Agent Skills
› NeverSight/learn-skills.dev
› figure-results-review
figure-results-review
GitHub用于在论文、报告或演示前审核ML/AI实验图表、表格及结果叙述。检查数据完整性、视觉一致性、统计证据及图文匹配度，确保结果展示清晰准确并符合目标会议规范。
Trigger Scenarios
用户拥有实验结果图表、表格或截图需要审核
需要检查论文声明与实际展示证据的一致性
调整图表视觉风格以符合特定会议或幻灯片规范
准备回复审稿人或向导师汇报的关键结果可视化
Install
npx skills add NeverSight/learn-skills.dev --skill figure-results-review -g -y
SKILL.md
Frontmatter
{
"name": "figure-results-review",
"description": "Review ML or AI experiment figures, tables, plots, captions, result narratives, and paper visual style before they are shown in a paper, advisor meeting, report, slide deck, rebuttal, or submission. Use this skill whenever the user has experimental results, plots, tables, metrics, screenshots, captions, draft result sections, or wants to audit figure style choices such as color, typography, markers, symbols, line widths, sizing, and venue-consistent visual conventions.",
"allowed-tools": "Read, Write, Edit, Bash, Glob, WebSearch, WebFetch",
"argument-hint": "[project-dir-or-results-file] [--mode paper|meeting|slide|rebuttal|diagnosis] [--venue <venue>]"
}
Figure Results Review
Audit figures, tables, plots, captions, and result narratives before they become paper evidence or meeting material.
Use this skill when:
- the user has a figure, table, plot, result screenshot, caption, result section, or slide with experimental evidence
- a paper claim needs to be checked against the actual displayed evidence
- a plot may be missing baselines, error bars, seeds, labels, units, or fairness context
- a table layout hides the main comparison or makes the result look weaker than it is
- paper figures need a consistent visual style: color palette, markers, symbols, line widths, fonts, sizing, and notation
- new results require deciding whether to update writing, rerun experiments, diagnose failures, or narrow claims
- a rebuttal needs a clean result table or concise visual answer
- an advisor meeting needs figures that make the decision obvious
Do not use this skill to design experiments from scratch. Use experiment-design-planner before results exist. Use result-diagnosis when the primary issue is why a result is surprising or broken. Use conference-writing-adapter when the main task is prose style after the evidence is already accepted.
Pair this skill with:
paper-evidence-boardwhen figures/tables must be linked to paper claims, sections, reviewer risks, and actionsresult-diagnosiswhen a plotted result is suspicious, unstable, negative, or contradictorybaseline-selection-auditwhen the visual exposes missing, weak, or unfair baselinesexperiment-design-plannerwhen the fix requires new experiments, ablations, controls, or metricsexperiment-report-writerwhen raw results need a structured report before figure reviewconference-writing-adapterwhen the final figure narrative or visual style must be adapted to a target venueresearch-project-memorywhen claim/evidence/risk/action updates should persist across sessions
Skill Directory Layout
<installed-skill-dir>/
├── SKILL.md
└── references/
├── caption-and-narrative.md
├── claim-support.md
├── memory-writeback.md
├── paper-visual-style.md
├── report-template.md
├── statistical-evidence.md
└── visual-integrity.md
Progressive Loading
- Always read
references/claim-support.md,references/visual-integrity.md, andreferences/statistical-evidence.md. - Read
references/paper-visual-style.mdwhen figures/tables are intended for a paper, slide deck, rebuttal, camera-ready, or venue-specific rewrite. - Read
references/caption-and-narrative.mdwhen revising captions, result prose, slide text, or paper figure callouts. - Read
references/report-template.mdbefore writing the final review. - Read
references/memory-writeback.mdwhen the project hasmemory/, component.agent/folders, or the user asks for persistent project memory. - If the expected plotting or table conventions depend on a target venue, benchmark, or recent paper style, verify with current accepted papers, official benchmark protocols, or user-provided exemplars.
- If the actual image/table cannot be inspected, audit the provided data/caption/prose and clearly mark visual-layout judgments as unverified.
Core Principles
- A figure is evidence for a specific claim, not decoration.
- Every plot/table should answer one reviewer question.
- The main comparison should be visually and numerically easy to find.
- Captions must state enough setup for the result to be interpreted without searching the paper.
- Statistical uncertainty, seeds, and variance matter when the claim depends on small differences.
- Compute, data, baseline, and protocol fairness must be visible when they affect interpretation.
- Paper figures should share a deliberate visual language. Style choices are part of writing because they control what reviewers notice first.
- A beautiful plot that does not support the claim should be revised or cut.
- New results must update claims, writing, reviewer risks, and next actions.
Step 1 - Recover Evidence Context
Collect:
- figure/table file path, screenshot, raw data, caption, or result prose
- paper claim or section the result is meant to support
- experiment setup: dataset, model, baseline, metric, seed, split, hyperparameters, protocol
- target audience: paper, advisor meeting, slide, rebuttal, internal report, or appendix
- target venue or benchmark expectations
- current paper location, if any
- linked project memory IDs such as
CLM-###,EVD-###,FIG-###,TAB-###,RSK-###, orACT-###
Rewrite the intended evidence relation:
This figure/table is supposed to show that [claim] because [metric/comparison/trend] under [setup].
If that sentence cannot be written, route to paper-evidence-board before polishing the visual.
Step 2 - Audit Claim Support
Read references/claim-support.md.
For each figure or table, answer:
- what exact claim does it support?
- is the displayed evidence sufficient for that claim?
- is the claim too broad for the measured setup?
- are baselines, ablations, controls, or diagnostics missing?
- does the result contradict another figure, table, or section?
- is the result main-paper material, appendix material, diagnostic material, or not ready?
Assign one status:
supports-claimsupports-narrower-claimambiguouscontradicts-claimdiagnostic-onlynot-ready
Step 3 - Audit Visual and Table Integrity
Read references/visual-integrity.md.
Check:
- axes, labels, units, scales, and transformations
- legend readability and method names
- ordering of methods, datasets, metrics, and ablations
- whether the main result is visually salient
- table grouping, bolding, decimals, missing values, and footnotes
- whether color, markers, line styles, or hatching remain readable in grayscale
- whether figure size works for one-column, two-column, slide, or appendix usage
- whether captions and labels match the actual plotted data
Flag any issue that could cause a reviewer to misread the result.
Step 4 - Audit Paper Visual Style
Read references/paper-visual-style.md when the output is paper-facing.
Check:
- color palette and colorblind/grayscale robustness
- stable method-to-color and method-to-marker mapping across all figures
- line width, marker size, hatch, symbol, and notation consistency
- font size, tick density, label length, and final-column readability
- figure dimensions for one-column, two-column, appendix, or slide use
- whether visual emphasis matches the paper's claim hierarchy
- whether the main method is recognizable without relying only on color
- whether theorem/method symbols in plots match the paper notation
If the paper has no visual style policy, propose one and record it in paper/.agent/ or .agent/conference-writing/project-style.md when appropriate.
Step 5 - Audit Statistical and Experimental Evidence
Read references/statistical-evidence.md.
Check:
- number of seeds or repeated runs
- variance, confidence intervals, standard deviation, or standard error
- significance or effect-size interpretation when differences are small
- data split and leakage risk
- metric definition and averaging
- baseline fairness and tuning budget
- compute or speed reporting when efficiency is part of the claim
- failure cases or negative results that should be shown
If the plot lacks necessary uncertainty, decide whether to rerun, add error bars, weaken the claim, or move the result to appendix/diagnostic status.
Step 6 - Review Caption and Result Narrative
Read references/caption-and-narrative.md when output text needs revision.
For each figure/table, produce:
- caption diagnosis
- revised caption or caption outline
- one-sentence paper callout
- claims to avoid in nearby prose
- reviewer question answered
- missing setup details to add
Captions should not oversell. They should state the setup, comparison, metric, and takeaway.
Step 7 - Route Fixes
For every issue, route to one or more actions:
edit-figure: labels, ordering, scale, legend, layout, or visual emphasisedit-table: grouping, decimals, bolding, footnotes, missing values, or row/column orderrewrite-caption: setup, metric, takeaway, caveat, or claim alignmentrewrite-results-text: nearby paper prose overclaims or misses the takeawaydefine-visual-style: missing or inconsistent paper visual style policyrestyle-figure: color, marker, line width, font size, symbol, panel layout, or emphasisrerun: missing seeds, variance, baseline, metric, or protocoldiagnose-result: suspicious, negative, unstable, or contradictory patternbaseline-audit: missing or unfair baselinenarrow-claim: evidence only supports a smaller statementmove-to-appendix: useful but not central enough for main papercut: visual does not support a paper need
Name the next skill when appropriate.
Step 8 - Write the Review Report
Read references/report-template.md.
If saving to a project and no path is given, use:
docs/results/figure_results_review_YYYY-MM-DD_<short-name>.md
The report must include:
- figure/table inventory
- claim-support status
- visual/table integrity issues
- visual style policy and consistency issues
- statistical evidence issues
- caption and narrative fixes
- reviewer-risk forecast
- routed actions and next skills
- memory update section
Step 9 - Write Back to Project Memory
Read references/memory-writeback.md when memory exists.
Update the smallest useful set of entries:
memory/evidence-board.md: figure/table evidence status, setup, and linked claimsmemory/claim-board.md: claims supported, narrowed, contradicted, or not readymemory/risk-board.md: reviewer risks from visual ambiguity, missing uncertainty, weak baselines, or overclaimingmemory/action-board.md: figure edits, reruns, caption fixes, result diagnosis, or claim revisionspaper/.agent/: figure/table map, paper locations, caption state, and stale visual warnings.agent/conference-writing/project-style.md: venue-facing figure style decisions when conference adaptation is active- worktree
.agent/worktree-status.md: result-generation or plotting tasks and exit conditions
Use certainty labels:
verifiedfor values checked against raw data, logs, or source figuresuser-statedfor user-supplied contextinferredfor reviewer-risk and narrative judgmentsunverifiedfor visual or statistical claims that could not be inspected
Final Sanity Check
Before finalizing:
- every figure/table has a linked claim and reviewer question
- main comparison is easy to find
- axes, units, legends, captions, and table labels are unambiguous
- colors, markers, fonts, symbols, and figure sizes are consistent across the paper
- uncertainty is present or the lack of uncertainty is justified
- baseline and compute fairness are visible when relevant
- overclaims are narrowed
- fixes are routed to concrete next actions or skills
- project memory is updated when present
Version History
- e0220ca Current 2026-07-05 21:35
