Prompt Variants

Prompt variants

It shows the system prompt variants of dotclaude, their byte budgets, and the gap in the compact prompt. Source labels are in Home.

The variants

The hooks module replaces the Claude Code sections intro, system, doing_tasks, actions, tools, tone, and lean_body with the variant of the session model. A variant holds only the text of one model. The rules that every model shares are the shared sections of features/prompt/shared.mjs, and variantSections puts them before the model notes (System prompt). The model is Opus 5.5, Sonnet 5.5, Haiku 5.5, Fable 5.1, or Mythos 5.1, which gets the Fable variant. A section with another id passes on unchanged. A model with no variant keeps the sections of Claude Code. Source: plugins/dotclaude/features/prompt/variants.mjs and plugins/dotclaude/features/prompt/shared.mjs.

Variant	Model notes	Verification line in the subagent block
Opus 5.5	One sentence before the first tool call, and updates only for new findings	off
Sonnet 5.5	none	on
Haiku 5.5	Finish each step before the next one	on
Fable 5.1 and Mythos 5.1	Targeted edits (one guide covers both models, unverified until the variant evals)	off

The shared verification and doing_tasks sections hold the verification rule and the extras rule for the main session, so no model notes repeat them (cuts V2 and V7 of task 29.3).

Each variant has at most one model_notes section with XML tags, and Sonnet 5.5 has none. The block that agent.spawn adds to a task prompt has the model lines, the block line (BLOCK_LINE), and the evidence line of design D30. A subagent does not get the shared sections, so the block carries both lines. The block line says that a failing check, a lint rule, a test, a bound, and a hook each stop the work, and that the agent changes the work to fit them or reports the block to the parent. The evidence line asks for a new check before Claude undoes finished work because of a claim of fault, also when the user tells it to delete the work.

When the variant is chosen

• The main session gets its variant from prompt.compose (tested, probe 2.7). The hook chooses the variant at the first compose of a session and keeps it. A model switch with /model does not change the prompt bytes, because a changed system prompt breaks the cache.
• A subagent has its agent body as its system prompt, because prompt.compose and prompt.section do not run for it (tested, probe 2.7). The agent.spawn hook adds a short block of the subagent model to the end of the task prompt, one time. The hook reads the model of the spawn, or the model of the parent when it has none. A model that only the agent file sets is not visible to the hook (inferred), so that subagent gets no block.

Size of the first request

A variant has no byte bound of its own. The text of the first request, with no tool schemas, has the targets FIRST_REQUEST_LOW_BYTES (8 KB) and FIRST_REQUEST_HIGH_BYTES (16 KB) in tools/limits.mjs. just measure prints the size and checks them, and it writes no file, where 1 KB is 1,024 bytes. A size above the high bound fails, and a size below the low bound is reported.

The compact prompt gap

Claude Code has no route to override the compact prompt. The prompt.compose and prompt.section hooks do not reach it. The settings snippet limits compaction instead: DISABLE_COMPACT turns it off, and the session stops at the context limit. The user then starts a fresh session with /clear.