Vertov Research

Screen Recordings, Screencasts and Video in Software Engineering

An Evidence Map of Bug Reporting, Documentation, Training and Collaboration

Author
Ankur Shrestha, Vertov, United States. ORCID 0009-0009-6036-5849
Date
14 November 2026
Type
Preprint, summarized on this page
DOI
10.5281/zenodo.24223562
License
CC BY 4.0

Abstract

Developers report bugs, teach one another, document designs and review work with recorded screens, but the research on these recordings is spread across communities that use different words for the same artifact. This paper maps that research. It is an AI-assisted evidence map. Searching open scholarly sources on 14 November 2026 returned 1466 records, or 1382 after deduplication; 198 studies were included and coded for activity, recording type, role, outcomes and strength of comparison. Tutorial and training studies (116) and requirements studies (39) are the largest groups; bug reporting accounts for 19 and code review for 1. Tool accuracy (69 studies) and satisfaction (63) are the most common outcomes; communication time was measured in 20 studies. Of the 27 studies with text, or text and screenshots, as the comparator, 19 report a direction and the results are mixed: 8 favor the recording, 8 are mixed, 1 found no difference and 2 favor the alternative. Only 6 put a recording in place of text for people, 4 of them at the edge of software engineering. The paper proposes four families of recording roles, an evidence-gap map and an open dataset that links key findings, data bases and some effect directions to quoted passages.

Keywords: screen recording; screencast; video; software engineering; evidence map; bug reporting; documentation; requirements engineering.

198
studies included, from 1,382 unique records
19
report a direction when a recording is compared with text, or with text and screenshots
8 of 19
favor the recording; 8 are mixed, 1 found no difference and 2 favor the alternative
19 and 1
studies of bug reporting and of code review, the activities closest to everyday practice

What the study found

Most studies are about tutorials, requirements and tools

Tutorial learning has by far the most studies (116 of 198), followed by documentation (46), requirements engineering (39) and onboarding and training (30). Bug reporting has 19 studies and code review has 1. A study can belong to more than one activity, so the bars add to more than 198.

Included studies by software engineering activity

Included studies by software engineering activityHorizontal bars: for each value, the number of the 198 included studies that carry it, with the count and the share beside each bar. A study with several values counts once in each, so the bars can add to more than the number of studies. The same numbers are in Table 2 of the paper.Bug reporting: 19 of 198Bug reporting19 · 9.6%Bug reproduction: 10 of 198Bug reproduction10 · 5.1%UI design feedback: 4 of 198UI design feedback4 · 2.0%Code review: 1 of 198Code review1 · 0.5%Documentation: 46 of 198Documentation46 · 23.2%Tutorial learning: 116 of 198Tutorial learning116 · 58.6%Onboarding training: 30 of 198Onboarding training30 · 15.2%Requirements: 39 of 198Requirements39 · 19.7%Remote collaboration: 13 of 198Remote collaboration13 · 6.6%Knowledge transfer: 22 of 198Knowledge transfer22 · 11.1%Program comprehension: 10 of 198Program comprehension10 · 5.1%Testing: 17 of 198Testing17 · 8.6%Other: 5 of 198Other5 · 2.5%0%50%100%

Share of the 198 included studies. A study with several values counts once in each, so bars can add to more than 198.

Source: Table 2 of the paper; counts from facts.json, sha256 6a546fd1951f. AI-assisted coding, reviewed by the author.
More on this finding

Most of the research is recent: 149 of the 198 studies were published in 2018 or later.

A recording plays one of four roles

Reading the studies by what the recording does, four roles stand out: evidence of a state, such as a defect or a test run (33 studies); instruction and demonstration, the largest by far (138); a record of knowledge kept to be found again (57); and a shared medium that people discuss or review (33). A study counts in every role it carries, so the counts overlap. The grouping is a proposal for organizing the literature, made after coding and not validated.

Four roles of a screen recording in software engineering work Four columns: evidence of a state, instruction and demonstration, record of knowledge, and shared medium. Each column gives a short definition and counts of included studies from the evidence map. Evidence of a state Shows a defect, a test run or a result so that someone else can inspect it. 33 included studies Roles coded: reporting (25), verification (11) Controlled comparison (level 4): 5 studies Consumed by a tool or agent: 16 studies Instruction and demonstration Shows how to do something, or why, to someone who will act on it later. 138 included studies Roles coded: instruction (102), demonstration (71), explanation (26) Controlled comparison (level 4): 39 studies Consumed by a tool or agent: 30 studies Record of knowledge Keeps a procedure, a design or a session so that it can be found again. 57 included studies Roles coded: documentation (57) Controlled comparison (level 4): 5 studies Consumed by a tool or agent: 14 studies Shared medium Gives two or more people a common object to discuss, review or react to. 33 included studies Roles coded: collaboration (16), elicitation (22) Controlled comparison (level 4): 8 studies Consumed by a tool or agent: 1 study
Figure 1. Roles of a recording in the included studies. A study counts in every role it carries, so the four counts overlap.
Text description of Figure 1

Four columns: evidence of a state, instruction and demonstration, record of knowledge, and shared medium. Each column gives a short definition and counts of included studies from the evidence map.

What gets measured is mostly whether the tool works

Tool accuracy is the most common outcome (69 studies), then satisfaction (63) and learning performance (45). What a person receiving a report cares about is measured far less often: communication time appears in 20 studies, reproduction accuracy in 5, clarification effort in 4 and error rate in 1.

Outcomes measured in the included studies

Outcomes measured in the included studiesHorizontal bars: for each value, the number of the 198 included studies that carry it, with the count and the share beside each bar. A study with several values counts once in each, so the bars can add to more than the number of studies. The same numbers are in Table 4 of the paper.Comprehension: 18 of 198Comprehension18 · 9.1%Task completion: 16 of 198Task completion16 · 8.1%Reproduction accuracy: 5 of 198Reproduction accuracy5 · 2.5%Communication time: 20 of 198Communication time20 · 10.1%Clarification effort: 4 of 198Clarification effort4 · 2.0%Error rate: 1 of 198Error rate1 · 0.5%Learning performance: 45 of 198Learning performance45 · 22.7%Cognitive load: 12 of 198Cognitive load12 · 6.1%Satisfaction: 63 of 198Satisfaction63 · 31.8%Tool accuracy: 69 of 198Tool accuracy69 · 34.8%Adoption use: 20 of 198Adoption use20 · 10.1%Other: 55 of 198Other55 · 27.8%0%50%100%

Share of the 198 included studies. A study with several values counts once in each, so bars can add to more than 198.

Source: Table 4 of the paper; counts from facts.json, sha256 6a546fd1951f. AI-assisted coding, reviewed by the author.

Comparisons with text are few and point in different directions

Text, or text with screenshots, is the point of comparison in 27 studies, and 19 of them state a direction. The direction is mixed: 8 favor the recording and 2 the alternative, while 8 are mixed and 1 found no difference. For 12 of the 19 the full text, or part of it, was read. For the others only the abstract was read, so the direction is what the study's own authors reported and has not been checked.

Direction of the reported result in the studies that compare a recording with text and report a direction

Direction of the reported result in the studies that compare a recording with text and report a directionHorizontal bars: for each value, the number of the 19 studies with text, or text and screenshots, as the comparator that report a direction that carry it, with the count and the share beside each bar. The counts are of studies by reported direction; they are not effect sizes and are not pooled. Table 5 of the paper lists the studies.Favors video: 8 of 19Favors video8 · 42.1%Mixed: 8 of 19Mixed8 · 42.1%No difference: 1 of 19No difference1 · 5.3%Favors alternative: 2 of 19Favors alternative2 · 10.5%0%50%100%

Share of the 19 studies with text, or text and screenshots, as the comparator that report a direction. Counts of studies by reported direction, not effect sizes; not pooled.

Source: Table 5 of the paper; counts from facts.json, sha256 6a546fd1951f. AI-assisted coding, reviewed by the author.
More on this finding

The studies are not all like-for-like contrasts. Some put a recording in place of text, some test a video added to a written specification, some observe reports with and without visual content and some measure a stated preference. Designs that add a video to a specification test the supplement, not a replacement. The paper does not turn the counts into a verdict, and states that it offers no simple rule for choosing a format.

Much of the controlled evidence comes from one research group

A single research line, around a group at Leibniz University Hannover, shares a co-author on 21 of the 198 included studies. That covers 20 of the 39 on requirements and 9 of the 12 requirements studies with a controlled design. The strongest favorable results sit in requirements engineering, so the strongest evidence for a recording is largely one group's work.

The activities closest to daily practice have the least evidence

Bug reporting, bug reproduction, knowledge transfer and code review have few controlled comparisons, while onboarding and training, which is mostly tutorials, has many. The gap map shades each activity and outcome by the strongest evidence level found. An empty cell means no study in this map, not that none exists.

Evidence-gap map: activity by outcome Evidence-gap map of software engineering activity (rows) by outcome (columns). Each cell shows the number of studies and is shaded by the strongest evidence in it: 65 cells have no study identified, 21 have only studies without a comparison, 18 have a non-controlled comparison at most, and 52 have a controlled comparison. An empty cell means no study in this map, not that none exist. Comprehension Task completion Reproduction accuracy Communication time Clarification effort Error rate Learning performance Cognitive load Satisfaction Tool accuracy Adoption use Other All studies Bug reporting 1 1 3 8 3 3 13 1 4 19 Bug reproduction 4 3 1 8 2 10 UI design feedback 2 2 1 1 1 4 Code review 1 1 1 Documentation 1 4 3 2 13 35 3 9 46 Tutorial learning 4 12 7 45 10 48 38 13 17 116 Onboarding training 3 1 24 7 18 2 1 30 Requirements 11 2 5 1 1 1 2 11 5 3 23 39 Remote collaboration 3 3 1 2 1 2 1 1 10 13 Knowledge transfer 1 2 2 6 7 7 11 22 Program comprehension 2 2 2 1 2 6 2 10 Testing 1 3 2 3 1 3 12 1 3 17 Other 1 3 2 5 Studies 18 16 5 20 4 1 45 12 63 69 20 55 No study identified in this map (not evidence that none exist) Studies without a comparison (evidence levels 1 and 2) At least one non-controlled comparison (level 3) At least one controlled comparison (level 4). The number in a cell is the count of studies; * = proposed value
Figure 2. Evidence-gap map: software engineering activity by outcome, shaded by the strongest evidence level in each cell.
Text description of Figure 2

Evidence-gap map of software engineering activity (rows) by outcome (columns). Each cell shows the number of studies and is shaded by the strongest evidence in it: 65 cells have no study identified, 21 have only studies without a comparison, 18 have a non-controlled comparison at most, and 52 have a controlled comparison. An empty cell means no study in this map, not that none exist.

Companion study: How GUI Agents Receive Instructions: An Exploratory Mapping of 85 Benchmark Families.

Limits to keep in mind

  • Literature coverage. The search drew on openly accessible sources. Four large databases (Web of Science, Scopus, IEEE Xplore and the ACM Digital Library) were left out, so some studies, especially paywalled and very recent ones, will be missing.
  • Evidence quality. Because 97 of the 198 studies were read only through their abstracts, some codes may be incomplete. The map describes how the literature is distributed. It does not estimate whether recordings work.
  • Classification and scope. A few publications exist in overlapping versions or sit near the eligibility line, and each judgment call can move a count. The protocol was not registered before coding.
  • Interpretation. These are patterns across the studies that were found. They say nothing about cause and effect, and the format comparisons tally studies instead of combining their results.
  • A preprint. This page summarizes a preprint. The PDF and the Zenodo record carry the full paper, with every table, the references and the appendices.

Publication details

The full paper, with every table, the reference list and the appendices, is the PDF above and the Zenodo record. This page is a summary of it and does not replace it. The study is a preprint.

Evidence map of 198 studies on screen recordings, screencasts and video in software engineering. 198 included studies coded for the activity, the type of recording, its role, the outcomes measured and the strength of the comparison, with a quoted passage behind key findings. AI-assisted coding, reviewed by the author and quote-checked.

The merged dataset, the evidence register (verbatim quotes and locations for key findings, data bases, methods and some effect directions), the codebook, the search strategy, the search log, the screening decisions, the correction log and the analysis scripts are released on Zenodo (dataset DOI: 10.5281/zenodo.24222760; license: CC BY 4.0). The full texts of the included studies are third-party works and are not redistributed; document hashes and URLs allow verification.

  • 14 November 2026. Version 1, released as a preprint.
Source files and sha256 behind the numbers on this page
  • 13-manuscript/manuscript-keys.md: 081c7e52bb9f32aeaf397516c7810cc21ef712c7f6ad276e9aee10c52913733c
  • 06-data/results/facts.json: 6a546fd1951fe06941f71eba80d79418557728651c5e4b4baec513005f3998b8
  • 13-manuscript/tables.json: 0acc02042a72c516333067ed3cb049b85563a24bf378c25a673c365a89a04f09
  • 13-manuscript/figures/captions.json: 3540ce14e504ba7f5c3088a98b03da229f962f53c00839ddc3249762f839ad1a
  • 13-manuscript/figures/taxonomy.json: 6bd96a5902452247025ff1a63dfd0778f129e19c9cd6e065900b3e03fe9b72ec
  • 13-manuscript/cited-keys.json: add3f40b4a32b79b4880027dab2f44bbb3fcabfbff81f6b3c6d620df3d911fe2
  • 13-manuscript/references.bib: e99dafc897d4af17048dfdd08bbcdefa68d4968ea125cfba4b54e0231c0412af
  • 13-manuscript/included-studies.bib: 3fd043e517dbd2e69b227ebc9a75305481969db3ae524dfaf89a626d3045c901
  • 13-manuscript/extra-references.bib: 948d8793f3d1ae54d8ff8836e4c6489761c3f10da69933c6a84b391446d3c393
  • 13-manuscript/params.tex: 743fef3de971233bd0833d60d33fcad3f53888c3ce089d1c36066dfcd6a04135
  • 13-manuscript/latex/template.tex: 0d11bc0a2e421b3426f15f919c402432d9c1c9f1d4a1969e109570f8f30ce01f
  • 15-ssrn-package/disclosures.txt: 6fa274782ec2d9f2b59f799d9b9543708e3cf7d0ebe9addd23e82a4eb1da3f9d
  • 13-manuscript/figures/fig1-flow.svg: 3ab08f5a121afdf2e6118e74737d973625c516b997966545438dae4c2068e0c2
  • 13-manuscript/figures/fig2-years.svg: 52341a6073053d7b0d827d78d141eb12092248c2349c899c51f8ed1b40f0c897
  • 13-manuscript/figures/fig3-activity-artifact.svg: c4f8f7d4ff06cf609fc008496d04f8f898eb7434d34913067f5f8e3e1f7d6629
  • 13-manuscript/figures/fig7-taxonomy.svg: c7d9d5f461e06a8092526dcad9756902fd27841cbd10653833f38ba6a6493415
  • 13-manuscript/figures/fig4-gap-outcome.svg: ab3bdb83a4d1472bf8bc3003529b97d72479803683311d27ff85f8bcb98fe49f

Disclosures

As stated in the paper.

  • Competing interests. The author is affiliated with Vertov (https://vertov.app), a company developing screen-recording and AI-agent collaboration software. Vertov has a commercial interest in the subject of this research.
  • Use of AI. Claude (Anthropic) was used to assist with literature research, data extraction and classification, analysis, verification checks and manuscript preparation. The methods and limitations of AI-assisted coding are described in Sections 3 and 6.
  • Funding. This research was funded by Vertov. No external funding was received.
  • Ethics. The study analyzed published research papers. No human participants and no personal data were involved.

Cite as

Shrestha, A. (2026). Screen Recordings, Screencasts and Video in Software Engineering: An Evidence Map of Bug Reporting, Documentation, Training and Collaboration. Preprint. Vertov Research. https://doi.org/10.5281/zenodo.24223562. https://vertov.app/research/screen-recordings-in-software-engineering

@misc{shrestha_screen_recordings_evidence_map,
  author = {Shrestha, Ankur},
  title = {Screen Recordings, Screencasts and Video in Software Engineering: An Evidence Map of Bug Reporting, Documentation, Training and Collaboration},
  year = {2026},
  note = {Preprint},
  publisher = {Vertov Research},
  doi = {10.5281/zenodo.24223562},
  url = {https://vertov.app/research/screen-recordings-in-software-engineering}
}

RIS

TY  - GEN
AU  - Shrestha, Ankur
TI  - Screen Recordings, Screencasts and Video in Software Engineering: An Evidence Map of Bug Reporting, Documentation, Training and Collaboration
PY  - 2026
PB  - Vertov Research
N1  - Preprint
DO  - 10.5281/zenodo.24223562
UR  - https://vertov.app/research/screen-recordings-in-software-engineering
ER  - 
Commands and recent recordings
Actions
Go to
↑ ↓ Select↵ OpenEsc Close