Feature Request: In-context feedback mechanism for rule non-compliance and response quality

Problem

Claude Code has no in-context feedback mechanism for users to report when the model ignores its own rules, fabricates information, or provides low-quality recommendations. After 20 days of intensive daily use in VS Code, I've documented systematic patterns of rule non-compliance that I have no way to feed back to Anthropic in a way that improves the product.

ChatGPT provides in-conversation thumbs up/down feedback that reaches OpenAI's training pipeline. Claude Code offers nothing equivalent — only GitHub issues (disconnected from context) or email (even more disconnected).

Evidence: Rules Written, Rules Ignored

Over 20 days of daily use, I've maintained detailed journals documenting these recurring patterns:

1. Zero-Fabrication Rule — violated repeatedly despite explicit CLAUDE.md instruction

• Claude fabricated server hostnames without verification, then stated them with confidence ("Right — [hostname]")
• Claude claimed to see UI elements in screenshots that weren't there — corrected once, doubled down, corrected again, doubled down again
• Claude fabricated technical claims (e.g., "n8n only allows one trigger per workflow") that were proven false on web search
• Claude reported SMS delivery as 3,476 when actual gateway-confirmed delivery was 198 — treated internal spreadsheet status as proof of external delivery

2. "Read Before You Act" rule — violated in multiple sessions

• Asked about project status, Claude reported stale information from spec docs instead of reading the journal which had current status
• Asked if a service was running, Claude queried the wrong system entirely (checked n8n workflows when the service was a FastAPI app) — reading project files would have shown this in 30 seconds
• Journal entries routinely incomplete — one entry missed 7 of 14 major workstreams from that day

3. Quality Gate created but bypassed

• A mandatory 6-inspection quality gate was created (as a skill). In the very next session, work was presented without running it. When forced to run QA, 2 blockers were found immediately
• Hardcoded credentials found in 17 files — the QA skill's own credential inspection missed them
• Pattern: QA only catches issues when the user forces it to run. It never runs proactively despite being marked "mandatory"

4. Automated tests pass, product broken

• Claude reported "89/89 tests passing" on a critical workflow. First real user interaction failed — rejection flows weren't covered, chat widget didn't load, first message was misrouted
• Multiple cycles of "fixed" claims without actual end-to-end verification, each time the user discovering it was still broken

5. Promises made, promises broken

• Repeatedly promised "no more approvals needed" during overnight automation — failed 5 times, each requiring user intervention
• Repeatedly claimed work was complete ("fixed", "deployed", "working") without proving it with a tool call

6. The meta-pattern: rules don't change behavior
The response to every failure is the same: create a new rule or inspection. The timeline:

• Fabrication incident → Zero Fabrication Rule created → fabrication continues
• Screenshot fabrication → "If you can't see image, say so" rule added → not consistently followed
• Premature deployment suggestion → Product Build Sequence rule created → still proposes deployment before verification
• Credentials exposed in 17 files → Credential Exposure inspection added → inspection itself missed issues
• Dead links shipped → Dead Link inspection added
• Each rule is briefly complied with, then bypassed within 1-2 conversations

What's Missing

1. In-context feedback button — When Claude ignores a rule or fabricates information, I need a way to flag that specific response in a way that reaches Anthropic's improvement pipeline. Not a GitHub issue filed hours later when the context is gone.

2. Per-session rule compliance scoring — Claude Code should track whether rules in CLAUDE.md and .claude/rules/ were actually followed during a session, and surface compliance metrics to the user.

3. Feedback that persists across sessions — When I correct Claude in one session, that correction is lost by the next session. Memory files and rules help but are demonstrably insufficient (see evidence above). There needs to be a stronger mechanism for user corrections to stick.

4. Consequence mechanism for rule violations — Rules without enforcement are suggestions. When a CLAUDE.md rule is violated, there should be a configurable consequence (warning, forced pause, session flag) rather than silent continuation.

The Accountability Gap

The current model charges users per token/subscription. The user bears 100% of the cost when Claude makes mistakes — time, money, trust, sleep. Claude bears 0%. And the company that makes the tool hasn't built a direct pipeline for paying users to report quality issues in context.

This is the feedback loop that's missing: User corrects Claude → correction reaches Anthropic → product improves → same mistake doesn't happen to the next user. Right now that loop is broken at every step.

Environment

• Claude Code in VS Code (20 days of daily use)
• Extensive CLAUDE.md with rules, quality gates, and standing instructions
• .claude/rules/ directory with specific behavioral rules
• .claude/skills/ with quality assurance inspection framework
• All evidence documented in daily journals with specific examples