Each recording below is a screen capture of the actual Liminal web UI while the harness ran a real turn — same API, same approval gates, same session logs as daily use. Prompts are shown verbatim; tool names match what the agent invoked.
How these were made:npm run marketing:capture:live at the monorepo root (requires your API key and a running web stack). Session JSONL is archived under assets/marketing/recordings/.
Live capture 01
TypeScript debounce + typecheck
~1.6 min · real harness run
Prompt
Write a well-typed TypeScript `debounce` function in `marketing-capture/debounce.ts` (create the folder), then run `npx tsc --noEmit` on that file only and explain in two sentences how it works.
Tools invoked
list_dir
activate_tool_family
read_file
write_file
run_shell
Liminal web UI · TypeScript debounce + typecheck
Live capture 02
Repo exploration
~5 min · real harness run
Prompt
Use grep_file to find where `AgentHarness` is defined under `packages/core`, read the main file, and summarize the ReAct loop in 3 bullet points. Do not edit any files.
Tools invoked
grep_file
list_dir
read_file
Liminal web UI · Repo exploration
Live capture 03
Web research
~2.2 min · real harness run
Prompt
Use web_search then web_fetch on OpenRouter's prompt caching docs. Reply in 3 bullets with URLs — keep it short.
Tools invoked
web_search
web_fetch
read_file
Liminal web UI · Web research
Live capture 04
Git status summary
~0.8 min · real harness run
Prompt
Run git_status and git_diff with stat summary. In 2–3 sentences, what is modified in this repo? Do not commit.
Tools invoked
activate_tool_family
git_status
git_diff
Liminal web UI · Git status summary
Install a local AI coding agent today
Install Liminal AI in one command. Free software, your API key, your machine — edit code, run tests, and ship faster without a SaaS middleman.