Run a read-only design study with an evidence legend

Produce a citation-bearing study of how something works, marking every claim [V]erified or [P]rovisional, before anyone touches it.

A design study is worth reading only if its confidence is legible. The trap is a study that mixes what was checked with what was assumed and marks neither. A two-symbol legend fixes it: [V] for a claim carrying a file:line or URL, [P] for a claim that is provisional — inferred, plausible, and explicitly flagged as unverified.

Start here

One prompt to start

Read-only -- I want a citation-bearing study of how this actually works before anyone touches it: let the next file to read be decided by what you just found, mark every claim [V]erified with a file:line or [P]rovisional if you're inferring it, and leave the reasoning traceable enough that I can audit it myself.

Needscompass

Beats

Run these in order. Each prompt is copy-pasteable straight into Claude Code.

1. compass:compass-investigate-dynamicallyalso in 1 other recipe

   • Root-cause a bugfix from a tracked issue

   The next thing worth reading is decided by the last observation; a pre-planned file list misses it.

   Run this beat on its own

   read-only please: I want to understand how this works before anyone changes it. Don't edit a single file.

2. compass:compass-ground-evidencealso in 2 other recipes

   • Verify a brief's premises before acting
   • Do oracle engineering and numerical V&V

   The [V]/[P] split is exactly this skill's file:line-or-flagged-prior-knowledge rule.

   Run this beat on its own

   write this up with a legend: [V] for anything you can cite a file:line for, [P] for anything you're inferring. Nothing unmarked.

3. compass:compass-map-relationshipsalso in 4 other recipes

   • Check cross-forge CI parity
   • Collapse duplication hand-synced across N sites
   • Migrate a return shape or type representation
   • Propagate a vendored artifact to N copies

   Multi-hop claims need a traversable triple index, not recollection.

   Run this beat on its own

   trace how a change in the shared script would reach each plugin's output

4. compass:compass-summarize-trace

   A study whose reasoning is not reconstructible cannot be audited by its reader

Worked example

Grounded in — why this beat order is trustworthy

a read-only study of how the four werkstoff `PreToolUse` hooks interact, covering `plugins/andon/hooks/`, `plugins/self-assess/hooks/`, `plugins/confab/hooks/`, and `plugins/cupertino/hooks/`, with each interaction marked [V] or [P].

Do / Don't

• ✓Let each observation decide what to read next -- a plan written before the first file is opened will miss it.
• ✓Mark every claim [V] with a real file:line or URL, or [P] if you're inferring it -- never leave a claim unmarked.
• ✓Build a traversable index for multi-hop claims rather than relying on recollection.
• ✓Make the reasoning trace itself reconstructible, not just the conclusion -- an unreconstructible study can't be audited.

• ×Don't follow a pre-planned file list once the investigation is underway -- the next thing worth reading is decided by the last observation.
• ×Don't leave a claim unmarked -- mixing what was checked with what was assumed, unmarked, is exactly the trap this recipe exists to avoid.
• ×Don't edit anything -- this is explicitly read-only, before anyone changes it.