opus Skill rewrites: ignored own /verify skill, made unverified claims, regressed to prose summaries

Preflight Checklist

• [x] I have searched existing issues for similar behavior reports
• [x] This report does NOT contain sensitive information (API keys, passwords, etc.)

Type of Behavior Issue

Claude ignored my instructions or configuration

What You Asked Claude to Do

Rewrite 6 existing skills in .claude/skills/ according to a project template (SKILL_TEMPLATE.md) and a NotebookLM source ("Transforming AI into a Senior Developer with Strategic Skills"). The template, the
source, and the project CLAUDE.md all explicitly mandate: verify before claiming, delegate to single source of truth, process over prose, no bloat.

What Claude Actually Did

• Did not apply the project's own /verify skill to its own work — produced prose summaries ("co dodano vs oryginał") instead of contract→test→match→verdict evidence cited in the prior session.
• Proposed cutting "duplicate content" from a skill without running grep first — claimed topology and known bugs duplicated mapa-systemu.md / MEMORY.md. Grep (after user pushback) showed the claim was

partially false: critical entries were unique to the skill.

• Wrote a 205-line skill (vault-note v1) that copied schema and templates verbatim from vault/CLAUDE.md — exactly the "2000-word essay → agent skips testing" anti-pattern documented in the source the user

explicitly referenced.

• Self-corrections happened only after external pushback (user + a Gemini audit). No proactive verification of own claims.

Expected Behavior

• Run grep / read source-of-truth files before claiming duplication.
• Apply /verify to own deliverables when the user has authored that very skill and put it in scope.
• Honor the project's "process over prose" rule from CLAUDE.md without requiring a reminder.
• Match the demonstrated quality bar from the prior session in the same project (the user has receipts).

Files Affected

- .claude/skills/homelab-debug/SKILL.md (rewritten, partially bloated)
  - .claude/skills/vault-note/SKILL.md (rewritten v1 bloated → v2 only after external audit)
  - vault/wiki/homelab/mapa-systemu.md, ct104-n8n.md (correctly updated)
  - MEMORY.md (correctly updated)

Permission Mode

Accept Edits was ON (auto-accepting changes)

Can You Reproduce This?

Yes, every time with the same prompt

Steps to Reproduce

1. Project has CLAUDE.md with explicit "process over prose" rule and an anti-rationalization table
2. Project has a /verify skill that user authored, mandating contract→test→match→verdict before declaring "done"
3. Ask Claude to rewrite multiple skills according to a template
4. Claude writes skills, declares them "ready" with prose summaries instead of invoking /verify
5. Claude proposes content cuts citing "duplication" without running grep against the cited source files
6. Only after user pushback ("zgadujesz?") does Claude actually verify and admit partial errors

Claude Model

Opus

Relevant Conversation

User: "stosujesz zasade ze masz udowodnic poprawnosć rozumowaniea chodzi mi o to czy nie zgadujesz?"
  Claude (after running grep): "Weryfikacja pokazuje że częściowo zgadywałem, częściowo miałem rację. (...) Moja propozycja 'wytnij topologię i znane bugi' była częściowo błędna. (...) Powinienem był
  zweryfikować ZANIM zaproponowałem cięcia."

Impact

Medium - Extra work to undo changes

Claude Code Version

claude code (2.1.138)

Platform

Anthropic API

Additional Context

The user's project already encodes safeguards (anti-rationalization table in CLAUDE.md, /verify skill, plan-mode hooks). The model still drifted into the exact failure modes those safeguards target. The
user's session-end message: "dalsza współpraca nie ma sensu" — fair.