reel-driven-development

Reel Driven Development (RDD) - turn recorded user walks (reels) into domain stories and outcome objects

PyPi
GitHub
Code
Docs

Demo

See a reel for yourself: the GenWiki demo reel (7 hops) on rdd.bitplan.com - true inspection mode, your verdicts stay on your device. Reviewers get a link by mail - no account, no password.

Installation

pip install reel-driven-development

or isolated with pipx:

pipx install reel-driven-development

Both put the rdd command and the tool entry points on the PATH.

Documentation

Wiki

Authors

• Wolfgang Fahl

Reel Driven Development

A reel is a recorded user walk - e.g. a screen-share demo where a domain expert walks through a system while talking. RDD treats the reel as a graph walk:

• every context switch (page, browser tab, application) and every relevant interaction (submenu, filter, zoom, sort) is a hop
• the narrative drives the sampling, never the clock
• every hop gets an evidence frame (screenshot) or a proven-absent note
• findings become outcome objects: a frustration spoken during a walk becomes a bug report, which is an acceptance criterion, which is an example

HopDetection was the first step - issue #1; the requirements epic is issue #7.

rdd - the one command name

rdd dispatches to the tools of the pipeline, so a user has to know one name - the name of what we do:

rdd without a subcommand lists them; the tool names stay available as entry points of their own.

rdd detect flags

rdd detect --help lists these together with the standard options of pybasemkit.

Detectors

Hop candidates come from the detectors of PySceneDetect, which benchmarks them.

Each detector is on offer at its library default and at half and double its threshold. ThresholdDetector is not on offer: it detects fades to a near-black level.

Cuts found on examples/recordings/genwiki-walk - 1501 frames, 60 s, about 1.4 s per run:

Output

A run writes to --out:

• hop-<hh>h<mm>m<ss>s[<ms>ms].jpg - the evidence frame of each hop, the full frame as recorded, named by its time offset so curation never renames a surviving frame (issue #21)
• hops.yaml - the hop records, whose field names are the property names of Concept:HopContent; node, url and summary stay empty because they come from the transcript and are never guessed from the picture
• config.yaml - the values that decided this hop set, so a run can be repeated from its own output (issue #4)

Example: GenWiki walk

rdd detect examples/recordings/genwiki-walk/genwiki-walk.mp4 --detector Content --out hops
genwiki-walk.mp4: 9 hops from Content over 1501 frames -> hops

Raw results of an earlier run in examples/recordings/genwiki-walk.

Known gaps

• content outside the walk - participant tiles, a clock, a playing video - can still decide hop boundaries (issue #5)
• there is no machine-readable progress stream for an agent driving a run (issue #36)
• transcript-anchored capture of a named moment is not implemented