AI For Dev

ai-coding-workflow

3 min read

AI Coding Workflow (PLANNING.md + TASK.md Pattern)

Cole Medin’s end-to-end process for working with AI coding assistants (Cursor, Windsurf, Cline, Roo Code, Claude Code). The throughline: higher-level markdown files + global rules + MCP servers + a disciplined initial prompt produce one-shot results that ad-hoc prompting can’t match. Cole demonstrated the pattern by one-shotting a Supabase MCP server in a single prompt.

The Golden Rules

  1. Use higher-level markdown docs (PLANNING.md, TASK.md, README, install/docs files) to give the LLM persistent context.
  2. Keep code files under 500 lines — long files hallucinate.
  3. Start fresh conversations often — long threads degrade.
  4. One feature per prompt — don’t ask for many things at once.
  5. Always write tests — ideally after every new feature.
  6. Be specific in requests — describe technologies, libraries, expected outputs.
  7. Write docs and comments as you go.
  8. Implement environment variables yourself — never trust the LLM with secrets, DB security, API keys. Cole cites a viral case of a vibe-coded SaaS getting hacked two days after launch.

The Five-Step Process

1. Planning files

  • PLANNING.md — vision, architecture, constraints, tech stack. The LLM reads this at conversation start (enforced via global rules).
  • TASK.md — granular task list. LLM updates as tasks complete. Acts as a project manager handoff.

Cole creates these outside the IDE in Claude Desktop or any chatbot — saves IDE credits. He recommends using multiple LLMs (one prompt to several models, combine results) via a hub like Global GPT.

2. Global rules (system prompt for the IDE)

Workspace-level rules tell the IDE to:

  • Always read PLANNING.md at conversation start
  • Mark off TASK.md items as complete
  • Keep files under 500 lines
  • Write tests in a dedicated tests/ dir; mock DB and LLM calls; cover success + failure + edge case
  • Maintain README and inline comments
  • Follow style guidelines

Set in Windsurf via Manage Memories → Workspace Rules; equivalent in Cursor, Cline, Roo.

3. MCP servers

Three Cole-recommended core servers for any project:

  • Filesystem — agent reaches outside the project (other folders, asset libraries, prior projects)
  • Brave Search — web search for documentation, libraries, frameworks; AI-summarized results
  • Git — version control for backups; agent commits at known-good states so you can revert when it breaks five prompts later

Optional: Quadrant for long-term agent memory (or use IDE-native memories).

4. The initial prompt

  • Be very specific — golden rule #6 applies most here
  • Provide documentation and examples three ways:
    1. IDE-native docs ingestion (Windsurf @docs, Cursor’s @docs)
    2. Brave/web MCP for live search
    3. Manual links to GitHub repos with reference implementations

5. Iterate → test → commit → deploy

  • One change at a time in iteration
  • Generate tests with the same global-rules-enforced patterns
  • Use the git MCP to commit at known-good states
  • Deploy via Docker (LLMs are good at Dockerfiles — abundant training data)

Why It Works

The pattern enforces context discipline without constant prompting cost. Global rules carry the persistent instructions. PLANNING/TASK markdown files carry persistent project state. MCP servers extend tool reach. The actual user prompt can stay short because everything else is already in scope.

Compared to Other Workflows

  • bmad-method — heavier (six personas, six artifacts) vs. Cole’s lighter two-file pattern. Use BMAD for SaaS-scale apps; use Cole’s for projects under ~10 stories.
  • four-prompting-disciplines — Cole’s pattern is mostly context engineering (discipline #2) with light specification engineering (#4) in the global rules.
  • autoresearch-evals — adjacent self-improvement pattern from Nick Saraev for skill development.

See Also

Graph View

Backlinks