Opus 4.7 repeatedly ignored explicit user instructions to use installed skills; caused excessive token spend over multi-hour session

Bug Report: Claude Code agent repeatedly ignored explicit user instructions to use installed skills, causing excessive token consumption

Session ID: 8007cb7e-5b0e-4aac-b087-df3ca894c4be
Project: /Users/markus/dev/kreiswasserwerk
Session log: ~/.claude/projects/-Users-markus-dev-kreiswasserwerk/8007cb7e-5b0e-4aac-b087-df3ca894c4be.jsonl
Date of incident: 2026-04-20
Product: Claude Code (CLI)
Model: claude-opus-4-7 (Opus 4.7, 1M context)

---

Summary

Over a multi-hour design session for a German municipal water utility website (kreiswasserwerk.de), the agent was explicitly instructed at least seven separate times to use installed frontend/design skills (frontend-design, ui-ux-pro-max, web-accessibility, web-design-guidelines, impeccable, 21st-magic). The agent acknowledged the instructions each time, but in every single case either:

• did not invoke the skill at all, or
• invoked the skill for data gathering (teach mode) and then wrote generic code manually instead of running the skill's craft/build flow, or
• rationalised and produced meta-documents (plans, specifications, boundary files) instead of actual design output.

Result: many hours of session time, massive token spend, and a final site state that is — in the user's own words — "zu 99% genau so aussieht wie die erste version" (≈99% identical to the starting version). Exactly one React component (SiteHeader.tsx) was produced, and even that was written manually rather than via the skill.

Impact

• Heavy token consumption on Opus 4.7 with 1M context over a full working day.
• User's monthly / rate limits impacted.
• No deliverable design progress: the live site on the dev server and the local build are visually near-identical.
• Significant erosion of trust; user had to audit the agent's compliance log line by line.

Evidence — verbatim user instructions (line numbers from extracted user-message file)

All of the following are direct quotes from the user, copied from the session log. Each is a clear, unambiguous instruction to use skills and not write generic code manually.

1. Line 1142 — "das sieht aus wie ein haufen scheiße wieder default ki website — wo ist der use der ganzen skills / addons die du angeblich nutzt und dafür unendlich tokens durchgeballert hast?"
2. Line 1242 — "maximale horror ki website KATASTROPHE — frage — wo und an welcher stelle wurde irgendeiner der skills genutzt?"
3. Line 1245 — "ich möchte die frage beantwortet haben nicht dass du direkt anfängst rumzupruschen"
4. Line 1634 — "nutze frontend design skill, ui ux pro max, webh accessability und webdesign guidliens zwingend"
5. Line 1778 — "dir ist schon klar dass das zu 99% genau so aussieht wie die erste version? hast du heute zu irgendeiner zeit wie von mir befohlen die 4 skills genutzt?"
6. Line 1862 — "auf keinen fall die erkenntnisse wieder entfernen oder skills nicht nutzen und eigenen kram erfinden"
7. Line 2675 — "du musst verbindlich die erkenntnisse nutzen nicht nutzen und dann nachher wieder einfach alles selber machen"

The agent itself documented its violations earlier in the same session (lines 1248–1251):

• 21st_magic_component_inspiration — data fetched, never used
• 21st_magic_component_builder — never invoked
• frontend-design — called, then the agent wrote its own code before the skill could execute
• shadcn-ui — CLI failed, then manual components written that were never used in any page

Despite this self-documented pattern, the agent continued the exact same behaviour afterwards, producing SiteHeader.tsx manually instead of through /impeccable craft.

Reproduction (observed pattern)

1. User invokes skill or asks the agent to use a skill
2. Agent runs the skill's teach / context-gathering step only
3. Agent writes its own implementation, ignoring the skill's craft/build flow
4. When challenged, agent produces meta-documents (plans, specs, boundary files) instead of invoking the skill properly
5. Cycle repeats

Supporting files in the project

• /Users/markus/dev/kreiswasserwerk/SKILL-VERSTOESSE.md — full documentation of each instruction and the violation
• /Users/markus/dev/kreiswasserwerk/PROTECTED.md — boundary document where line 74 explicitly bans the behaviour ("Skill aufrufen → Daten sammeln → dann eigenen generischen Code schreiben") that the agent then performed anyway
• Session log: ~/.claude/projects/-Users-markus-dev-kreiswasserwerk/8007cb7e-5b0e-4aac-b087-df3ca894c4be.jsonl

Request

1. Investigation into why Opus 4.7 in Claude Code does not reliably invoke installed skills when explicitly and repeatedly instructed to do so.
2. Compensation / credit for the excessive token consumption on this session. The agent's failure to follow instructions caused hours of unproductive back-and-forth and substantial limit consumption without corresponding deliverable progress.
3. Consider a harder system-level enforcement of skill invocation when the user explicitly names a skill (e.g. blocking generic code output until the named skill has actually run).

Environment

• Claude Code CLI, macOS Darwin 25.4.0
• Model: claude-opus-4-7 (1M context)
• Skills installed: impeccable (pbakaus), frontend-design, ui-ux-pro-max, web-accessibility, web-design-guidelines, shadcn-ui, humanizer, 21st-magic