@dzhechkov/harness-cli 0.3.201 → 0.3.202

package/README.md

@@ -531,11 +531,21 @@ pre-emptive probe, not reactive detection, is the real defense.
 
 ### Claim check (`dz claim-check`)
 
-Makes the Integrity Rule — *"no fake claims; verify before claiming success"* — a runnable check
-instead of prose. Every quantitative claim (coverage, test/skill/command counts, benchmarks) must be
+**Why you want this.** Docs accumulate numbers nobody can reproduce — things like `99% accuracy`,
+`10× faster`, `100% coverage` — and an AI agent writing your README is especially good at inventing
+them. A rule like *"no fake claims; verify before claiming success"* enforced only by discipline
+eventually loses. This makes it a runnable check with an exit code.
+
+**What it does.** Every quantitative claim (coverage, test/skill/command counts, benchmarks) must be
 tagged `MEASURED` / `CLAIMED` / `SYNTHETIC` / `ESTIMATED` / `UNVALIDATED` / `BASELINE`, and a
 `MEASURED` claim must name its reproducer (`npm test`, a coverage report, `npm view`, a git ref).
-"100% / perfect" framing is always high severity.
+A `100%` or `perfect` framing is always high severity — it is the claim most likely to be untrue.
+(The backticks are load-bearing: this paragraph quotes the claims it forbids, and a backticked
+literal reads as code, not as an assertion.)
+
+So `coverage 92%` is flagged; `coverage 92% (MEASURED — npm test)` is not. `2136 tests` is flagged;
+a shields.io badge URL is not (the URL is machinery, not prose). A number inside backticks still
+counts: ``accuracy reached `0.95` `` is a claim.
 
 ```bash
 dz claim-check                       # scans root README, every package README, features/*/08_qe_report.md
@@ -557,6 +567,27 @@ dz publish --filter harness-cli --claim-check error  # a high finding fails THIS
 dz publish --filter harness-cli --claim-check off    # disable the gate entirely
 ```
 
+#### From Claude Code — just ask in plain language
+
+There is no slash command: `dz` is a CLI, and the agent runs it for you.
+
+> "Проверь наши README на непроверяемые утверждения перед релизом"
+>
+> "Run dz claim-check and fix the high findings by tagging them MEASURED with a reproducer"
+>
+> "Before we publish, make sure no doc claims 100% of anything"
+
+Claude runs `dz claim-check --json`, reads the findings, and edits the offending lines. Two habits
+worth asking for explicitly: **do not silence a finding by deleting the number** (a claim you cannot
+reproduce should be removed or downgraded, not hidden), and **do not weaken the rules** to make the
+gate green.
+
+#### In CI
+
+```bash
+dz claim-check --fail-on high || exit 1     # block a release on a "100%" claim
+```
+
 Ported from rUv's [`ruview`](https://www.npmjs.com/package/@ruvnet/ruview) `src/guardrails.js` (MIT),
 which was itself written after a documented AI-slop incident. The detection semantics are kept
 verbatim — including a deliberately `\b`-free `100%` regex that looks like a bug and is not.

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@dzhechkov/harness-cli",
-  "version": "0.3.201",
+  "version": "0.3.202",
   "description": "The dz CLI — install AI skills for Claude Code, Codex, OpenCode, Hermes, OpenClaude, GitHub Copilot. 35 commands, 13 presets, 6 platform targets.",
   "type": "module",
   "license": "MIT",