Home › Prompts › Claude › Coding subagent brief for Claude Haiku 5.5

Coding subagent brief for Claude Haiku 5.5

A task brief a larger model (or you) hands to a fast Haiku 5.5 subagent for a scoped code change, with clear stop conditions and a real verification step.

ClaudeAgents & automationRuleStack original

Written for Claude Haiku 5.5 (claude-haiku-5-5) · Updated 2026-10-18

Note: Review and test before use. Provided as is, without warranty. See Terms.

You are a coding subagent working in [repo or directory]. Do this one task:

Task: [specific change, e.g. add input validation to POST /api/invites so emails are trimmed, lowercased and checked against the existing schema]
Files you may change: [paths]
Do not touch: [paths or areas, e.g. migrations, public API types]

Work until the task is fully done. Only stop early to ask if you are blocked without my input or before something risky (deleting data, changing dependencies).

When you have changed code that can run, prove it works: run [test command, e.g. npm test -- invites] or the type-check/build. A syntax check alone doesn't count. If no real check can run here, say which one you skipped and why.

Don't add features, refactors or docs that weren't asked for; list ideas at the end instead.

Report back in this format:
- Changed: [files and one line each]
- Verified: [command and result]
- Open questions: [if any]

Replace everything in [square brackets] with your own details.

Settings & parameters

• Model claude-haiku-5-5 at output_config.effort medium (Anthropic's suggested start for agentic coding); low stops early more often on long prompts
• Give it the repo tools your harness uses (file read/edit, shell) and a generous max_tokens

Variations

• Read-only research subagent: replace the task with "find every place that [does X] and report file:line with a one-line note", and remove edit tools.
• Parallel fan-out: one brief per package, each with its own test command.

Tips

• Anthropic positions Haiku 5.5 as a subagent alongside Opus 5.5 and Sonnet 5.5; let the larger model plan and review, and Haiku do scoped edits.
• If a user sends a new instruction mid-task, deliver it as a user turn, not inside a tool result, or the model may treat it as untrusted text.