AGENT SKILL

evidence-capture

Record an affected UI flow or package existing screenshots and video as delivery evidence.

License 许可
MIT
最近更新
2026-11-14
Stars
★ 0
Forks
0

01 前言 Front Matter

name
evidence-capture
description
Record an affected UI flow or package existing screenshots and video as delivery evidence.

02 资源文件 1

03 SKILL.md 正文

The orchestrator retains product acceptance. Live Computer Use is assigned only to a role with actual browser capability; deterministic host tools capture and package media under that role's authorized plan. Use an assigned bounded script or media role only when a writer is needed; a running recording needs no model agent. Keep one driver per browser session. Reuse the active execution method when a browser-capable role is needed. For existing media, verify the candidate and review receipt in step 1, then continue at step 5; recording steps apply only when making new captures.

  1. For final evidence of a visual implementation, require completed Impeccable work and independent visual-capable critique with findings resolved on this candidate (design-quality). Packaging existing media reuses its review receipt; it does not start a redesign or a new implementation review. If the receipt is missing, label visual acceptance unverified. A behavior-only recording without a visual change does not require a design exercise. Baselines and review screenshots are diagnostic and may precede that gate. Read the project's existing verification skill. Bind the actual repository, app, revision, authorized origin/path, test state and viewport. Obtain a baseline before editing; label any recovered old revision as a recapture.
  2. Select a representative walkthrough for the affected behavior, plus distinct risk states. Honor an explicit request for exhaustive coverage. Reuse valid captures while candidate, data and environment match; changing an unrelated file does not itself require rerecording all flows. Choose the viewport and populated/empty state that demonstrate the requested task and match its reference. A portrait browser capture does not demonstrate the desktop composition. Confirm authentication, organization and usable sample data before recording; report a missing state instead of substituting an unrelated flow. Preserve private data and use authorized fixtures only.
  3. Discover the active host's recording and interaction tools once. Keep its authorization and browser mechanism. Read host-execution.md before composing a batch. Use the existing host-authorized execution plan; do not introduce a second plan schema or browser bridge.
  4. Start recording, perform the authorized flow, wait for the final visible outcome, then stop; also stop in a finally path on failure. If the host only supports fixed-duration clips, ensure the completed flow fits inside the clip or use shorter meaningful segments. A later action cannot complete an earlier recording. Trim idle time only when it preserves the demonstrated sequence. Group stable dependent operations into code; return compact step results and evidence references. Yield when navigation, state or permission changes need judgment. Refresh stale element references; never replay a possibly committed write to recover from a tool error.
  5. Check that the media decodes, has nonzero duration and shows actual actions. Sample beginning, transition and the actual end of the file; verify that the terminal outcome is present within the media, not just in a later screenshot or tool log. Inspect more when anomalies appear. Record exactly what was sampled. Capture existence, full playback, functional success and visual quality are separate claims.

Return completed steps, tool failures, media references, candidate/viewport/state and gaps. Keep raw frames, recordings and credentials out of model text output. The executor reports complete/incomplete, never product PASS/FAIL. The orchestrator combines it with before/after probes and independent visual review. Missing recording support remains a gap; screenshots or generated video cannot impersonate a walkthrough.

See host-execution.md for limits and invocation.

04 同仓库更多 Skills