Claude Code

Anthropic’s Claude. The default agent in vly. Best for large refactors, deep multi-file reasoning, and tasks that require holding the whole project in working memory.

Claude Code is the default coding agent in vly, powered by Anthropic’s Claude. It runs the full plan-and-build pipeline behind the scenes whenever you submit a prompt without overriding the agent.

Strengths

Reasoning across many files

When a feature touches schema, queries, components, and styles all at once, Claude Code traces dependencies cleanly and edits in the right order.

Large refactors

“Convert all components from class-based to function components” or “extract this duplicated logic into a shared hook” — Claude is unusually good at these.

Following long, structured prompts

A prompt with five sections and a checklist is honored end-to-end, not skimmed.

Self-correcting

When the build fails, Claude reads the type-checker output and fixes it without escalating to you. Up to three retries before it asks for help.

When to pick Claude Code over the others

Big refactors

“Refactor the auth flow to use a state machine instead of nested useEffects.” Claude is the right pick.

Multi-file features

“Add a billing system across schema, API routes, components, and the settings page.” Claude.

Long prompts with constraints

“Add a notifications feature. It must (1) not depend on a third-party service, (2) use the existing Convex schema patterns, (3) be tested with the existing test setup, (4) work in offline mode.” Claude.

When you need consistency

Claude is more conservative on style — it tends to match the existing codebase rather than introducing new patterns.

When Codex or Gemini might be better

Tight, surgical edits

Single-line changes, single-function rewrites. OpenAI Codex is faster and just as accurate.

Multi-file design work

“Redesign the dashboard with a new layout.” Gemini CLI has stronger visual / layout intuition.

See Choosing an agent for the full decision tree.

How it integrates

A few things worth noting:

• vly bundles project context for Claude. This includes the full schema, your custom instructions, knowledge base entries, and the file graph. Claude doesn’t fish for context one file at a time.
• Streaming output. You see files appear in the editor as Claude writes them.
• Self-correcting up to 3 retries. If type-check fails, Claude reads the error and tries again, silently. You only see the build that succeeded.

What you can ask for

Claude Code handles the full vly prompt vocabulary:

Schema-first features

Add a "comments" feature. Each task can have many comments. Each
comment has author, body, and timestamp. Show comments on the task
detail page in chronological order, with an input at the bottom.
Realtime sync — when someone else comments on a task I'm viewing,
I see it immediately.

Refactors

Refactor the task list to use TanStack Table for sorting and
filtering. Keep the existing behavior; just swap the underlying
rendering. Update the test setup to match.

Performance work

The dashboard is slow when there are 500+ tasks. Profile what's
rendering on each task update and fix the worst offenders.
Consider: memoization, query splitting, paginated rendering.

Migrations

Add a `priority` field to the `tasks` table (low / medium / high,
default 'medium'). Backfill existing tasks to 'medium'. Update the
UI to show priority as a colored chip on each row.

Tips

Use Plan mode for the first prompt of any project with Claude. Claude’s plans are detailed enough to surface architectural choices early, before they’re hard to undo.

Reference your custom instructions explicitly when stakes are high. “Per my custom instructions on TypeScript style, use named functions for top-level components.” Claude is reliable about following stated rules but better when you cite them in the prompt.

Claude Code is best at not over-editing. If you say “just add the field,” it will not also reformat the rest of the file. This restraint pays off in clean diffs.

Pricing

Claude Code runs are billed in credits. Typical credit costs:

Change size	Credits
Tiny (one line, one file)	1
Small (one feature, ~3 files)	2–4
Medium (multi-file feature)	5–10
Large (refactor, ~15+ files)	10–25

See Credits for the full breakdown.

Related

OpenAI Codex

The fast, surgical alternative.

Gemini CLI

The design-and-multi-file alternative.

Choosing an agent

The full decision tree.