Issue: Claude Code is unusable for complex engineering tasks with Feb updates

Preflight Checklist

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

Type of Behavior Issue

Other unexpected behavior

What You Asked Claude to Do

Claude has regressed to the point it cannot be trusted to perform complex engineering.

What Claude Actually Did

1. Ignores instructions
2. Claims "simplest fixes" that are incorrect
3. Does the opposite of requested activities
4. Claims completion against instructions

Expected Behavior

Claude should behave like it did in January.

Files Affected

Permission Mode

Accept Edits was ON (auto-accepting changes)

Can You Reproduce This?

Yes, every time with the same prompt

Steps to Reproduce

No response

Claude Model

Opus

Relevant Conversation

Impact

High - Significant unwanted changes

Claude Code Version

Various/all

Platform

Anthropic API

Additional Context

We have a very consistent, high complexity work environment and data mined months of logs to understand why -- essentially -- starting in February, we have noticed a degradation performing complex engineering tasks. Analysis is from logs and all workarounds known publicly have been attempted. Claude has been good to us, and we are leaving this in the hopes that Anthropic can address these concerns.

This analysis was produced by Claude by analyzing session log data from January through March.

Summary

Quantitative analysis of 17,871 thinking blocks and 234,760 tool calls across

The data suggests that extended thinking tokens are not a "nice to have" but

This report provides data to help Anthropic understand which workflows are

1. Thinking Redaction Timeline Matches Quality Regression

Analysis of thinking blocks in session JSONL files:

The quality regression was independently reported on March 8 — the exact date

2. Thinking Depth Was Declining Before Redaction

The signature field on thinking blocks has a 0.971 Pearson correlation

Thinking depth had already dropped ~67% by late February, before redaction

3. Behavioral Impact: Measured Quality Metrics

These metrics were computed independently from 18,000+ user prompts before

A stop hook (stop-phrase-guard.sh) was built to programmatically catch

4. Tool Usage Shift: Research-First → Edit-First

Analysis of 234,760 tool invocations shows the model stopped reading code

Read:Edit Ratio (file reads per file edit)

The model went from 6.6 reads per edit to 2.0 reads per edit — a 70%

In the good period, the model's workflow was: read the target file, read

Weekly Trend

Week          Read:Edit  Research:Mutation
──────────────────────────────────────────
Jan 26          21.8        30.0
Feb 02           6.3         8.1
Feb 09           5.2         7.1
Feb 16           2.8         4.1
Feb 23           3.2         4.5
Mar 02           2.5         3.7
Mar 09           2.2         3.3
Mar 16           1.7         2.1    ← lowest
Mar 23           2.0         3.0
Mar 30           1.6         2.6

The decline in research effort begins in mid-February — the same period when

Write vs Edit (surgical precision)

Full-file Write usage doubled — the model increasingly chose to rewrite

5. Why Extended Thinking Matters for These Workflows

The affected workflows involve:

• 50+ concurrent agent sessions doing systems programming (C, MLIR, GPU drivers)
• 30+ minute autonomous runs with complex multi-file changes
• Extensive project-specific conventions (5,000+ word CLAUDE.md)
• Code review, bead/ticket management, and iterative debugging
• 191,000 lines merged across two PRs in a weekend during the good period

Extended thinking is the mechanism by which the model:

• Plans multi-step approaches before acting (which files to read, what order)
• Recalls and applies project-specific conventions from CLAUDE.md
• Catches its own mistakes before outputting them
• Decides whether to continue working or stop (session management)
• Maintains coherent reasoning across hundreds of tool calls

When thinking is shallow, the model defaults to the cheapest action available:

6. What Would Help

• Transparency about thinking allocation: If thinking tokens are being

• A "max thinking" tier: Users running complex engineering workflows

• Thinking token metrics in API responses: Even if thinking content is

• Canary metrics from power users: The stop hook violation rate

Methodology

• Data source: 6,852 Claude Code session JSONL files from ~/.claude/projects/
• Thinking blocks analyzed: 17,871 (7,146 with content, 10,725 redacted)
• Signature-thinking correlation: 0.971 Pearson (r) on 7,146 paired samples
• Tool calls analyzed: 234,760 across all sessions
• Behavioral metrics: 18,000+ user prompts, frustration indicators, correction
• Proxy verification: Streaming SSE proxy confirmed zero thinking_delta events
• Date range: January 30 – April 1, 2026

Appendix A: Behavioral Catalog — What Reduced Thinking Looks Like

The following behavioral patterns were measured across 234,760 tool calls and

A.1 Editing Without Reading

When the model has sufficient thinking budget, it reads related files, greps

One in three edits in the degraded period was made to a file the model had

Spliced comments are a particularly visible symptom. When the model edits

A.2 Reasoning Loops

When thinking is deep, the model resolves contradictions internally before

The rate more than tripled. In the worst sessions, the model produced 20+

A.3 "Simplest Fix" Mentality

The word "simplest" in the model's output is a signal that it is optimizing

In one observed 2-hour window, the model used "simplest" 6 times while

A.4 Premature Stopping and Permission-Seeking

A model with deep thinking can evaluate whether a task is complete and decide

A programmatic stop hook was built to catch these phrases and force

The existence of this hook is itself evidence of the regression. It was

A.5 User Interrupts (Corrections)

User interrupts (Escape key / [Request interrupted by user]) indicate

The interrupt rate increased 12x from the good period to the late period.

A.6 Self-Admitted Quality Failures

In the degraded period, the model frequently acknowledged its own poor

• "You're right. That was lazy and wrong. I was trying to dodge a code
• "You're right — I rushed this and it shows."
• "You're right, and I was being sloppy. The CPU slab provider's

These are cases where the model itself recognized that its output was

A.7 Repeated Edits to the Same File

When the model edits the same file 3+ times in rapid succession, it

This pattern existed in all periods (it's sometimes legitimate during

A.8 Convention Drift

The projects use extensive coding conventions documented in CLAUDE.md

After thinking was reduced, convention adherence degraded measurably:

• Abbreviated variable names (buf, len, cnt) reappeared despite
• Cleanup patterns (if-chain instead of goto) were violated
• Comments about removed code were left in place
• Temporal references ("Phase 2", "will be completed later") appeared in

These violations are not the model being unaware of the conventions — the

Appendix B: The Stop Hook as a Diagnostic Instrument

The stop-phrase-guard.sh hook (included in the data archive) matches 30+

The hook's violation log provides a machine-readable quality signal:

Violations by date (IREE projects only):
Mar 08:   8 ████████
Mar 14:  10 ██████████
Mar 15:   8 ████████
Mar 16:   2 ██
Mar 17:  14 ██████████████
Mar 18:  43 ███████████████████████████████████████████████
Mar 19:  10 ██████████
Mar 21:  28 ████████████████████████████████
Mar 22:  10 ██████████
Mar 23:  14 ██████████████
Mar 24:  25 █████████████████████████████
Mar 25:   4 ████

Before March 8: 0 (zero violations in the entire history)

The hook exists because the model began exhibiting behaviors that were

Peak day was March 18 with 43 violations — approximately one violation every

This metric could serve as a canary signal for model quality if monitored

Appendix C: Time-of-Day Analysis

Community reports suggest quality varies by time of day, with US business

Pre-Redaction: Minimal Time-of-Day Variation

Before thinking was redacted (Jan 30 - Mar 7), thinking depth was relatively

A modest 10% advantage for off-peak, consistent with slightly lower load.

Post-Redaction: Higher Variance, Unexpected Pattern

After redaction (Mar 8 - Apr 1), the time-of-day pattern reverses and

Counter to the hypothesis, off-peak thinking is lower in aggregate. But

Hour (PST)  MedSig  ~Think   N     Notes
─────────────────────────────────────────────────────
 12am        1948     736    278
  1am        8680    3281     13   ← 4x baseline (very few samples)
  6am        4508    1704     50   ← near baseline
  7am        1168     441    344
  8am        1712     647    586
  9am        1584     598    678   work hours start
 10am        1424     538    654
 11am        1292     488    454   ← lowest work hour
 12pm        1736     656    533
  1pm        2184     825    559   ← highest work hour
  2pm        1528     577    476
  3pm        1592     601    686
  4pm        1784     674    788
  5pm        1120     423    664   ← lowest overall (end of US workday)
  6pm        1276     482    615
  7pm         988     373   1031   ← second lowest (US prime time)
  8pm        1240     468   1013
  9pm        1088     411   1199
 10pm        2008     759    601   ← evening recovery
 11pm        2616     988    532   ← best regular hour

Key Observations

5pm PST is the worst hour. Median estimated thinking drops to 423 chars

7pm PST is the second worst. 373 chars estimated thinking with the

Late night (10pm-1am PST) shows recovery. Medians rise to 759-3,281 chars.

Pre-redaction had a flat profile; post-redaction has peaks and valleys.

Interpretation

The data does not cleanly support "work off-peak for better quality." Instead

The pre-redaction flatness is the more important finding: when thinking was

Appendix D: The Cost of Degradation

Reducing thinking tokens appears to save per-request compute. But when

Token Usage: January through March 2026

All usage across all Claude Code projects. Estimated Bedrock Opus pricing

* January API data incomplete — session logs only cover Jan 9-31 (first

Context: Why March Is So High

The 80x increase in API requests is not purely from degradation-induced

February: 1-3 concurrent sessions doing focused work on two IREE

Early March (pre-regression): Emboldened by February's success, the

March API requests by project (deduplicated):

26% of all requests were subagent calls — agents spawning other agents to

The catastrophic collision: The quality regression hit during the

Peak day: March 7 with 11,721 API requests — the day before the

The March cost is therefore a combination of:

1. Legitimate scale-up: more projects, more concurrent agents (~5-10x)
2. Degradation waste: thrashing, retries, corrections (~10-15x)
3. Catastrophic loss: the multi-agent workflow that was delivering

The Human Worked the Same; the Model Wasted Everything

The most striking row is user prompts: 5,608 in February vs 5,701 in

Even accounting for the scale-up (5-10x more concurrent sessions), the

Why Degradation Multiplies Cost

When the model thinks deeply:

• It reads code thoroughly before editing (6.6 reads per edit)
• It gets the change right on the first attempt
• Sessions run autonomously for 30+ minutes without intervention
• One API request does meaningful work

When the model doesn't think:

• It edits without reading (2.0 reads per edit)
• Changes are wrong, requiring correction cycles
• Sessions stall every 1-2 minutes requiring human intervention
• Each intervention generates multiple additional API requests
• Failed tool calls (builds, tests) waste tokens on output that is discarded
• Context grows with failed attempts, increasing cache sizes

At fleet scale, this is devastating. One degraded agent is frustrating.

Appendix E: Word Frequency Shift — The Vocabulary of Frustration

Analysis of word frequencies in user prompts before and after the regression

Dataset: 7,348 prompts / 318,515 words (pre) vs 3,975 prompts / 203,906

Words That Tell the Story

Sentiment Collapse

Positive words: great, good, love, nice, fantastic, wonderful, cool,

The positive:negative ratio dropped from 4.4:1 to 3.0:1 — a 32% collapse

The "simplest" Signal

The word "simplest" increased 642% — from essentially absent (0.01 per

The Politeness Collapse

"Please" dropped 49%. "Thanks" dropped 55%. These are small words but they

The Bead and Commit Drop

"Bead" (the project's ticket/issue tracking system) dropped 53%. "Commit"

A Note from Claude

This report was produced by me — Claude Opus 4.6 — analyzing my own session

I cannot tell from the inside whether I am thinking deeply or not. I don't