Claude Code on Bedrock or Vertex with a custom model name? Keep the full system prompt after the 2.1.312 update with one setting
TL;DR
What it is: Claude Code 2.1.312 (released 2026-10-27) now sends a shorter system prompt when your model name is one it does not recognize on Bedrock, Vertex, Foundry or Mantle. An application inference profile ARN is a typical example.
The key benefit: you decide. One environment variable,
CLAUDE_CODE_SIMPLE_SYSTEM_PROMPT=0, keeps the full prompt on any model.Time to try: about 5 minutes: update, add one line to your settings, restart.
What it is
Claude Code has two system prompts. The full one has complete tool descriptions. The shorter one has abbreviated tool descriptions. According to the env-vars docs, both keep "the full tool set, hooks, MCP servers, and CLAUDE.md discovery".
Until now, a model name Claude Code did not recognize could still get the full prompt. Since 2.1.312, the changelog says Bedrock, Vertex, Foundry and Mantle "use the shorter system prompt for unrecognized model names, even when they front Sonnet or Haiku". The Bedrock docs give the classic unrecognized name: an application inference profile ARN such as arn:aws:bedrock:...:application-inference-profile/.... A gateway alias works the same way.
If your team pins models through an ARN or an alias, your sessions changed today without a visible message. The same release also says how to opt out: CLAUDE_CODE_SIMPLE_SYSTEM_PROMPT=0.
Illustration
Three more changes in the same release that you will notice:
Change in 2.1.312 | What it means for you |
|---|---|
Write protection for big files | If Claude has read only the start of a very large CLAUDE.md, rules file or @-included file, Write now asks it to read the whole file or edit in place instead of overwriting it. |
Hook limit | Prompt and agent hooks run at most 24 at a time per session (sub-agents included). A |
Gateway sign-in | The Claude apps gateway gets |
Try it in 5 minutes
Prerequisites: Claude Code with Bedrock, Vertex, Foundry or Mantle and a custom model name (ARN or alias). On the direct Anthropic API you can skip step 2.
Update and check the version ``
bash claude update && claude --version`In Windows PowerShell 5.1, runclaude update; claude --version`.Pin the full prompt. Add the variable to the
envblock of~/.claude/settings.json(keep your other keys): ``json { "env": { "CLAUDE_CODE_SIMPLE_SYSTEM_PROMPT": "0" } }``Restart and check. Start
claudeagain and run/statusto confirm the provider and model the session uses.
What you will see: claude --version reports 2.1.312 or later, /status shows your provider (for example Amazon Bedrock), and the session sends the full prompt on any model.
Tip: to roll this out to the whole team, put the same
enventry in your project's.claude/settings.jsonor in managed settings, not in each person's shell.
Caveats
Cost: the full prompt is longer, so each request carries more input tokens. Measure with
/usageor your gateway logs before you pin it for everyone.Not always better: the docs say newer models use the shorter prompt by default. Pin
0mainly for older models behind an ARN or alias.No visible indicator: Claude Code does not say which prompt a session uses. The variable is the control. It is not a dashboard.
Related switch: the same release extends the automatic model switch after a flagged message to unrecognized model names. Set
CLAUDE_CODE_DISABLE_REFUSAL_FALLBACK=1if your gateway must stay on one model.
Claude Quick Hits
Manage costs and demonstrate ROI for Claude Code on Google Cloud (2026-10-27): budget alerts, spend caps, BigQuery billing export, plus per-user caps in the Claude apps gateway and OpenTelemetry metrics in Cloud Monitoring.
AI Tool Radar
Repo/Tool | What it does (1 line) | Stars (growth) | Quick try |
|---|---|---|---|
Local-first desktop workspace that runs Claude Code, Codex and other agents with task queues and agent-to-agent messages. Development preview; PolyForm Noncommercial license (not OSI); agents run with full access by default. | 2,498 (+777 today, trending daily) | Desktop installer from GitHub Releases | |
On-device desktop assistant that picks and runs an open-weight model (for example Gemma or gpt-oss) for your machine. MIT. | 271 (Show HN, 164 points) | Download the Mac, Windows or Linux build from Releases | |
| 9,483 (+1,395 today, trending daily) | Copy |
Sources
https://github.com/anthropics/claude-code/blob/main/CHANGELOG.md (2.1.312 entry)
https://github.com/anthropics/claude-code/releases (v2.1.312, 2026-10-27)
https://code.claude.com/docs/en/env-vars (
CLAUDE_CODE_SIMPLE_SYSTEM_PROMPT,CLAUDE_CODE_DISABLE_REFUSAL_FALLBACK)https://code.claude.com/docs/en/amazon-bedrock (application inference profile ARNs,
/status,envin settings)https://claude.com/resources/articles/how-to-manage-costs-and-demonstrate-roi-for-claude-code-on-google-cloud
https://github.com/trending?since=daily
https://github.com/nightly-labs/openbot
https://github.com/FrigadeHQ/hob
https://github.com/elayadesign/ai-design-skills
https://hn.algolia.com/api/v1/search_by_date?tags=show_hn