This page gives the system prompt variants of claudetto 0.1.0. Read it to learn which prompt each model gets, and how the prompt size is bounded. It gives the model of each variant, the verification line of each subagent block, and the size bounds of the first request. It also gives the gap in the compact prompt. The rules that every model shares are on the page Prompt.

The prompt slice holds one variant for each model family. The source is src/plugins/claudetto/slices/prompt/domain/variants.ts.

Variant	Model	Model notes	Verification line in the subagent block
opus	Opus 5.5	One sentence before the first tool call, and updates only for new findings	off
sonnet	Sonnet 5.5	none	on
haiku	Haiku 5.5	Finish each step before the next one	on
fable	Fable 5.1, and Mythos 5.1	Edit a file at the lines that need the change	off

• A variant holds the text of one model only.
• Each variant has at most one <model_notes> section. Sonnet 5.5 has none.
• The shared verification and doing_tasks sections hold the rules for the main session. No variant repeats them.
• A model id that contains mythos gets the fable variant. The mapping is in src/plugins/claudetto/slices/prompt/domain/apply.ts.
• The variant of each model is set in VARIANTS, in the same file.

• The prompt.compose hook chooses the variant at the first compose of a session. It keeps that variant. The code is in src/plugins/claudetto/slices/prompt/adapters/mod.ts.
• A model switch with /model does not change the prompt bytes. A changed system prompt would break the prompt cache.
• A subagent uses its agent body as its system prompt. The agent.spawn hook gets the model that the engine resolved for the subagent from the result of next(e). Then the session.append hook adds a short block of the model notes to the end of the task row of the subagent, one time. So an agent with no model on the call gets the block of the model that runs it, also when its agent file or CLAUDE_CODE_SUBAGENT_MODEL chooses that model. The block of each model is in VARIANTS.
• The block of a subagent carries the block line and the evidence line. The block of Sonnet 5.5 and Haiku 5.5 also carries the verification line, and the block of Haiku 5.5 also carries the step line. The block line says that a failing check, a lint rule, a test, a bound, and a hook each stop the work. The evidence line asks for a new check before Claude undoes finished work because of a claim of fault.

The first request has a token bound, START_TOKENS (9,960 tokens) in tools/limits.ts. tools/measure.ts passes it to claudetto-start as --bound. The text of the first request, with no tool schemas, has two targets in tools/limits.ts.

• FIRST_REQUEST_LOW_BYTES is 8 KB.
• FIRST_REQUEST_HIGH_BYTES is 16 KB.

just measure runs tools/measure.ts, which reads both bounds. A size above the high bound fails the check. A size below the low bound is reported. In this page, 1 KB is 1,024 bytes.

The prompt.compose and prompt.section hooks do not reach the compact prompt of Claude Code. The handoff slice replaces the built-in compaction by default instead (Handoff). The context_management section agrees with the state of the session. The states are handoff, built-in, and off (DISABLE_COMPACT is on). The built-in state keeps the built-in text, and the compact prompt of Claude Code stays unchanged. The settings snippet does not set DISABLE_COMPACT (Settings Snippet).

The 0.1.0 basis changed from dotclaude 1.0.0. That version set DISABLE_COMPACT to 1 and autoCompactEnabled to false in its snippet, so a session stopped at the context limit. See openspec/changes/claudetto-0-1-0/design.md.

• Evidence: the index of the evidence pages and their source files.
• Prompt: the rules that every model shares.
• Handoff: the replacement of the built-in compaction.