workflow-mastery

Workflow Mastery for .NET

Core Principles

1. Parallel over sequential — Run 3-5 Claude sessions simultaneously using git worktrees. Build a feature in one, fix a bug in another, run tests in a third. The single biggest productivity unlock.
2. Plan then execute — For any non-trivial task, start in plan mode, iterate until the plan is bulletproof, then switch to auto-accept. A good plan means Claude 1-shots the implementation.
3. Verification closes the loop — Give Claude a way to prove its work: dotnet build, dotnet test, get_diagnostics via MCP. This single practice 2-3x the quality of the output.
4. Context is a budget, not a dumping ground — The context window fills fast: a typical .cs file is 500-2000 tokens, and 50 file reads can burn a large share of the budget. Spend tokens like sprint capacity — deliberately.
5. Automate the repetitive — If you do it more than once a day, make it a hook, a slash command, or a subagent. Pre-allow safe permissions. Eliminate friction.
6. Compound your knowledge — Every correction becomes a rule in MEMORY.md (see instinct-system skill). Every PR review adds a learning. Over time, Claude's mistake rate drops because your project's knowledge base grows.

Patterns

Parallel Sessions with Git Worktrees

The biggest productivity multiplier. Each worktree gets its own Claude session, its own files, zero conflicts.

# Create worktrees for parallel work
git worktree add ../my-project-feature origin/main
git worktree add ../my-project-bugfix origin/main
git

Practical .NET workflow:

Tips:

• Name your terminal tabs by task so you never lose track
• Use shell aliases (alias zf='cd ../my-project-feature') for one-keystroke switching
• Enable terminal notifications so you know when a session needs input

Auto-Format Hook for .NET

Catch formatting issues on every file write — eliminates the "CI failed on formatting" loop.

Why || true: The hook should never block Claude's workflow. If formatting fails (e.g., on a non-C# file), silently continue.

Pre-Allow Safe .NET Permissions

Stop clicking "allow" for every dotnet command. Add these to .claude/settings.json:

Check this into git so the whole team gets frictionless workflows.

Plan Mode Strategy

For any task touching 3+ files or involving architecture decisions:

Advanced pattern: Have one Claude write the plan, then spin up a second Claude session to review it as a staff engineer:

When things go sideways: The moment implementation deviates from the plan, STOP. Don't push through. Switch back to plan mode, understand what changed, re-plan, then resume.

Verification Loop for .NET

For the full 7-phase verification pipeline (build, diagnostics, anti-patterns, tests, security, format, diff review) with structured PASS/FAIL reporting, see the verify skill.

Boris's #1 tip: "Give Claude a way to verify its work." The short version: always tell Claude to run dotnet build, dotnet test, get_diagnostics, and dotnet format --verify-no-changes before declaring done. The verify skill has the complete pipeline with short-circuit rules and report templates.

Compounding Knowledge via Corrections

For the full correction capture system — detection, generalization, categorized storage, and periodic audits — see the instinct-system skill. The short version: after every correction, capture a generalized rule in MEMORY.md so the same mistake never recurs.

Prompting Techniques for .NET

Challenge Claude's work:

Demand proof:

After a mediocre fix:

For EF Core migrations:

Subagent Patterns for .NET

The kit ships 10 specialist agents — route to them before writing your own: dotnet-architect, code-reviewer, refactor-cleaner, test-engineer, security-auditor, build-error-resolver, ef-core-specialist, api-designer, performance-analyst, devops-engineer. Each carries pre-loaded skills and domain context a generalist session lacks.

Use them:

For workflows the kit does not cover, create project-specific subagents in .claude/agents/ — a markdown file with a role, a numbered job list, and a required report format (PASS with summary / FAIL with specifics). Keep one concern per agent so its output stays reviewable.

When to offload vs. stay in main context: see Context Discipline below — subagents are also your context isolation chambers, not just task runners.

Context Discipline

The rules in .claude/rules/agents.md already mandate MCP-first navigation (find_symbol before file reads, get_diagnostics over builds). This section is the strategy layer on top: how to budget, when to offload, and how to recover.

Token Economics

A Roslyn MCP query costs 30-150 tokens; a file read costs 500-2000+. To understand OrderService, four MCP calls (find_symbol → get_public_api → find_references → get_type_hierarchy) cost ~310 tokens; reading the four related files costs ~2900. Then read only the method you'll modify. Reserve full file reads for files you are about to edit.

Subagent Offloading Decision Matrix

Ask subagents for compressed answers: "Trace the auth flow from login to token validation. Return numbered steps with file:line references." You get ~300 tokens of findings instead of 15k tokens of raw files.

File Reading Prioritization

Budget Planning and Recovery

Before a complex task, sketch the spend: understand ~5k (MCP + subagent), plan ~2k, implement ~15k (read targets + write + iterate), verify ~3k — leaving the bulk of the window for conversation.

Lazy Skill Loading

Don't front-load skills "just in case" — 15 skills at ~300 tokens each is ~4500 tokens spent before any work starts. Load modern-csharp at session start if relevant; pull ef-core, testing, etc. the moment the topic actually arises.

Anti-patterns

Don't Skip Plan Mode for Complex Tasks

Don't Work in a Single Session When You Could Parallelize

Don't Accept the First Solution

Don't Load Everything Because the Window Is Large

Decision Guide