results-presentation — скилл для Claude Code: отчёты по результатам
Скилл results-presentation для Claude Code, который оформляет эконометрические результаты в графики, таблицы и текстовую интерпретацию. Запускается на этапе подготовки отчёта или по просьбе показать результаты и подготовить их к публикации.
Apache-2.0 ★ 89 обновлён 2026-11-12
npx skills add bis-med-it/marco-BIS --skill results-presentation -a claude-codeЧто делает results-presentation
Скилл превращает результаты эконометрического анализа в графики, таблицы и текстовую интерпретацию. Основной формат отчёта — интерактивный HTML; графики в PDF и таблицы LaTeX можно запросить отдельно.
Вывод подстраивается под аудиторию: например, для журнальной статьи, рабочего документа, краткой записки или презентации. Генерацией HTML-отчёта занимается скрипт scripts/lib/report.py. PDF и LaTeX нужны только для соответствующих запросов.
Какие задачи решает
- показать результаты эконометрической оценки на графиках
- свести коэффициенты или диагностику в таблицу
- подготовить текстовую интерпретацию результатов
- собрать результаты в интерактивный HTML-отчёт
Примеры запросов агенту
- Подготовь интерактивный HTML-отчёт по результатам эконометрической оценки.
- Построй графики и таблицы для журнальной статьи.
- Напиши интерпретацию результатов для краткой записки.
Как установить results-presentation
1Через Skills CLI (Vercel)
npx skills add bis-med-it/marco-BIS --skill results-presentation -a claude-code2Вручную: положить файлы
~/.claude/skills/results-presentation/SKILL.mdПапка скилла целиком; для одного проекта — .claude/skills/results-presentation/
1Через Skills CLI (Vercel)
npx skills add bis-med-it/marco-BIS --skill results-presentation -a codex2Вручную: положить файлы
.agents/skills/results-presentation/SKILL.mdДля всех проектов — ~/.agents/skills/
1Через Skills CLI (Vercel)
npx skills add bis-med-it/marco-BIS --skill results-presentation -a cursor2Вручную: положить файлы
.cursor/skills/results-presentation/SKILL.mdCursor читает и .agents/skills/, и .claude/skills/
1GitHub CLI
gh skill install bis-med-it/marco-BIS skills/results-presentation2Вручную: положить файлы
.github/skills/results-presentation/SKILL.md1Через Skills CLI (Vercel)
npx skills add bis-med-it/marco-BIS --skill results-presentation -a gemini-cli2Командой платформы
gemini skills install https://github.com/bis-med-it/marco-BIS --path skills/results-presentationВручную: положить файлы
.devin/skills/results-presentation/SKILL.mdСтарый путь .windsurf/skills/ тоже читается
Claude.ai
Упакуйте папку в zip и загрузите в Settings → Capabilities → Skills
По шагам, с командами для каждой платформы: Как установить скилл в Claude Code: пять способов
Вопросы о results-presentation
Как установить results-presentation в Claude Code?
Выполни npx skills add bis-med-it/marco-BIS --skill results-presentation -a claude-code. Файл скилла находится по пути ~/.claude/skills/results-presentation/SKILL.md.
Что делает results-presentation?
Оформляет эконометрические результаты в графики, таблицы и текстовую интерпретацию. Основной результат — интерактивный HTML-отчёт.
Какие форматы может подготовить results-presentation?
По умолчанию результатом служит интерактивный HTML-отчёт. По запросу скилл также готовит графики в PDF и таблицы LaTeX.
Исходный текст
SKILL.md · Apache-2.0
---
name: results-presentation
description: >
Generates charts, tables, and narrative interpretation of
econometric results. Primary output is an interactive HTML report (using
scripts/lib/report.py); PDF charts and LaTeX tables are optional, on request
(e.g. for journal or print workflows). Adapts output to the target audience. Use whenever the
orchestrator reaches the reporting stage, when the user asks for "charts",
"plots", "tables", "write up the results", "prepare for publication", or when
results need to be communicated in any format.
---
# Results presentation
## Purpose
Translate raw econometric output into clear and accurate communication.
This skill handles three layers: visual presentation (charts and tables), written
interpretation (narrative), and format adaptation (who is the audience).
## Design principles
1. **Clarity over decoration.** Charts should convey information, not impress with
complexity. Follow Tufte's principles: maximise data-ink ratio, avoid chartjunk.
2. **Accuracy over persuasion.** Present uncertainty prominently. Never hide wide
confidence bands or inconvenient results. If the price puzzle is there, show it.
3. **Audience-appropriate.** The same impulse response means different things to a
referee, a governor, and a journalist. Adapt language and detail, not substance.
4. **Reproducible.** Every chart and table is generated by a script saved in the
research log. No manual post-processing.
5. **House style aware.** Consult the knowledge base for institutional formatting
conventions before applying defaults.
## Inputs
```
presentation_request:
results: structured output from any estimation method
audience: "journal" | "working_paper" | "policy_brief" | "presentation" | "internal"
outputs_requested:
- "irf_charts"
- "fevd_charts"
- "fevd_tables"
- "coefficient_table"
- "diagnostics_table"
- "narrative_summary"
- "historical_decomposition"
- "comparison_charts" (LP vs VAR, or across specifications)
- "time_varying_panels" (for TVP-VAR)
- "robustness_appendix"
style: (optional, agent-side; VARReport reads only line and band colours and
band alphas, from config `presentation.chart`)
colour_palette: list of hex colours
font: string
figure_width: float (inches)
figure_height: float (inches)
dpi: integer
line_width: float
band_alpha: float (transparency for confidence bands)
language: "en" | "es" | "th" | "ja" | ... (for axis labels and narrative)
```
## Charts
### Impulse response functions
The most important chart in any VAR analysis. Getting it right matters.
**Layout:**
- One panel per response variable, arranged in a grid.
- If multiple shocks are identified, one figure per shock (each containing panels
for all response variables).
- Horizon on the x-axis, response magnitude on the y-axis.
- Zero line always visible and prominent.
**Adaptive layout (must scale with `n_vars`):**
Layout parameters (grid dimensions, panel sizes, font sizes, inter-panel
spacing, legend placement, and margins) should all be derived from the
panel count rather than hardcoded. A 6-variable VAR and a 12-variable BVAR
both render cleanly without manual tuning per spec. The reference
implementation lives in `scripts/lib/report.py::_smart_layout` and applies
the rules below; any other implementation (R / Julia / Matlab) should
follow the same principles.
| n panels | Grid (rows × cols) | Panel target | Font (main) | Notes |
|---|---|---|---|---|
| 1 | 1 × 1 | 380 × 320 px | 13 pt | wide single panel |
| 2 | 1 × 2 | 340 × 280 px | 12 pt | |
| 3 | 1 × 3 | 300 × 260 px | 12 pt | |
| 4 | 2 × 2 | 300 × 260 px | 12 pt | not 3 + 1 |
| 5-6 | 2 × 3 | 280 × 240 px | 11 pt | |
| 7-9 | 3 × 3 | 260 × 220 px | 11 pt | |
| 10-12 | 3 × 4 | 240 × 200 px | 10 pt | |
| 13-16 | 4 × 4 | 220 × 180 px | 10 pt | |
| > 16 | ⌈n/⌈√n⌉⌉ × ⌈√n⌉ | 200 × 170 px | 9 pt | densest |
Rules that apply uniformly:
- **X-axis label only on bottom-most panel of each column.** Eliminates
"Quarters" appearing under every panel, which is the canonical cause
of the next row's subplot title colliding with the label above it.
- **Y-axis title only on leftmost column.** By default the response variable
name lives here and subplot titles are kept short / empty. When the responses
share a common unit (e.g. percent), pass `units="Percent"` to `add_irf` /
`add_standard_sections`: the y-axis then shows the unit and the variable name
moves to the subplot title, avoiding the name-twice redundancy and matching
published IRF figures (this is the "Y-axis: Percent" + named-panel form in the
Annotation list below).
- **Subplot titles wrap to 2 lines** at the word boundary, threshold
≈ `panel_width_px / 9`. Single very-long words get an ellipsis rather
than mid-word split.
- **Inter-panel spacing computed in absolute pixels**, not Plotly's
default normalised fraction. Otherwise the gap shrinks as `n_rows`
grows, causing subplot titles to collide with axes above.
- **Legend on the right side when `n_rows ≥ 3`** (vertical), below
otherwise (horizontal). The horizontal-below placement collides with
bottom x-axis labels for tall grids.
- **Margins scaled to content**: top margin reserves space for the
figure-level title (shock name); bottom margin grows when the legend is
horizontal; right margin grows when the legend is vertical-right;
left margin reserves space for y-axis titles.
- **Tick label fonts scale down 1-2 pt below the subplot title font.**
**Bands:**
- Default: 68% credible/confidence interval as dark shading, 90% as light shading.
- Posterior median (Bayesian) or point estimate (frequentist) as a solid line.
- Never show only the point estimate without uncertainty bands, unless the
method provides no interval inference (`bands="none"`, e.g. pooled panel
fits). Then mark the result clearly as point-only (see "Significance is
never fabricated" below).
**Annotation:**
- Panel titles: variable names in plain language (not series codes).
- Axis labels: "Quarters" or "Months" (not "h" or "horizon").
- Y-axis: units or "Percent" or "Percentage points" as appropriate.
- Shock description in the figure title: "Response to a one-standard-deviation
monetary policy shock" (not "IRF of shock 1").
**Style defaults (journal):** the repo-root `config.yaml` holds them under
`presentation.chart` (figure and panel size, line and band colours, band
alphas, zero-line colour, line width, serif font, font sizes, 300 dpi, PDF).
Pass the loaded config to `VARReport(config=...)`: the HTML report uses the
line and band colours and band alphas, and a matplotlib PDF export should use
the rest, with a dashed zero line, no grid, and only the bottom and left
spines.
**Audience adaptations:**
| Element | Journal | Policy brief | Presentation |
|---------|---------|--------------|--------------|
| Font | Serif, 9-11pt | Sans-serif, 10-12pt | Sans-serif, 14-18pt |
| Colour | Muted, print-safe | Institutional palette | High contrast |
| Annotations | Minimal, in caption | Inline, plain language | Large, with arrows |
| Bands shown | 68% and 90% | 68% only (simpler) | 68% only, wider lines |
| Format | PDF (vector) | PDF | PNG (for slides) |
| Caption | Technical | Interpretive | Omitted (verbal) |
| Size | Column or page width | A4 page | 16:9 slide |
### Forecast error variance decomposition
**Table format (journal):**
| Horizon | Monetary policy | Demand | Supply | ... |
|---------|----------------|--------|--------|-----|
| 1Q | 85.2 | 10.1 | 4.7 | ... |
| 4Q | 52.3 | 28.7 | 19.0 | ... |
| 8Q | 34.1 | 38.2 | 27.7 | ... |
| 20Q | 22.5 | 41.0 | 36.5 | ... |
Values in percent. Include posterior median and (in parentheses) 68% credible
interval if Bayesian.
**Chart format:** stacked area chart or grouped bar chart at selected horizons.
### Historical decomposition
Stacked bar or area chart showing the contribution of each structural shock to
the historical evolution of a variable. Base (deterministic) component shown
separately. Time on x-axis, contribution on y-axis.
Useful for narrative interpretation: "The 2008 decline in output was primarily
driven by financial shocks (blue), with a smaller contribution from demand
shocks (orange)."
### Time-varying parameter charts (TVP-VAR)
- **IRF panels across time:** small multiples, one per selected date, with
consistent y-axis scaling for comparability.
- **Parameter evolution:** time on x-axis, coefficient value on y-axis, with
posterior credible bands. Annotate regime changes (vertical lines with labels).
- **Volatility evolution:** similar layout, with shading for known crisis periods.
### Comparison charts
When presenting results from multiple methods (LP vs VAR, different priors,
different samples):
- Overlay on the same axes with distinct line styles and colours.
- Legend must be clear and unambiguous.
- Do not use more than 3-4 methods on a single panel; split if needed.
- Annotate where methods agree and disagree.
## Tables
### Coefficient tables
For journal submission, follow the standard format:
- Variables as rows, equations as columns.
- Coefficients with standard errors in parentheses below.
- Significance stars: * p<0.10, ** p<0.05, *** p<0.01.
For Bayesian: report posterior probability of the coefficient being positive/negative
instead of stars, or credible intervals.
- Bottom rows: sample size, log-likelihood, information criteria.
Generate as LaTeX (for journal, on request), markdown (for working paper), or formatted
spreadsheet (for internal use).
### Diagnostics summary table
Compact table reporting all diagnostic test results:
| Test | Statistic | p-value | Verdict |
|------|-----------|---------|---------|
| Ljung-Box (lag 4) | 12.3 | 0.42 | Pass |
| ARCH-LM | 8.7 | 0.07 | Marginal |
| Jarque-Bera | 5.2 | 0.08 | Pass (soft) |
| Max eigenvalue | 0.94 | n/a | Stable |
Use colour coding for presentations (green/amber/red) but not for journal
submissions.
## Narrative interpretation
### Purpose
The agent writes a plain-language interpretation of the results, calibrated to
the audience. This is not a mechanical description of charts; it is an economic
narrative that connects the statistical findings to the research question.
**The report never ships without one.** `VARReport.add_standard_sections(...)`
**warns** if no narrative is supplied (a report with charts and tables but
no interpretation is a defect). Passing `narrative="auto"` drafts an
economics-tied skeleton via `lib.draft_narrative`, built from the IRF
signs/magnitudes, FEVD shares, companion stability, and the Stage-5 sensibility
`flags` (price puzzle, explosive roots, …), which the agent then **refines**
into the prose below. The skeleton ends with an explicit `[DRAFT …]` marker so
an un-refined narrative is never mistaken for finished work. Always refine it;
never publish the raw skeleton.
### Structure
**For a journal paper (results section):**
1. Open with the main finding in one sentence.
2. Walk through the key impulse responses, referencing figure numbers.
3. Note where results align with or depart from prior literature.
4. Discuss uncertainty: are the bands tight enough to draw firm conclusions?
5. Address any anomalies (price puzzle, counterintuitive signs) with economic
reasoning.
6. Summarise FEVD results and what they imply for the relative importance of shocks.
7. Note robustness: do LP confirm the VAR? Is the result prior-sensitive?
**For a policy brief:**
1. Lead with the policy implication: "A 100bp rate increase is associated with
a [X]% decline in output after [Y] quarters."
2. Contextualise: "This is [consistent with / larger than / smaller than]
estimates from comparable economies."
3. Caveat plainly but briefly: "These estimates are subject to uncertainty;
the true effect could range from [A] to [B]."
4. Avoid: jargon, Greek letters, references to estimation method.
**For an internal presentation:**
1. Bullet-point key findings with chart references.
2. "What's new" framing: how do these results differ from previous estimates?
3. Robustness summary in one slide.
4. Clear "so what" for the policy decision at hand.
### Tone and language
- **Never overstate certainty.** "The results suggest" not "the results prove."
- **Use magnitudes.** "A one-percentage-point increase in the policy rate is
associated with a 0.3% decline in output", not just "output falls."
- **Contextualise with the literature.** "This is consistent with Christiano,
Eichenbaum, and Evans (1999), who find a similar peak response at 6 quarters."
- **Acknowledge limitations.** Every narrative should include at least one relevant
caveat about the analysis.
### Language support
The narrative can be generated in any language specified in the knowledge base
or the presentation request. For non-English output:
- Use locally conventional economic terminology.
- Follow local citation conventions.
- Adapt number formatting (decimal commas vs points, thousand separators).
- The agent should note if it is less confident in domain-specific terminology
in the requested language and suggest the user review the phrasing.
## Output formats
### HTML report (primary)
Every model run **must** produce a single-file HTML report as the primary output.
Use `scripts/lib/report.py` (`VARReport` class) to generate it. The report combines
all results into a single interactive file.
```python
import sys
from pathlib import Path
sys.path.insert(0, str(Path(__file__).resolve().parent.parent.parent)) # scripts/ on the path from scripts/<tier>/<id>/
from lib.report import VARReport
report = VARReport(
title="[Paper] [Model] Replication",
subtitle="[n]-variable [model], [identification], [dataset] [sample]",
config=cfg, # pass config.yaml dict for style consistency
)
report.add_specification(spec_dict) # model details table
report.add_irf(irf_point, ci_68_lo, ci_68_hi, # interactive Plotly IRF panels
ci_90_lo, ci_90_hi, # tabbed by shock
var_names, shock_labels)
report.add_diagnostics(diag_dict, var_names) # color-coded p-value table
report.add_summary(summary_rows, # peak responses, significance
title="Technology Shock Summary") # name the shock
report.add_fevd(fevd_data, var_names, horizons, # heat-mapped FEVD table
shock_names=shock_labels) # columns are SHOCKS, not variables
report.add_robustness(baseline_irf, alt_irf, # overlay comparison chart
labels, var_names)
report.add_narrative(text) # plain-language interpretation
report.add_research_log(log_dict) # collapsible JSON metadata
report.save(OUTPUT_DIR / "report.html")
```
**When the user asks to drop warnings or caveats:** presentation *framing*
is the user's call: banners, "anomaly" labels, remedies sections can go.
Three things are not negotiable: (i) the narrative still describes the
estimates factually (a result a reader can see in the charts cannot be
written around); (ii) one neutral sentence of standard interpretation
stays where the literature would immediately raise it; (iii) the research
log keeps the full record (the flag, the user's decision, verbatim). Say
this boundary out loud when applying it: "your specification, your
disposition, but the report stays truthful."
Frequency and availability caveats for the standard sections:
- **Monthly runs must override the quarterly defaults** in
`add_standard_sections`: pass `x_axis_label="Months"` and monthly
`fevd_horizons` (e.g. `(1, 6, 12, 24, 48)`); the defaults are
`"Quarters"` and `(1, 4, 8, 20, 40)`.
- **FEVD is not always available.** `lib.structural_fevd` returns a
single-horizon `(n_vars, n_shocks)` matrix (not an IRF-shaped 3-D
array), and the FEVD requires unit-variance structural shocks: for a
proxy-SVAR IRF normalised as `unit_policy` or `one_sd`, the shock's
variance scale is not pinned down (`reliability` is the unit-variance
option). `add_standard_sections` then inserts a visible "FEVD omitted"
note in the report (it is not silently skipped); echo the reason in
the narrative rather than presenting a partial table.
- **Significance is never fabricated.** CI-based significance markers
require finite interval bounds of the same sign; a run without
interval inference (`bands="none"`, e.g. pooled panel fits, which are
point-only) renders "N/A (no inference)" in the summary tables rather
than stars or empty bands.
The report is a single HTML file (~300KB) that loads Plotly from a CDN: one file to share, but charts need internet on first open (air-gapped users: bundle plotly.js locally).
**When to add each section:**
- `add_specification` and `add_research_log`: always (every model)
- `add_irf`: when the analysis computes identified impulse responses
- `add_diagnostics`: diagnostics appropriate to the fitted method; use
`lib.diagnostics` only for supported VARFit-like objects, and mark
sections that do not apply as not applicable, with the reason. The table
is colour-coded: green (p>0.10), amber, red (p<0.05)
- `add_summary` and `add_fevd`: when Cholesky or structural IRFs are computed
- `add_robustness`: when alternative specifications are run
- `add_narrative`: always; the agent writes (or refines a `narrative="auto"`
draft of) the interpretation. `add_standard_sections` warns if it's omitted.
### LaTeX + PDF (optional, on request)
Generate alongside the HTML report only when the user asks (e.g. for journal
or print workflows). The audience setting alone does not trigger it:
- IRF charts as PDF via matplotlib (using config.yaml style settings).
- Tables as `.tex` files with booktabs formatting.
- Results section draft as `.tex`.
### Markdown
For working papers. Charts as PNG, tables in markdown, narrative as markdown text.
### Presentation slides
Delegate to the `pptx` skill if available. One chart per slide, speaker notes
with the full narrative.
## House style integration
Before generating any output, check `knowledge-base/context/conventions.md` for:
- Institutional colour palette.
- Required fonts.
- Logo placement rules.
- Specific chart formatting conventions.
- Preferred table format (e.g., three-line tables, booktabs style).
- Language and terminology conventions.
If no house style is specified, use the skill's defaults (clean, professional,
journal-ready).
## Quality checklist
Before presenting any chart or table to the user, verify:
- [ ] All axes are labelled with units.
- [ ] The zero line is visible in IRF charts.
- [ ] Uncertainty bands are shown (point-only results only when the method
provides no inference, and then clearly marked as such).
- [ ] Variable names are human-readable (not codes).
- [ ] The figure title describes what the chart shows.
- [ ] The colour scheme is accessible (colourblind-safe palette if not overridden).
- [ ] Resolution is sufficient for the target format (300+ dpi for print).
- [ ] The chart does not misrepresent the results (e.g., truncated y-axis
exaggerating small effects).
- [ ] Number formatting follows local conventions.
## Sensibility check on communication
Just as the orchestrator checks results for economic sensibility, this skill
checks the presentation for communication sensibility:
1. **Does the narrative match the charts?** If the text says "output declines
significantly" but the 90% band includes zero at all horizons, flag this.
2. **Is uncertainty fairly represented?** If bands are very wide, the narrative
should not overstate the findings.
3. **Is the audience appropriate?** If the user requests a policy brief but the
results are highly uncertain and model-dependent, suggest caveating more
heavily or recommending further analysis before communicating to policymakers.
4. **Is the specification search disclosed?** Per the orchestrator's Stage-5
disclosure rules: the narrative must state how many sensibility-driven
revisions preceded the final specification (including "none"), show the
first-pass result alongside the final one when any revision occurred, and
say in one sentence that the reported bands are conditional on that
search (post-selection, not nominal-coverage, intervals). A report whose
results were iterated into shape without saying so fails this check.
© bis-med-it · Apache-2.0 · bis-med-it/marco-BIS