Why Our AI Coding Workflow Split Into Three Distinct Modes — And the Hotkeys That Switch Between Them

TL;DR: One tool can’t win at autonomy, context persistence, and daily editing simultaneously. We mapped each workflow to the tool that wins its axis: cc for greenfield, cx for 3–5 hr refactors, wind for daily type-heavy work. Total: ~$135/mo for 2 seats. Full workflow →

The Context

Two-dev team, 5 repos, 6–8 hrs/day AI-assisted. Tried forcing everything into one tool (Cursor, then Windsurf, then Claude Code). Each failed at something: Cursor lost context on long refactors; Claude Code too verbose for quick edits; Codex chat UI friction for terminal work. Stopped chasing “the one tool” and built a workflow map instead.

What We Tested

Workflow	Typical Duration	Context Need	Autonomy Need	Tool Winner	Why Loser Failed
Greenfield feature (new API, component, infra)	45–90 min	Low (fresh context)	High (test-write-debug loops)	Claude Code	Cursor: no autonomous loops; Codex: chat UI friction
Long refactor (migration, schema change, auth rewrite)	3–5 hrs	Critical (dev server, DB, tunnels)	Medium (guided steps)	Codex	Cursor: loses context at 90 min; Claude Code: session-only
Daily editing (type-heavy, bug fixes, small features)	15–60 min	Low	Low	Windsurf	Cursor: $20 mystery limits; Claude Code: verbose for quick edits
Model-flexible heavy lift (Terraform, security review)	2–4 hrs	Medium	Medium	Cline (BYOK)	All others: locked model routing

The Pivot Point

April 2026: Terraform refactor across 3 envs. Needed Sonnet for architecture, DeepSeek V3 for repetitive blocks, Opus for security review. Tried in Cursor — three separate chats, each hitting limits. Tried in Codex — can’t switch models mid-session. Cline + OpenRouter did it in one 4-hr session, $4.20 API cost. Realized: workflow ≠ tool. Workflow = (task type × context need × autonomy need × model need). Built the map.

What We Use Now

Terminal aliases in shared dotfiles (synced via chezmoi):

# ~/.config/chezmoi/dot_zshrc.tmpl
alias cc='claude-code'              # Greenfield, autonomous
alias cc-opus='cc --model opus --reason'  # Flagged Opus use
alias cx='codex'                    # Long refactors, persistent
alias wind='windsurf'               # Daily editing, Cascade
alias cline='code --extension cline'      # Model-flexible heavy lift

Workflow decision tree (printed on desk):

New feature / greenfield? → cc
  └─ Needs Opus (auth/infra/payments)? → cc-opus "reason"
Refactor / migration / debug > 90 min? → cx
  └─ Needs model switching? → cline
Daily edits / bug fixes / type-heavy < 60 min? → wind
  └─ Concurrent FE/BE? → wind (Cascade 3 agents)

Cost ceiling: $135/mo team (Claude Code pool $100 + Windsurf 2×$15 + Codex in existing ChatGPT Plus). Cline BYOK variable (~$10-20/mo). Hard cap enforced by monthly trackmycodes review.

When You’d Choose Differently

• Solo dev: Pick ONE primary (Cursor or Windsurf or Claude Code) + one backup. Complexity not worth it.
• Team > 5: Standardize on one tool for collaboration (Cursor Business or Windsurf Teams) — workflow splitting creates onboarding friction.
• Enterprise compliance: Copilot Business or Cursor Business — audit trails, SSO, data residency.

Tool Crucible Rating

Dimension	Rating (1–5)	Notes
Overall	4.5	Map > tool; eliminates “which tool?” paralysis
Ease of Use	3	Requires team discipline + shared aliases
Value	5	Right tool per task = no overage, no context loss
Support	N/A	Internal framework

---

This is part of our AI Coding Tool Evaluation series. See full workflow: The Three-Mode AI Coding Workflow Map (2026 Edition)