Real prompts. Real tools. No mockups.

Prompt

You are recording a marketing demo. Show how Liminal ships code with verification. 1) Call plan() with exactly 3 steps: implement slugify, add tests, run tests. 2) Create marketing-capture/slugify.ts exporting slugify(input: string): string — lowercase, trim, collapse whitespace to single hyphens, strip characters outside [a-z0-9-]. 3) Create marketing-capture/slugify.test.ts using node:test (import test, assert from node:test / node:assert) with 3 cases: "Hello World" → "hello-world", leading/trailing spaces, punctuation stripped. 4) Run: node --test marketing-capture/slugify.test.js marketing-capture/slugify.test.ts (use whichever path exists after you write the file). If tests fail, fix the implementation once and re-run. End with PASS/FAIL and one example slugify("Foo Bar!") result.

Prompt

Read-only repo archaeology for a marketing capture — do not edit any files. 1) repo_map on packages/core/src with depth 2. 2) grep_file path packages/core/src/agent.ts pattern "class AgentHarness" — note the line number of the match. 3) read_file path packages/core/src/agent.ts offset=<that line minus 5> limit=120 line_numbers true — inspect how a turn dispatches tools and consumes results. 4) In exactly 4 bullets, explain the ReAct loop as implemented here: user message → model tool calls → tool results → next model turn → turn_end.

Prompt

You are recording a marketing demo. Show how Liminal ships code with verification. 1) Call plan() with exactly 3 steps: implement slugify, add tests, run tests. 2) Create marketing-capture/slugify.ts exporting slugify(input: string): string — lowercase, trim, collapse whitespace to single hyphens, strip characters outside [a-z0-9-]. 3) Create marketing-capture/slugify.test.js using node:test (import test, assert from node:test / node:assert) with 3 cases: "Hello World" → "hello-world", leading/trailing spaces, punctuation stripped. 4) Run: node --test marketing-capture/slugify.test.js (write the test as .js so node:test runs without a compile step). If tests fail, fix the implementation once and re-run. End with PASS/FAIL and one example slugify("Foo Bar!") result.

Prompt

Research task for marketing footage: 1) web_search for OpenRouter prompt caching cache_control (official docs first). 2) web_fetch the best official OpenRouter documentation URL from results (not a forum post). 3) Answer in exactly 3 bullets. Each bullet must include: a full https URL you fetched, and one exact parameter or JSON field name for caching copied from that page (e.g. cache_control). Max 90 words total.