Agentic editing of terminal screencasts

asciinema is naturally suited to agentic screencast editing. A .cast recording is plain text (JSON Lines), one event per line of the form [interval, code, data], where interval is seconds since the previous event. Editing reduces to arithmetic on those intervals (and optionally to substitution on the payloads, e.g. for redaction), so a small tool can expose trimming, speeding, and cutting as cheap operations that a language model can reason about and combine.

As a demonstration, I recorded an ~85-minute Claude Code session running an ML fine-tuning task with the ml-research plugin and turned it into a 40-second GIF of the highlights without leaving Claude Code. The edit was driven by short natural-language instructions and one custom skill (cast-edit) that wraps the format with a small Python tool.

The session

The recording started with asciinema rec before issuing /ml-research:ml-research-task with a prompt to fine-tune Qwen/Qwen2-0.5B on trl-lib/Capybara via HF Jobs. Claude ran for about 85 minutes, dispatched a researcher subagent, ran a local smoke test, submitted two HF Jobs (a 5-step smoke and a full epoch), and produced a recap table with the final recipe. The raw .cast is 5087s, 12,577 events, ~88 KB.

The edit

The conversation, paraphrased:

1. "give me an overview, remove the trailing /quit, then remove the notable dead-air spans". Claude ran castedit.py analyze for a few-hundred-token summary of idle gaps and typing runs, located the / keypress at t=5083.15s and applied a trailing cut, then applied idle_cap=2.0 to clamp every gap longer than two seconds. The 5087s recording dropped to 1336s.

2. "create a 40-second cast that makes user input, intermediate responses, and the final response easy to follow, and speed up everything else. First give me an overview of what is relevant and what to speed up". Claude re-analyzed the 1336s cast, identified the readable segments (the typed prompt at t≈12s, the first tool calls at t≈16–50s, an auto-recap at t≈1115s, and the final summary table at t≈1199–1304s), and proposed an eight-region plan with explicit speedup factors per segment:

   Window (s)	Content	Factor
   0–11.5	terminal idle before prompt	10x
   11.5–16.5	user types the /ml-research:ml-research-task prompt	1x
   16.5–50	first reply, Bash(uv --version && hf auth whoami), "Preflight passes"	5x
   50–1115	subagent work, spinners, polling	250x
   1115–1120	first ※ recap block	1x
   1120–1199	waiting for the full HF Jobs run	50x
   1199–1304	final "Done" response, stages table, recipe	7x
   1304–1336	trailing pings	30x

   First render landed at 32s, eight seconds short of target. Claude reallocated the headroom to the two readable stretches with the most text on screen, dropping the early-intermediate factor from 5x to 3.5x and the final-response factor from 7x to 5x. Second render landed at 40.4s. 3. "render to gif". agg with --renderer resvg and a font fallback that pins STIX Two Math for the U+23F5 auto-mode arrows, which agg's default swash renderer otherwise displays as tofu.

The result

The 8.2 MB GIF gives the typed prompt, the first tool calls, the auto-recap, and the final summary table a few seconds of legible screen time each. The 19-minute waiting stretch compresses to under five seconds.

What made it work

Two design choices in the cast-edit skill kept this practical.

The cast never enters the model's context. A real recording is megabytes of ANSI output. castedit.py analyze parses the file and returns a compact summary: idle gaps over threshold, detected typing runs with reconstructed text snippets, and an estimate of how much idle-capping would reclaim. The model reasons over a few hundred tokens, not the raw stream.

A small, explicit JSON plan covers the editing surface. The edit command takes three knobs: idle_cap clamps every gap to a maximum, speed_regions divides intervals in a time range by a factor, and cuts drops events in a range. Each knob is a direct manipulation of intervals, so the output stays lossless to rendering and no absolute timestamps need recomputing. The model emits the plan as JSON and the engine applies it.

The accompanying SKILL.md is operator notes and guardrails: the rounding behavior that keeps total drift sub-millisecond, why cuts cannot shrink pure idle gaps, and the agg flags that avoid Claude Code's known glyph artifacts.

The full editing loop, from "give me an overview" to validated GIF, was a handful of natural-language turns. Most of the guidance is in the skill description telling the model what knobs exist and how to combine them. The model supplies the judgment about which seconds matter to a viewer.