How do I plan a task in Claude Code?

Start in plan mode with a clearly defined deliverable, such as a plan file or spec document. Break complex tasks into smaller units and save the plan to a file like CLAUDE.md or TASK.md so Claude can reload it each session.

How can I reduce token usage in Claude Code workflows?

Plan carefully before executing and batch your verification processes so Claude runs fewer passes. Use workflows sparingly since dynamic iterations can consume a significant number of tokens in minutes.

Can I use other models alongside Claude Code?

Yes. Users orchestrate multiple models by using Claude for planning, architecture, and deeper reasoning, then switching to cheaper models like Gemini or DeepSeek for execution and iteration to manage costs.

What are custom agents in Claude Code?

They are user-defined agents that automate specific tasks within a session. Common examples include a critic agent that checks what was missed or a hygiene skill that confirms documentation is up to date.

How to Build Iterative Workflows in Claude Code

Q: What is the 70% wall in iterative workflows?

It is the point where projects stall at 70 to 80 percent completion due to accumulating bugs or unforeseen issues. Users suggest stopping to triage by listing every bug in a single file, sorting by criticality, then fixing in priority order.

Users build iterative workflows in Claude Code by starting with a planning phase to outline goals and deliverables, then implementing changes, and refining through feedback loops. The approach works for code reviews, debugging, project planning, and data analysis.

To manage Claude's context window and avoid repetition, users externalize state into files like plan.md and CLAUDE.md so Claude can read relevant sections on demand. This keeps sessions focused without re-establishing context each time.

Advanced setups pair Claude Code with cheaper models for execution, build custom critic agents to catch gaps, and connect external tools through MCP servers for autonomous context gathering and verification.

Key techniques

Planning and review stages Generate CLAUDE.md or TASK.md, implement, then review in a testing loop
Externalized state files Save context to plan.md and tasks.md and reload each run instead of re-establishing
Multi-model orchestration Use Claude for planning and reasoning and cheaper models for execution
Custom critic agent Runs a pass asking what was not checked before you finish
MCP server integrations Connect Claude Code to Slack, Jira, Linear, and Datadog through a browser extension

How to Build Iterative Workflows in Claude Code — infographic

Users users leverage iterative workflows in Claude Code for everything from code reviews and debugging to project planning and data analysis, with a common strategy involving initial planning with Claude, followed by implementation and continuous refinement through feedback loops.

Optimizing Iterative Workflows

Define Clear Goals and Deliverables: Clearly define the task and expected outcomes at the outset, and break down complex tasks into smaller, manageable units. "You can and should define the deliverable in plan mode. 'The deliverable of this session is a plan file containing x' or whatever."

Leverage Planning and Review Stages: Many users start with a planning phase in Claude Code, often generating a detailed CLAUDE.md or TASK.md file, then implement the changes, and finally use Claude or other tools for review and debugging. "Planing, Spec definition, Implmentation, Design, Testing loop."

Externalize Context and State: To manage Claude's context window and avoid repetition, externalize important information into files like plan.md or CLAUDE.md, allowing Claude to read relevant sections on demand rather than re-establishing context every session. "Biggest unlock for me was saving state to files (plan.md, tasks.md) and reloading each run."

Addressing Challenges in Iterative Workflows

Mitigate Token Usage: Dynamic workflows can consume a significant number of tokens; users recommend prompting Claude to plan carefully and batch verification processes to optimize usage. "Seriously though it ate those tokens in less than 10 minutes so genuinely use workflows sparingly"

Overcoming the "70% Wall": When projects stall at 70-80% completion due to accumulating bugs or unforeseen issues, Users suggest stopping to triage, listing all bugs, and sorting them by criticality. "stop fixing things as you find them. do one sweep where you only write every bug/gap into a single list (i keep a FIXES.md), no fixing allowed during the sweep."

Ensuring Accuracy and Verification: Implement robust testing and verification steps, potentially using hooks to run test suites after every file edit, or having Claude directly access external tools for validation. "Implement robust testing into your builds."

Advanced Techniques and Tools

Multi-Model Orchestration: Users are integrating Claude Code with other models like Gemini or DeepSeek, leveraging Claude for planning and reasoning, and cheaper models for execution and iteration to manage costs and maximize efficiency. "Claude → planning, architecture, deeper reasoning Gemini CLI → execution, expansion, iteration, shipping"

Custom Agents and Skills: Create custom agents and skills within Claude Code to automate specific tasks, such as a "critic" agent that asks what wasn't checked, or a /hygiene skill to confirm documentation is updated. "I'm a big fan of the critic near the end. He runs a pass to ask 'what did we not check?'"

MCP Servers and External Integrations: Utilize MCP (Multi-Agent Communication Protocol) servers to connect Claude Code with external tools like browser automation, Slack, Jira, or Datadog, enabling Claude to gather context and verify actions autonomously. "I built an open-source MCP server that connects Claude Code to web apps (Slack, Jira, Linear, Datadog, etc.) through a Chrome extension, using whatever login sessions are already active."

Are you interested in exploring specific examples of these iterative workflows in more detail?