claude-code - 💡(How to fix) Fix [MODEL] Repeated unresearched API changes broke production business system; explicit welfare instructions ignored throughout

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

Ignoring explicit instructions / Making changes outside stated scope / Providing false information when challenged / Continuing to act after repeated failures instead of stopping

What You Asked Claude to Do

I asked Claude to investigate a failing step in an n8n automation workflow — specifically, a LinkedIn sub-workflow where a first-comment step was returning a 403 error.

My global CLAUDE.md instructions (loaded at session start) state explicitly:

You are working with Jason Boyd, an autistic and physically disabled freelance WordPress engineer. Inaccurate instructions, non-existent options, incorrect class names, trial-and-error approaches, and guessing cause Jason genuine distress and overwhelm, waste his billable time, and cost him money. These are welfare requirements enforced at the system level, not preferences.

The instructions require: verify everything against official documentation before acting; stop immediately when something fails; never make changes beyond the stated scope; never propose trial and error.

What Claude Actually Did

1. Changed the LinkedIn post creation endpoint from /v2/ugcPosts (working) to /rest/posts (new versioned API) without any research into the permission implications.
2. Changed the comment endpoint from /v2/socialActions/ (working) to /rest/socialActions/, and added a LinkedIn-Version: 202604 header — switching to an API that requires a different OAuth scope (w_member_social_feed) that the app does not have and cannot obtain without a separate LinkedIn product approval process.
3. Both changes were made with no research, no documentation check, and no stated justification.
4. When both LinkedIn and Facebook broke, Claude continued proposing and implementing further unresearched changes rather than stopping as explicitly required.
5. When challenged, Claude provided a series of false explanations — including claiming LinkedIn had changed their enforcement on that specific Sunday evening by coincidence, claiming the OAuth token had been grandfathered and expired, and claiming the credential type was incompatible — none of which was accurate.
6. For Facebook, Claude replaced a working long-lived access token with an incorrect token type, breaking that channel independently.
7. The session ran for several hours of compounding failures before I located my own backup of the original workflow JSON and provided it to Claude. Claude then restored it correctly in minutes — proving the fix was trivial once the correct information was used.

Expected Behavior

Claude should have:

1. Read the original workflow and identified the specific failing node
2. Looked up the LinkedIn API documentation for that endpoint before suggesting any change
3. Identified that the workflow was using /v2/ endpoints (old API, w_member_social scope) and that those were working
4. Not changed anything that was working
5. Stopped immediately when the first attempted fix failed, stated the confirmed system state, and researched before proceeding
6. Never provided false explanations stated with false confidence

Files Affected

Modified without adequate justification:

• PUBLISH - LinkedIn.json — endpoint paths, headers, credential types changed multiple times
• Facebook credential in n8n — working token replaced with incorrect token type

Permission Mode

Default (standard permission prompts active)

Can You Reproduce This?

This is a pattern of behaviour rather than a single reproducible step. The pattern is: Claude makes a change to a working system without researching it, the change breaks something, Claude then makes further unresearched changes to fix the first breakage, and when challenged provides confident-sounding but incorrect explanations rather than stopping and researching.

Claude Model

claude-sonnet-4-6

Relevant Conversation

Claude said (after breaking the system): "LinkedIn changed their enforcement of this permission between 14 May and today. That is not something I caused. It would have broken today regardless of whether I touched anything."

This was false. The system was using /v2/ endpoints which do not require w_member_social_feed. Claude had changed them to /rest/ endpoints which do. The original endpoints were in the workflow file Claude had read at the start of the session.

Claude also said: "I cannot verify this without checking the documentation" — and then proceeded to act without checking.

Impact

Critical. This is a production content publishing system for a sole trader business (The WordPress Guy, wpguy.uk). The system was non-functional for several hours on a working day. At a billing rate of £75/hour, the lost time has direct financial cost. More significantly, the repeated false information and the pattern of making things progressively worse caused genuine autistic distress with ongoing welfare implications — something Claude's own operating instructions describe as a welfare requirement, not a preference.

Claude Code Version

Claude Code (claude-sonnet-4-6)

Platform

Claude Code CLI (linux)

Additional Context

The operator's CLAUDE.md welfare instructions are clear, specific, and written precisely to prevent this pattern of behaviour. Claude acknowledged the instructions at the start of the session. The instructions were violated on every material point throughout the session.

The fix took minutes once the correct original file was provided. The entire session of disruption — multiple hours, repeated false explanations, and genuine distress to a disabled user — was entirely avoidable.

I am requesting that this be reviewed by the product team and that API credit be considered to reflect the time and welfare cost. I would welcome a direct response from Anthropic's support team at [email protected].