Qwen Delegate

Qwen Delegate

Delegate token-cheap tasks to the qwen CLI (Qwen Code, model coder-model), preserving Anthropic PRO tokens for tasks that genuinely require Claude's reasoning.

Note: The qwen CLI is cloud-backed, not a local LLM runner (it does not use Ollama, LM Studio, or similar). It is an agent CLI that runs against Qwen/Alibaba cloud infrastructure, authenticated via OAuth or Alibaba Cloud Coding Plan. Throughout this skill, references to "Qwen" mean the cloud-backed qwen CLI unless explicitly stated otherwise. Delegation avoids Anthropic API calls, not all cloud calls.

Delegation Criteria

Delegate to Qwen when all three conditions are met:

1. Task type matches one of the delegatable categories below
2. No deep reasoning required — transformation, generation, or summarisation suffices
3. Output requires Claude validation — Qwen result is always reviewed before delivery

Delegatable Task Categories

Category	Examples
Summarisation	Summarise a document, meeting notes, PR description
Formatting	Convert Markdown→HTML, reformat JSON/YAML, fix indentation
Translation	Translate text to another language
Boilerplate generation	Scaffold a class, generate CRUD stubs, create a config template
Code generation	Write a function from a spec, implement a known algorithm
Refactoring	Rename identifiers, extract a method, apply a pattern
Text transformation	Convert lists to tables, rewrite in different tone
Simple analysis	Count items, extract fields, detect language

Do NOT delegate when:

• The task requires architectural decisions or multi-file reasoning
• Security-sensitive code is involved and correctness is critical
• The user explicitly asked for Claude's opinion or judgment
• The task is faster to do directly than to delegate (< 5 lines)

Command Patterns

Pattern 1 — Text-only task (no file access)

qwen "PROMPT" \
  --output-format text \
  --approval-mode yolo \
  --max-session-turns 1 \
  --exclude-tools run_shell_command,edit,write_file

Use for: summarisation, translation, formatting, explanation tasks. Use --max-session-turns 1 only for pure text transformations; Qwen's internal thinking counts as turns — if the task is complex, increase to 2.

Pattern 2 — File context via stdin

cat FILE_PATH | qwen "TASK DESCRIPTION" \
  --output-format text \
  --approval-mode yolo \
  --max-session-turns 2 \
  --exclude-tools run_shell_command,edit,write_file

Use for: summarising a file, translating a document, reformatting content.

Pattern 3 — Code generation or refactoring (without file access)

qwen "TASK DESCRIPTION" \
  --output-format text \
  --approval-mode yolo \
  --max-session-turns 3 \
  --exclude-tools run_shell_command,edit,write_file

Use for: generating a function, writing an algorithm, refactoring a snippet. Minimum 3 turns — Qwen uses thinking turns internally; --max-session-turns 1 causes exit code 53 on code tasks.

Pattern 4 — Code task with file access (multi-step)

qwen "TASK DESCRIPTION" \
  --output-format text \
  --approval-mode yolo \
  --max-session-turns 5

Use for: scaffolding a module, refactoring across multiple files, tasks that require reading existing code.

Why no --exclude-tools here: This pattern intentionally allows Qwen to use read_file, glob, and grep_search to read existing code before generating output. Write/edit/shell tools (edit, write_file, run_shell_command) are still blocked by default in --approval-mode yolo without explicit confirmation — Qwen Code will prompt before any write/execute action. Validate all generated code with Claude before applying it to the project.

Pattern 5 — Custom system prompt for specialised tasks

qwen "PROMPT" \
  --output-format text \
  --system-prompt "FOCUSED SYSTEM PROMPT" \
  --approval-mode yolo \
  --max-session-turns 2 \
  --exclude-tools run_shell_command,edit,write_file

Use for: tasks that benefit from a specific persona or tight output constraints (e.g., "You are a JSON formatter. Return only valid JSON.").

Execution Workflow

1. Identify — Determine if the task matches the delegation criteria above.
2. Select pattern — Choose the command pattern that fits (text-only, stdin, code, custom prompt).
3. Compose prompt — Write a clear, self-contained prompt for Qwen. Include all necessary context inline or via stdin.
4. Execute — Run the qwen command via Bash tool.
5. Validate — Review Qwen's output critically:
   • Correctness (logic, syntax, facts)
   • Completeness (all requirements addressed)
   • Quality (style, conventions, security)
6. Fix or augment — Apply corrections directly; do not re-delegate fixes to Qwen.
7. Present — Deliver the validated result to the user. Optionally note: "Generated by Qwen CLI (coder-model), validated by Claude."

Prompt Composition Guidelines

• Be explicit: include all context Qwen needs in the prompt (do not assume it has conversation history).
• For code tasks, specify: language, function signature, expected behaviour, edge cases.
• For text tasks, specify: output format, length constraints, tone.
• Keep prompts focused — one task per delegation call.

Example: Summarisation

cat docs/api-reference.md | qwen \
  "Summarise this API reference in 5 bullet points, focusing on authentication and rate limits." \
  --output-format text \
  --approval-mode yolo \
  --max-session-turns 1 \
  --exclude-tools run_shell_command,edit,write_file

Example: Code generation

qwen "Write a Python function named 'parse_iso_date' that accepts a string and returns a datetime object. Handle invalid input by raising ValueError with a descriptive message. Include type hints." \
  --output-format text \
  --approval-mode yolo \
  --max-session-turns 2 \
  --exclude-tools run_shell_command,edit,write_file

Example: JSON formatting

echo '{"name":"test","version":"1","deps":["a","b"]}' | qwen \
  "Format this JSON with 2-space indentation and sort keys alphabetically." \
  --output-format text \
  --approval-mode yolo \
  --max-session-turns 1 \
  --exclude-tools run_shell_command,edit,write_file

Validation Gate (Mandatory)

Every delegation result must pass Claude's review before being used or shown to the user:

• Code: Check logic, syntax, edge cases, security implications.
• Text: Check accuracy, completeness, tone, formatting.
• Structured output: Validate schema, field types, required fields.

If Qwen's output is incorrect or incomplete, correct it directly — never present unvalidated Qwen output.

Attribution

When presenting delegated results, optionally include a one-line note:

Drafted by Qwen CLI (coder-model), reviewed and validated by Claude.

This maintains transparency without being verbose. Omit attribution for minor transformations (formatting, renaming) where it adds no value.