---
name: visualize-results
description: Create charts and diagrams from saved Bedrock Bench runs, including cost versus success, timing, token usage, compatible run history, and recorded tool sequences.
---

Create the plot or diagram that answers the user's question using saved evidence.
Use the packaged CLI at `../../scripts/bench.py`, relative to this directory.
Keep outputs in the user's selected project or results directory.

For the existing library, report, or replay, use
[inspect-results](../inspect-results/SKILL.md). For a new chart, read the
[chart data guide](references/chart-data.md) for the fields and accounting rules.

- Resolve the exact run sources from the user's request, selected view, or
  `runs RESULTS_DIRECTORY`. Identify live measurements, reference checks, and
  synthetic demos explicitly.
- Use `report RUN_JSON [RUN_JSON ...] --out REPORT_DIRECTORY` to validate
  compatible completed runs and generate `comparison.json`. Keep per-run
  aggregates as separate observations for a history plot; the combined report
  aggregates repetitions and is not a time series.
- Use `replay RUN_JSON --out REPLAY_DIRECTORY [--attempt ATTEMPT_ID]` for public
  prompts and recorded tool exchanges. `REPLAY.json` includes evidence sources,
  limits, and unavailable traces. An absent trace is not a zero-tool attempt.
- For paired differences or uncertainty, use
  [compare-experiments](../compare-experiments/SKILL.md) to calculate the
  statistics before plotting them. Use
  [diagnose-failures](../diagnose-failures/SKILL.md) for failure categories and
  [audit-benchmark](../audit-benchmark/SKILL.md) for grader-control outcomes.
  Their JSON and tabular outputs retain provenance and missing-data warnings;
  preserve those in the chart. Read the
  [research chart recipes](references/research-charts.md) for these views.
- Prefer an available host visualization skill for interactive inline plots.
  Follow its rendering contract and preserve the artifact in the authorized
  output directory. Do not depend on a particular host, plugin version, or
  installed cache path. Without inline support, create a self-contained HTML/SVG
  chart or a static image and open/link it using the available file tools.
- Use Mermaid when a small static sequence or architecture diagram answers the
  question. For a detailed AWS architecture drawing, use an available AWS diagram
  skill when appropriate. Distinguish the task's intended architecture from
  resources actually verified in saved synthesis/grader evidence.
- A tool diagram shows observed commands, replies, and outcomes. Recorded order
  does not prove causation; parallel or asynchronous tools must remain distinct.
  Do not infer hidden reasoning, invent intermediate steps, or run recorded code.
- Include units, sample counts, source/run identifiers, and accounting coverage.
  Preserve unknowns. Use one common timing basis for compared marks and label it.
  A single attempt is one observation, not a latency distribution or a model ranking.
- Check the resulting view: accurate labels and denominators, working primary
  interaction, readable light/dark appearance, and narrow layout when inline.
  Use standard plotting tools for publication/export figures.

Only when this skill is invoked in the current turn, offer contextual next
steps after its chart or evidence explanation. Do not carry this behavior into
ordinary chat or plugin/repository maintenance. Use the shared
[follow-up guidance](../benchmark-agent-tasks/references/follow-ups.md).
New benchmark execution belongs to
[benchmark-agent-tasks](../benchmark-agent-tasks/SKILL.md); analyzing results
does not authorize more model calls.