impossible, no\nstring\u002Fnumeric mismatches.\n\n### Tidy data\n\nReshape to tidy form: one variable per column, one observation per\nrow. Use `melt()` for wide-to-long, `pivot_table()` for long-to-wide.\nSplit mixed columns (e.g. `city_state` into two). Normalize repeated\nentities into separate tables. See Level 3 for messy-pattern\nrecipes.\n\n### Exploratory data analysis\n\nStart univariate (histograms, value counts, box plots), then\nbivariate (scatter, grouped bars, correlation). Look for outliers,\nskewness, class imbalance, and unexpected patterns. Use `.groupby()`\nto understand subgroups. Summarize findings in plain language before\nmoving to modeling.\n\n### Statistical reasoning\n\nState the question before choosing a test. Check assumptions\n(normality, equal variance, independence) before applying parametric\ntests. Report effect sizes alongside p-values — statistical\nsignificance is not practical significance. Use confidence intervals\nto communicate uncertainty.\n\n### Feature selection\n\nRemove zero- and near-zero-variance features. Drop one of each highly\ncorrelated pair (r > 0.9). Use domain knowledge first, then\ndata-driven methods (mutual information, permutation importance).\nWatch for leakage — never use target-derived features in training.\n\n### Model selection workflow\n\nStart simple: a mean\u002Fmedian baseline, then linear or logistic\nregression. Add complexity only if simpler models underperform and\nyou understand why. Use train\u002Fvalidation\u002Ftest split or\ncross-validation; never evaluate on training data. Compare on a\nsingle metric chosen before fitting. Document why a model was chosen,\nnot just which one won.\n\n### Communicating results\n\nLead with the finding, not the method. Use plain language: \"customers\nwho used feature X retained 23% more\" beats \"the coefficient was\n0.23.\" Show uncertainty — ranges, intervals, caveats. Include a \"so\nwhat\" — what action does this analysis support? Pin library\nversions, set random seeds, and script everything for\nreproducibility.\n\n### Preferred libraries\n\n| Task | First choice | Alternative |\n|------|-------------|-------------|\n| DataFrames | polars | pandas |\n| Visualization | matplotlib + seaborn | plotly |\n| Statistics | scipy.stats | statsmodels |\n| ML | scikit-learn | xgboost\u002Flightgbm |\n| Notebooks | jupyter lab | marimo |\n\n## Gotchas\n\nAgent-specific failure modes — provider-neutral pause-and-self-check items:\n\n- **Starting with a complex model before a simple baseline.** Training an XGBoost model before computing a mean predictor or running a logistic regression means you can't know how much of the model's value came from complexity vs. data. Build a baseline first; complexity is only justified when the baseline is inadequate.\n- **Evaluating on training data.** \"The model is 98% accurate\" on the data it trained on is not a finding — it is a description of memorization. Always evaluate on held-out data and report the held-out metric.\n- **Confusing statistical significance with practical significance.** With n > 10,000, almost any real difference is statistically significant. A p-value of 0.001 on a 0.1% conversion difference means \"this is real\" but does not mean \"this matters.\" Always report effect size alongside p-values.\n- **Choosing a statistical test before checking its assumptions.** Running a t-test on non-normally distributed data with small samples produces unreliable p-values. Check normality (Shapiro-Wilk for n \u003C 5000), equal variance (Levene), and independence before choosing a parametric test; fall back to non-parametric alternatives when assumptions fail.\n- **Reporting findings without uncertainty.** \"Feature X predicts churn\" without a confidence interval or cross-validation variance means the finding might not replicate. Always communicate uncertainty ranges alongside point estimates.\n- **Truncating the bar chart y-axis to make differences look larger.** A bar chart with a y-axis starting at 0.85 makes a 0.03 difference look dramatic. Unless the baseline is well-understood (like conversion rate from 30% to 32%), bar charts should start at zero.\n- **Not setting random seeds before any analysis.** Non-reproducible analysis is not science. Set `random_state=42` (or equivalent) on all sampling, model initialization, and cross-validation calls. Document the seed in the notebook or script so results can be reproduced exactly.\n\n## Full reference\n\n### Tidy data — messy patterns and fixes\n\n**Headers as values (years across columns).** Melt to long form:\n\n```python\n# pandas\ndf_tidy = df.melt(id_vars=\"country\", var_name=\"year\", value_name=\"population\")\n\n# polars\ndf_tidy = df.unpivot(index=\"country\", variable_name=\"year\", value_name=\"population\")\n```\n\n**Multiple variables in one column** (`male_18-25`). Split:\n\n```python\n# pandas\ndf[[\"gender\", \"age_group\"]] = df[\"demographic\"].str.split(\"_\", expand=True)\n```\n\n**Variables in both rows and columns** (metric in rows, dates in\ncolumns). Melt dates first, then pivot metrics:\n\n```python\nlong = df.melt(id_vars=[\"id\", \"metric\"], var_name=\"month\", value_name=\"value\")\ntidy = long.pivot_table(index=[\"id\", \"month\"], columns=\"metric\", values=\"value\").reset_index()\n```\n\n**Multiple observational units in one table.** Normalize:\n\n```python\nsongs = df[[\"song_id\", \"title\", \"artist\"]].drop_duplicates()\nplays = df[[\"song_id\", \"play_date\", \"user_id\"]]\n```\n\nWide form is fine when feeding correlation matrices, display pivot\ntables, or ML feature matrices. Store and clean tidy; reshape for\nthe consumer.\n\n| Operation | pandas | polars |\n|-----------|--------|--------|\n| Wide to long | `df.melt()` | `df.unpivot()` |\n| Long to wide | `df.pivot_table()` | `df.pivot()` |\n| Split column | `str.split(expand=True)` | `str.split().list.get()` |\n| Deduplicate | `df.drop_duplicates()` | `df.unique()` |\n\n### Statistical methods\n\nAlways compute descriptive stats first (`df.describe()`,\n`value_counts(normalize=True)`, `df.skew()`). Mean is sensitive to\noutliers; median is robust. Standard deviation reports spread in\noriginal units; IQR is the robust analogue.\n\n| Distribution | Use when | Example |\n|---|---|---|\n| Normal | Continuous, symmetric | Heights, measurement error |\n| Binomial | Fixed trials, two outcomes | Conversion rate |\n| Poisson | Counting events in fixed window | Tickets per hour |\n| Exponential | Time between events | Customer arrivals |\n| Log-normal | Positive, right-skewed | Income, file sizes |\n\n**Choosing a test:**\n\n| Question | Data | Test | scipy call |\n|---|---|---|---|\n| One mean vs known | Continuous | One-sample t | `ttest_1samp(x, mu)` |\n| Two means | Continuous, normal | Independent t | `ttest_ind(a, b)` |\n| Two means, non-normal | Continuous | Mann-Whitney U | `mannwhitneyu(a, b)` |\n| Paired before\u002Fafter | Continuous | Paired t | `ttest_rel(b, a)` |\n| 3+ means | Continuous | One-way ANOVA | `f_oneway(a, b, c)` |\n| 3+ groups, non-normal | Continuous | Kruskal-Wallis | `kruskal(a, b, c)` |\n| Two categoricals | Counts | Chi-square | `chi2_contingency(t)` |\n\nCheck normality with `shapiro` (n \u003C 5000), equal variance with\n`levene`, and chi-square cell counts >= 5 (otherwise Fisher's exact).\n\nA p-value is the probability of data this extreme if H0 is true — not\nthe probability that H0 is true. A 95% CI means 95% of intervals\nfrom repeated studies contain the true value. p \u003C 0.05 with a tiny\neffect is significant but irrelevant; p = 0.06 is not \"trending.\"\nFor multiple comparisons use Bonferroni or FDR correction.\n\nFor regression, use `statsmodels` OLS for continuous outcomes and\ncheck residual normality, homoscedasticity, and VIF \u003C 5 for\nmulticollinearity. Use logistic regression for binary outcomes;\nexponentiate coefficients for odds ratios.\n\nPractical guidance: small samples (n \u003C 30) call for non-parametric\ntests; large samples (n > 10000) make almost everything significant,\nso focus on effect size. Understand missingness mechanism before\nimputing. Bayesian methods (PyMC, ArviZ) when you actually need\nP(hypothesis | data).\n\n### Visualization guidelines\n\n| Relationship | Chart | When to use |\n|---|---|---|\n| Compare categories | Bar | \u003C 15 categories |\n| Many categories | Horizontal bar | Long labels or > 10 |\n| Trend over time | Line | 1–5 series |\n| Distribution | Histogram | Shape, skew, outliers |\n| Group distributions | Box \u002F violin | Compare medians + spread |\n| Two continuous | Scatter | Correlation, clusters |\n| Part of whole | Stacked bar | Composition over time |\n| Correlation matrix | Heatmap | Many pairwise relations |\n| Geographic | Choropleth \u002F point | Location matters |\n\nDefault rule: bar (categorical) or scatter (continuous) when unsure.\n\nBasic seaborn setup uses `sns.set_theme(style=\"whitegrid\",\npalette=\"colorblind\")`. Categorical palettes: `\"colorblind\"` or\n`px.colors.qualitative.Safe`. Sequential: `viridis`, `cividis`.\nDiverging: `coolwarm`, `RdBu`. Avoid rainbow\u002Fjet and red-green-only\ndistinctions. Never rely on color alone — add patterns, markers, or\nlabels. Test with a colorblindness simulator.\n\nAnnotate the insight, not just the data: peak labels, threshold\nlines, events that explain trend changes. Use direct labels instead\nof legends with 2–3 series. Plotly for interactive (notebooks,\ndashboards); matplotlib\u002Fseaborn for static (reports, papers).\n\nExport checklist: title states the finding, both axes labeled with\nunits, legend or direct labels, colorblind-friendly palette, 300 dpi\nfor print and SVG\u002FPDF for scalable, fonts readable at final size.\n\nCommon mistakes: truncated y-axis on bar charts, dual y-axes, 3D\ncharts, pies with > 4 slices, missing units, default titles like\n\"Figure 1,\" and overplotting (use alpha, hexbin, or 2D density).\n\n### Worked examples\n\n- **CSV of transactions, churn patterns:** Load, inspect\n  shape\u002Fdtypes\u002Fnulls, build a tidy customer time series, run EDA on\n  churn-rate distributions and cohorts, t-test usage frequency\n  between churned and retained groups (with effect size), produce\n  annotated visualizations of key drivers.\n- **Top features for house prices:** Load, clean nulls, encode\n  categoricals, check correlations, fit a baseline linear regression,\n  use permutation importance to rank features, report top 10 with\n  CIs and a bar chart, note multicollinearity concerns.\n- **A\u002FB conversion difference significant?** Compute rates and\n  sample sizes, run chi-square (verify expected cell counts >= 5),\n  report p-value and 95% CI for the difference, visualize rates with\n  error bars, state whether the effect is practically meaningful.\n","# Data Science\n\n## Intro\n\nA solid data analysis flows from import through cleaning, tidying,\nexploration, modeling, and communication — in that order. Tidy data,\nhonest uncertainty, and plain-language findings beat clever models on\nmessy inputs.\n\n## Overview\n\n### Import and cleaning\n\nInspect shape, dtypes, nulls, and duplicates immediately after\nloading. Print `df.info()` and `df.describe()` before any analysis.\nHandle missing data explicitly — drop, fill, or flag, but never ignore\nsilently. Parse dates at load time (`parse_dates=`) and set correct\ndtypes early. Validate assumptions: no negatives where impossible, no\nstring\u002Fnumeric mismatches.\n\n### Tidy data\n\nReshape to tidy form: one variable per column, one observation per\nrow. Use `melt()` for wide-to-long, `pivot_table()` for long-to-wide.\nSplit mixed columns (e.g. `city_state` into two). Normalize repeated\nentities into separate tables. See Level 3 for messy-pattern\nrecipes.\n\n### Exploratory data analysis\n\nStart univariate (histograms, value counts, box plots), then\nbivariate (scatter, grouped bars, correlation). Look for outliers,\nskewness, class imbalance, and unexpected patterns. Use `.groupby()`\nto understand subgroups. Summarize findings in plain language before\nmoving to modeling.\n\n### Statistical reasoning\n\nState the question before choosing a test. Check assumptions\n(normality, equal variance, independence) before applying parametric\ntests. Report effect sizes alongside p-values — statistical\nsignificance is not practical significance. Use confidence intervals\nto communicate uncertainty.\n\n### Feature selection\n\nRemove zero- and near-zero-variance features. Drop one of each highly\ncorrelated pair (r > 0.9). Use domain knowledge first, then\ndata-driven methods (mutual information, permutation importance).\nWatch for leakage — never use target-derived features in training.\n\n### Model selection workflow\n\nStart simple: a mean\u002Fmedian baseline, then linear or logistic\nregression. Add complexity only if simpler models underperform and\nyou understand why. Use train\u002Fvalidation\u002Ftest split or\ncross-validation; never evaluate on training data. Compare on a\nsingle metric chosen before fitting. Document why a model was chosen,\nnot just which one won.\n\n### Communicating results\n\nLead with the finding, not the method. Use plain language: \"customers\nwho used feature X retained 23% more\" beats \"the coefficient was\n0.23.\" Show uncertainty — ranges, intervals, caveats. Include a \"so\nwhat\" — what action does this analysis support? Pin library\nversions, set random seeds, and script everything for\nreproducibility.\n\n### Preferred libraries\n\n| Task | First choice | Alternative |\n|------|-------------|-------------|\n| DataFrames | polars | pandas |\n| Visualization | matplotlib + seaborn | plotly |\n| Statistics | scipy.stats | statsmodels |\n| ML | scikit-learn | xgboost\u002Flightgbm |\n| Notebooks | jupyter lab | marimo |\n\n## Gotchas\n\nAgent-specific failure modes — provider-neutral pause-and-self-check items:\n\n- **Starting with a complex model before a simple baseline.** Training an XGBoost model before computing a mean predictor or running a logistic regression means you can't know how much of the model's value came from complexity vs. data. Build a baseline first; complexity is only justified when the baseline is inadequate.\n- **Evaluating on training data.** \"The model is 98% accurate\" on the data it trained on is not a finding — it is a description of memorization. Always evaluate on held-out data and report the held-out metric.\n- **Confusing statistical significance with practical significance.** With n > 10,000, almost any real difference is statistically significant. A p-value of 0.001 on a 0.1% conversion difference means \"this is real\" but does not mean \"this matters.\" Always report effect size alongside p-values.\n- **Choosing a statistical test before checking its assumptions.** Running a t-test on non-normally distributed data with small samples produces unreliable p-values. Check normality (Shapiro-Wilk for n \u003C 5000), equal variance (Levene), and independence before choosing a parametric test; fall back to non-parametric alternatives when assumptions fail.\n- **Reporting findings without uncertainty.** \"Feature X predicts churn\" without a confidence interval or cross-validation variance means the finding might not replicate. Always communicate uncertainty ranges alongside point estimates.\n- **Truncating the bar chart y-axis to make differences look larger.** A bar chart with a y-axis starting at 0.85 makes a 0.03 difference look dramatic. Unless the baseline is well-understood (like conversion rate from 30% to 32%), bar charts should start at zero.\n- **Not setting random seeds before any analysis.** Non-reproducible analysis is not science. Set `random_state=42` (or equivalent) on all sampling, model initialization, and cross-validation calls. Document the seed in the notebook or script so results can be reproduced exactly.\n\n## Full reference\n\n### Tidy data — messy patterns and fixes\n\n**Headers as values (years across columns).** Melt to long form:\n\n```python\n# pandas\ndf_tidy = df.melt(id_vars=\"country\", var_name=\"year\", value_name=\"population\")\n\n# polars\ndf_tidy = df.unpivot(index=\"country\", variable_name=\"year\", value_name=\"population\")\n```\n\n**Multiple variables in one column** (`male_18-25`). Split:\n\n```python\n# pandas\ndf[[\"gender\", \"age_group\"]] = df[\"demographic\"].str.split(\"_\", expand=True)\n```\n\n**Variables in both rows and columns** (metric in rows, dates in\ncolumns). Melt dates first, then pivot metrics:\n\n```python\nlong = df.melt(id_vars=[\"id\", \"metric\"], var_name=\"month\", value_name=\"value\")\ntidy = long.pivot_table(index=[\"id\", \"month\"], columns=\"metric\", values=\"value\").reset_index()\n```\n\n**Multiple observational units in one table.** Normalize:\n\n```python\nsongs = df[[\"song_id\", \"title\", \"artist\"]].drop_duplicates()\nplays = df[[\"song_id\", \"play_date\", \"user_id\"]]\n```\n\nWide form is fine when feeding correlation matrices, display pivot\ntables, or ML feature matrices. Store and clean tidy; reshape for\nthe consumer.\n\n| Operation | pandas | polars |\n|-----------|--------|--------|\n| Wide to long | `df.melt()` | `df.unpivot()` |\n| Long to wide | `df.pivot_table()` | `df.pivot()` |\n| Split column | `str.split(expand=True)` | `str.split().list.get()` |\n| Deduplicate | `df.drop_duplicates()` | `df.unique()` |\n\n### Statistical methods\n\nAlways compute descriptive stats first (`df.describe()`,\n`value_counts(normalize=True)`, `df.skew()`). Mean is sensitive to\noutliers; median is robust. Standard deviation reports spread in\noriginal units; IQR is the robust analogue.\n\n| Distribution | Use when | Example |\n|---|---|---|\n| Normal | Continuous, symmetric | Heights, measurement error |\n| Binomial | Fixed trials, two outcomes | Conversion rate |\n| Poisson | Counting events in fixed window | Tickets per hour |\n| Exponential | Time between events | Customer arrivals |\n| Log-normal | Positive, right-skewed | Income, file sizes |\n\n**Choosing a test:**\n\n| Question | Data | Test | scipy call |\n|---|---|---|---|\n| One mean vs known | Continuous | One-sample t | `ttest_1samp(x, mu)` |\n| Two means | Continuous, normal | Independent t | `ttest_ind(a, b)` |\n| Two means, non-normal | Continuous | Mann-Whitney U | `mannwhitneyu(a, b)` |\n| Paired before\u002Fafter | Continuous | Paired t | `ttest_rel(b, a)` |\n| 3+ means | Continuous | One-way ANOVA | `f_oneway(a, b, c)` |\n| 3+ groups, non-normal | Continuous | Kruskal-Wallis | `kruskal(a, b, c)` |\n| Two categoricals | Counts | Chi-square | `chi2_contingency(t)` |\n\nCheck normality with `shapiro` (n \u003C 5000), equal variance with\n`levene`, and chi-square cell counts >= 5 (otherwise Fisher's exact).\n\nA p-value is the probability of data this extreme if H0 is true — not\nthe probability that H0 is true. A 95% CI means 95% of intervals\nfrom repeated studies contain the true value. p \u003C 0.05 with a tiny\neffect is significant but irrelevant; p = 0.06 is not \"trending.\"\nFor multiple comparisons use Bonferroni or FDR correction.\n\nFor regression, use `statsmodels` OLS for continuous outcomes and\ncheck residual normality, homoscedasticity, and VIF \u003C 5 for\nmulticollinearity. Use logistic regression for binary outcomes;\nexponentiate coefficients for odds ratios.\n\nPractical guidance: small samples (n \u003C 30) call for non-parametric\ntests; large samples (n > 10000) make almost everything significant,\nso focus on effect size. Understand missingness mechanism before\nimputing. Bayesian methods (PyMC, ArviZ) when you actually need\nP(hypothesis | data).\n\n### Visualization guidelines\n\n| Relationship | Chart | When to use |\n|---|---|---|\n| Compare categories | Bar | \u003C 15 categories |\n| Many categories | Horizontal bar | Long labels or > 10 |\n| Trend over time | Line | 1–5 series |\n| Distribution | Histogram | Shape, skew, outliers |\n| Group distributions | Box \u002F violin | Compare medians + spread |\n| Two continuous | Scatter | Correlation, clusters |\n| Part of whole | Stacked bar | Composition over time |\n| Correlation matrix | Heatmap | Many pairwise relations |\n| Geographic | Choropleth \u002F point | Location matters |\n\nDefault rule: bar (categorical) or scatter (continuous) when unsure.\n\nBasic seaborn setup uses `sns.set_theme(style=\"whitegrid\",\npalette=\"colorblind\")`. Categorical palettes: `\"colorblind\"` or\n`px.colors.qualitative.Safe`. Sequential: `viridis`, `cividis`.\nDiverging: `coolwarm`, `RdBu`. Avoid rainbow\u002Fjet and red-green-only\ndistinctions. Never rely on color alone — add patterns, markers, or\nlabels. Test with a colorblindness simulator.\n\nAnnotate the insight, not just the data: peak labels, threshold\nlines, events that explain trend changes. Use direct labels instead\nof legends with 2–3 series. Plotly for interactive (notebooks,\ndashboards); matplotlib\u002Fseaborn for static (reports, papers).\n\nExport checklist: title states the finding, both axes labeled with\nunits, legend or direct labels, colorblind-friendly palette, 300 dpi\nfor print and SVG\u002FPDF for scalable, fonts readable at final size.\n\nCommon mistakes: truncated y-axis on bar charts, dual y-axes, 3D\ncharts, pies with > 4 slices, missing units, default titles like\n\"Figure 1,\" and overplotting (use alpha, hexbin, or 2D density).\n\n### Worked examples\n\n- **CSV of transactions, churn patterns:** Load, inspect\n  shape\u002Fdtypes\u002Fnulls, build a tidy customer time series, run EDA on\n  churn-rate distributions and cohorts, t-test usage frequency\n  between churned and retained groups (with effect size), produce\n  annotated visualizations of key drivers.\n- **Top features for house prices:** Load, clean nulls, encode\n  categoricals, check correlations, fit a baseline linear regression,\n  use permutation importance to rank features, report top 10 with\n  CIs and a bar chart, note multicollinearity concerns.\n- **A\u002FB conversion difference significant?** Compute rates and\n  sample sizes, run chi-square (verify expected cell counts >= 5),\n  report p-value and 95% CI for the difference, visualize rates with\n  error bars, state whether the effect is practically meaningful.\n",{"name":5,"metadata":53,"description":60},{"processkit":54},{"id":55,"created":56,"version":57,"category":58,"apiVersion":59},"SKILL-data-science",1775433600000,"1.1.0","data-ai","processkit.projectious.work\u002Fv2","Data analysis workflow from import through modeling and communication. Use when analyzing a dataset, exploring data, building a statistical model, selecting features, or communicating findings to stakeholders.\n","main","7231f100e266e30d2c1699ecde5c00602471dc3c","src\u002Fcontext\u002Fskills\u002Fdata-ai\u002Fdata-science","https:\u002F\u002Fgithub.com\u002Fprojectious-work\u002Fprocesskit","https:\u002F\u002Fprojectious-work.github.io\u002Fprocesskit\u002F","https:\u002F\u002Fgithub.com\u002Fprojectious-work","2026-11-11 23:24:18",{"source":8,"skillName":5,"commands":69,"prompt":79,"zipAvailable":80},[70,73,76],{"tool":71,"command":72},"npx","npx skills add projectious-work\u002Fprocesskit --skill data-science",{"tool":74,"command":75},"bunx","bunx skills add projectious-work\u002Fprocesskit --skill data-science",{"tool":77,"command":78},"pnpm","pnpm dlx skills add projectious-work\u002Fprocesskit --skill data-science","安装技能，技能地址：https:\u002F\u002Fgithub.com\u002Fprojectious-work\u002Fprocesskit\u002Fblob\u002Fmain\u002Fsrc\u002Fcontext\u002Fskills\u002Fdata-ai\u002Fdata-science\u002FSKILL.md",true,[82,88,92,96],{"path":83,"resourceType":84,"resourceTypeDesc":85,"sizeBytes":86,"sourceHtmlUrl":87},"references\u002Fstatistical-methods.md","REFERENCE","参考文档",5253,"https:\u002F\u002Fgithub.com\u002Fprojectious-work\u002Fprocesskit\u002Fblob\u002Fmain\u002Fsrc\u002Fcontext\u002Fskills\u002Fdata-ai\u002Fdata-science\u002Freferences\u002Fstatistical-methods.md",{"path":89,"resourceType":84,"resourceTypeDesc":85,"sizeBytes":90,"sourceHtmlUrl":91},"references\u002Ftidy-data-principles.md",3942,"https:\u002F\u002Fgithub.com\u002Fprojectious-work\u002Fprocesskit\u002Fblob\u002Fmain\u002Fsrc\u002Fcontext\u002Fskills\u002Fdata-ai\u002Fdata-science\u002Freferences\u002Ftidy-data-principles.md",{"path":93,"resourceType":84,"resourceTypeDesc":85,"sizeBytes":94,"sourceHtmlUrl":95},"references\u002Fvisualization-guidelines.md",5196,"https:\u002F\u002Fgithub.com\u002Fprojectious-work\u002Fprocesskit\u002Fblob\u002Fmain\u002Fsrc\u002Fcontext\u002Fskills\u002Fdata-ai\u002Fdata-science\u002Freferences\u002Fvisualization-guidelines.md",{"path":97,"resourceType":98,"resourceTypeDesc":99,"sizeBytes":14,"sourceHtmlUrl":100},"scripts\u002F.gitkeep","SCRIPT","脚本","https:\u002F\u002Fgithub.com\u002Fprojectious-work\u002Fprocesskit\u002Fblob\u002Fmain\u002Fsrc\u002Fcontext\u002Fskills\u002Fdata-ai\u002Fdata-science\u002Fscripts\u002F.gitkeep",[102,109,116,122,145,165],{"skillNo":103,"name":104,"nameZh":6,"description":6,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":105,"categories":106,"keywords":107,"publishedAt":108},"01M7665G2ED3624R3X93304265","standard-ui-design","https:\u002F\u002Fgithub.com\u002Fprojectious-work\u002Fprocesskit\u002Fblob\u002Fmain\u002F.agents\u002Fskills\u002Fstandard-ui-design\u002FSKILL.md",[],[],"2026-11-12 13:55:43",{"skillNo":110,"name":111,"nameZh":6,"description":112,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":113,"categories":114,"keywords":115,"publishedAt":108},"01M76651QTC6W0HJ4AFW2MMR3V","standard-agent-use","Company standard for working with AI agents. Use when deciding what to delegate to agents, capturing open work, keeping agent memory or knowledge notes, improving an agent workflow, or introducing agent work in a team or for a client.","https:\u002F\u002Fgithub.com\u002Fprojectious-work\u002Fprocesskit\u002Fblob\u002Fmain\u002F.agents\u002Fskills\u002Fstandard-agent-use\u002FSKILL.md",[],[],{"skillNo":117,"name":118,"nameZh":6,"description":6,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":119,"categories":120,"keywords":121,"publishedAt":108},"01M76650NN4T8P241021NBR64S","standard-agent-building","https:\u002F\u002Fgithub.com\u002Fprojectious-work\u002Fprocesskit\u002Fblob\u002Fmain\u002F.agents\u002Fskills\u002Fstandard-agent-building\u002FSKILL.md",[],[],{"skillNo":123,"name":124,"nameZh":6,"description":125,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":126,"categories":127,"keywords":129,"publishedAt":48},"01M6X5MTC89B4ZGXW3JVSQCPF8","skill-management","Manage Skill package manifests as v1 ontology entities.","https:\u002F\u002Fgithub.com\u002Fprojectious-work\u002Fprocesskit\u002Fblob\u002Fmain\u002Fsrc\u002Fcontext\u002Fskills\u002Fprocesskit\u002Fskill-management\u002FSKILL.md",[128],{"categoryNo":24,"categoryName":25,"taxonomyType":21,"taxonomyTypeDesc":22},[130,131,132,133,134,135,42,136,137,138,139,140,124,141,142,143,144],"management","skill management","ontology","v1","create","as","transition","get","manage","update","alpha","list","package","skill package","manifests",{"skillNo":146,"name":147,"nameZh":6,"description":148,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":149,"categories":150,"keywords":152,"publishedAt":48},"01M6X5MCV74NHW5MEDNHZRF2C8","schema-management","Manage generated processkit schema contracts through deterministic MCP rebuild and inspection workflows.","https:\u002F\u002Fgithub.com\u002Fprojectious-work\u002Fprocesskit\u002Fblob\u002Fmain\u002Fsrc\u002Fcontext\u002Fskills\u002Fprocesskit\u002Fschema-management\u002FSKILL.md",[151],{"categoryNo":24,"categoryName":25,"taxonomyType":21,"taxonomyTypeDesc":22},[153,154,130,155,156,157,158,159,160,161,162,163,42,137,138,140,164],"schema","schema management","contract","rebuild","contracts","workflows","wed","generated","MCP","validation","deterministic","mode",{"skillNo":166,"name":167,"nameZh":6,"description":168,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":169,"categories":170,"keywords":171,"publishedAt":48},"01M6X5M5G6YJ96AMZGSCZJ1D13","proposition-management","Record claims, risks, beliefs, world facts, WSJF estimates, and assumptions through the Proposition interface.","https:\u002F\u002Fgithub.com\u002Fprojectious-work\u002Fprocesskit\u002Fblob\u002Fmain\u002Fsrc\u002Fcontext\u002Fskills\u002Fprocesskit\u002Fproposition-management\u002FSKILL.md",[],[172,130,173,174,134,175,176,177,42,137,178,179,139,180,140,181,182,183,184,185],"proposition management","proposition","interface","claims","record","world","creation","event","WSJF","index","updates","query","facts","log"]