Boris Cherny's Claude Code workflow, broken down

Most developers use Claude Code like a souped-up autocomplete. The boris cherny claude code workflow doesn't look like that at all. Cherny, who built the tool at Anthropic, made public an approach built around three habits missing from most setups: maintaining multiple parallel sessions on different surfaces, linking each repeated daily operation to a slash command that automatically triggers a sub-agent, and deliberately picking the model based on how much brain power the problem needs. This article breaks down each practice from his public GitHub gist and the InfoQ profile published January 10, 2026. We'll separate what you can adopt today without restructuring your environment from what requires deeper changes.

The Boris Cherny Claude Code workflow in one sentence

Put simply: the boris cherny claude code workflow treats Claude Code not as a chat but as a system of agents to orchestrate, where repeated operations are automated and powerful models reserved for tasks that actually need them.

His public gist, "boris-team-tips.md", lists ten distinct techniques. Each one targets a specific friction point in an agentified workday: how to open a session, when to upgrade models, how to name a command so it stays memorable, how to split a task into parallel sub-problems. These aren't decorative tricks. Each technique cuts a precise cognitive cost.

Three axes structure the whole thing. Parallelization first: multiple contexts open simultaneously, each isolated on a distinct task. Command automation next: anything that comes back more than once in a day becomes a slash command, not text to reformulate every time. Model selection finally: Opus with extended thinking for hard decisions, lighter models for routine tasks.

The InfoQ profile from January 10, 2026 confirms this approach has been his default work mode since Q4 2025, not a one-off experiment. This status shift (advanced feature to daily practice) is the most useful temporal marker for calibrating your own adoption.

Slash commands as one-keystroke workflows

A slash command in Claude Code isn't a text shortcut. It's a prompt template paired with a sub-agent trigger. When you type /commit, you're not pasting text into a window: you're launching an agent configured to read the current diff, understand the change context, and produce a structured commit message following your conventions.

Cherny maps each repeated daily operation to this mechanism. Generating commit messages, writing PR descriptions, code simplification passes, post-implementation checks: each has its command. The January 2026 InfoQ profile quantifies between five and ten invocations of this type per day, depending on active session volume.

The concrete mechanics: a slash command (/cmd) corresponds to a prompt file stored in .claude/commands/. This file specifies the instruction, automatically injected context scope, and any files to include. The sub-agent reads this context in isolation, produces a result, and returns it without polluting the main session. The official Claude Code documentation details the format of these command files.

This pattern changes two things in your developer productivity setup. First, it reduces prompt quality variance: you reuse a tested and refined instruction rather than formulating a new one each time. Second, it reduces mental load: typing /simplify costs less attention than writing "look at this code and suggest a simplification considering current performance constraints".

To adopt this pattern with the lowest barrier, start with commit generation. Create .claude/commands/commit.md with a precise instruction about your convention (Conventional Commits, mandatory scope, non-vague messages), and associate it with /commit. It's a ten-minute investment for a measurable daily gain on your slash command automation.

Running multiple sessions in parallel

The first recommendation in the public gist is explicit: maintain multiple Claude Code sessions open simultaneously. Not cascaded (one session waiting for the previous one), but parallel with distinct scopes.

In his documented agentic coding workflow, Cherny uses at least three surfaces: local terminal, web interface, and mobile. The local terminal handles active code. The web interface processes research or non-code content generation. Mobile covers light reviews and context questions while moving around.

Two constraints structure these parallel AI sessions. First constraint: context isolation. Two sessions reading the same file pose no problem, but two sessions writing to the same branch without coordination create conflicts. The documented rule is one session per branch, or one session per task type (implementation, tests, review). Second constraint: token budget. Each session consumes independently. Three simultaneous sessions represent three separate budgets, which you need to anticipate in your AI coding assistant setup if you work under a strict ceiling.

The gain is concrete: while one session generates tests for a module, another can work on documentation for a different function, and a third can review an open PR. The development day compresses without degrading the quality of each individual task.

Model selection: Opus with thinking vs. lighter models

Cherny applies an explicit model selection rule. Opus with extended thinking is required for multi-constraint problems and ambiguous architectural decisions. Lighter models cover routine edits, boilerplate, and verification passes.

The distinction isn't "difficult versus easy". It's more precise: Opus with extended thinking brings a measurable gain when the problem requires holding multiple contradictory constraints in memory simultaneously. A refactoring that preserves three behavioral invariants while changing data structure, or an architecture decision that arbitrates between five performance and maintainability trade-offs: these are cases where thinking budget changes final answer quality.

For routine edits (renaming a variable, adding a case to a switch, generating a unit test for a well-defined function), a lighter model produces the same result faster and cheaper. Mobilizing Opus to write a getter test is poorly allocated budget.

In 2026, Claude nomenclature clarifies this choice: Opus 4.7 for deep reasoning, Sonnet 4.6 for throughput. These identifiers were formalized in the 2025-2026 period, and extended thinking is available on Opus with a dedicated token budget for internal reasoning before the visible response.

The operational rule: identify three problem types that merit Opus in your context (architecture decisions, non-deterministic behavior debugging, security reviews). Everything else goes to Sonnet or equivalent. This explicit classification avoids the temptation to upgrade models by default on all sessions.

What evolved in his approach in 2026

The January 10, 2026 InfoQ article provides a precise temporal marker: on that date, Boris Cherny presents his workflow as a codified system of daily practices, not a list of experimental tricks. This status shift is structural.

Before Q4 2025, sub-agent orchestration was documented as an advanced feature, reserved for users capable of manually managing agent coordination. What the January 2026 InfoQ profile documents in the boris cherny claude code workflow is the shift of this pattern to default mode. Every repeated operation is a command. Every session is scoped. Every complex problem is decomposed before being delegated.

The most structural change concerns slash commands. In the initial gist (pre-2026), they were presented as optional optimizations for advanced users. In the January 2026 InfoQ version, they're the central mechanism. You no longer query Claude Code ad hoc. You invoke it via named commands whose behavior is predictable and reproducible.

This shift has a practical implication if you're evaluating what level of setup investment justifies a serious agentic coding workflow. Cherny's 2026 answer is clear: commands and parallelization aren't optimizations for long projects. They're the base configuration for any daily work with Claude Code.

A second evolution concerns session surfaces. In 2025 setup notes, mobile was mentioned as a marginal option. In 2026, it's listed among the three standard surfaces with documented usage (light reviews, context questions while traveling). This signals the workflow was adjusted based on real usage, not theoretical design.

Key takeaways

The boris cherny claude code workflow distinguishes three adoption levels. Slash commands for commits and PR descriptions have the lowest barrier: one config file, ten minutes, immediate daily gain. Parallel AI sessions adopt in one day with a clear scope rule. Complete sub-agent orchestration, with systematic model selection and three simultaneous surfaces, requires restructuring, but that's the level documented as standard in 2026.

---

Cherny's slash commands aren't text shortcuts, they're sub-agents with preset context. The CLI blueprint in the welcome kit shows how to structure tools the same way, so Claude can chain them without you reformulating prompts every time.

→ Get the welcome kit