← ClaudeAtlas

interaction-testslisted

Organize a project's agent-facing documentation (CLAUDE.md, docs/, rules, memory, docstrings) as a verdict-bearing INDEX over pulled findings files, then VERIFY it works by running blind interaction tests — probes measuring whether a fresh agent actually recalls what the docs know. Use when agent-facing docs grew past a screen, when an agent rebuilt or re-litigated something already decided, when CLAUDE.md is large or contradicts other docs, when documentation drift is suspected, or when asked to audit, restructure, shrink, or verify documentation an agent reads.
rupel190/claude-interaction-tests · ★ 0 · Testing & QA · score 67
Install: claude install-skill rupel190/claude-interaction-tests
# Interaction Tests **An interaction test does not test code. It tests how an agent interacts with the context structure you gave it** — whether a fresh agent, mid-task, carrying its own vocabulary, recalls the thing that stops it doing the wrong thing. ⛔ **A read-through cannot substitute.** A read-through is what produced the docs. Someone who edited a file line-by-line the same day still misses that it overstates its own source; a probe catches it in one shot. This skill is itself built the way it tells you to build: this file is the index, `reference/` holds the detail. **Pull the reference file when its subject is in play.** ## The five moves ``` 1 MAP where does each KIND of fact live — → reference/organize.md INCLUDING the kinds decided outside the repo 2 ORGANIZE index (loaded) over findings (pulled) → reference/organize.md 3 PROTECT rules + guards so it cannot decay → reference/drift-protection.md the eight guard patterns a machine CAN check, and where each RUNS → reference/guards.md 4 TEST blind probes with controls → reference/probes.md 5 READ diagnose by failure mode, then fix → reference/taxonomy.md ``` `reference/planning-rows.md` covers making decisions **citable**, which is what lets a probe say *"row C4 closed this"* and lets you find contradictions precisely. ## The four laws Everything else follows from these. **1. The index must be useful when the pull does NOT happen