Knowledge, prompts & checklists

Three optional inputs steer the design pass: a checklist (what to test), domain-knowledge files (app-specific facts), and prompt overrides (replace any built-in prompt).

Checklists

A checklist is a plain-Markdown file passed to design / explore via --checklist <file> that steers what the bot tests. Each intent becomes a coverage target — the run is scored on how much of the checklist it actually exercised (semantic checklist coverage in the Pilot verdict).

Format — one test intent per - bullet; # lines are notes / section dividers and are ignored:

# Header & navigation
- The top navigation shows the Platform, Docs, Blog and Cairn links
- Clicking the "Plune" logo returns to the home page

# Hero
- The hero shows the tagline "An AI that walks your system and leaves a trail of tests"

Caution

Use single # lines for section notes — never ## headings. A ## line flips the parser into “headings-as-items” mode and your - bullets are silently dropped.

Ready-to-run examples ship in the repo: examples/plan.md — a full checklist for this very page (plune.ai/cairn) — plus examples/checklist-history.md.

Domain knowledge — ./knowledge/*.md

Put Markdown files in ./knowledge/ with a url: front-matter. When a run’s URL matches, the file content is injected into the design pass — credentials hints, validation rules, business context:

---
url: https://app.example.com/checkout
---

- Discount codes are case-insensitive; `SAVE10` is a valid test code.
- The "Pay" button stays disabled until both address fields validate.

Prompt overrides — ./prompts/<name>.md

Drop a file named after any built-in prompt to override it without rebuilding the package:

prompts/
  design.md        # replaces the built-in design prompt
  judge.md         # replaces the LLM-judge prompt

Overrides work fully offline. With Langfuse connected, prompts can also be managed there (see ADR 0004).