Claude Code 2.1.267: three captured instruction sets

Captured on 2026-09-10 UTC (2026-09-11 in Sydney), using direct CLI calls:
  claude -p hello --model claude-fable-5-1
  claude -p hello --model opus
  claude -p hello --model sonnet

A temporary local HTTP forwarding proxy, selected with ANTHROPIC_BASE_URL,
captured outbound /v1/messages JSON bodies and forwarded them to
https://api.anthropic.com. Anthropic returned real greetings; this was not
a mock response or a request for the model to recite its instructions.
Existing file-backed Claude authentication was supplied through the environment.
IS_SANDBOX was also inherited from the configured Claude provider environment.
The direct commands did not set --system-prompt or --append-system-prompt.

*.txt: top-level system text blocks in order, joined by exactly two newlines.
Text contents are verbatim, not a paraphrase. No trailing newline was added.
*-system-blocks.json: original top-level system array, reserialized as JSON;
this preserves block boundaries and cache metadata, not the original wire bytes.
*-instruction-messages.json: original messages array, reserialized as JSON.
It includes the supplied hello, CLI attribution reminder, and additional
system-role environment/agent/skill instructions. These are not included
in the character counts for the top-level system text.
*.diff: pairwise text diffs. A final newline is added to each input for diff
formatting only; the downloadable prompt files are unchanged.
manifest.json: exact hashes, sizes, model IDs, tool names, generation settings,
and request-file modification timestamps (not independently attested times).

No auth headers, account metadata, full tool schemas, or full raw requests
are published. The text retains non-sensitive local scratch/memory paths.
Fable and Sonnet used the same scratch working directory; Opus used a sibling.
These are observations of this configured installation, not an attestation
of universal defaults. There was one successful greeting per model; no
counterbalanced repeated experiment, cross-account test, or behavior benchmark.
The dated binary hash was observed during the audit, not independently
recorded at each invocation. All invocations reported Claude Code 2.1.267.