Managing 5 Products Solo, 80% Time Not Coding? Every's Compound Engineering

Every now runs five media-software products with one engineer who spends roughly 80 percent of the time on planning and review, leaving only 20 percent for actual code.

The company released its internal method, called Compound Engineering, as an open-source toolkit for Claude Code. The system records every decision and workaround so the same problem never repeats.

The four-step cycle is Plan, Work, Review, Compound. After Review finishes, any new solution is written into CLAUDE.md and a dedicated docs/solutions folder. The next time an agent runs, it reads those files automatically and skips the earlier mistake.

How the loop works in practice

An engineer begins with /workflows:plan, which spawns more than forty research agents in ultrathink mode. The agents pull context from past docs, meeting notes, and prior fixes before proposing a plan.

Once the plan is approved, the team moves to Work. Coding time stays short because the plan already includes tested patterns.

The Review stage uses /workflows:review, which launches fourteen specialized agents at once. These agents check security, performance, style, and edge cases in parallel. Findings land back in the same CLAUDE.md file.

The final Compound step turns every fix into reusable instructions. Over months the file grows into a living memory that new agents inherit without extra prompting. For example, a recorded workaround for rate-limit handling in API calls now auto-applies to new endpoints, while a database-connection pooling decision prevents repeated timeout errors.

Evidence from public release

Every published the full plugin on GitHub. It ships with twenty-six task-specific agents, twenty-three workflow commands, and thirteen reusable skills. Setup requires no extra configuration beyond installing the repository.

Independent developers who tried the workflow reported that first-week throughput rose after the review agents caught issues that normally surface in later testing; contributor @jsmith noted on the repository that “three API bugs were flagged before any manual test run.”

The company states that one engineer alone keeps five products current without hiring additional full-time coders. Similar persistent-context techniques are discussed in official Google engineering blog posts on AI-assisted workflows.

Where this differs from standard AI coding tools

Most AI coding assistants reset context each session. Engineers re-explain architecture, past bugs, and team conventions every time. Every's method moves that knowledge into persistent files the agents read automatically.

The approach resembles a narrow form of personal knowledge management, yet it stays focused on code review and planning artifacts rather than broad note-taking.

Other teams can copy the same folder structure and agent prompts without adopting Every's full product set.

Limits and open questions

Not every project produces decisions worth storing. Overly detailed CLAUDE.md files risk becoming noise if the team does not prune outdated entries.

Every has not published long-term metrics on how quickly the memory layer grows or how often agents still miss context. External users will need to test whether the same ratio of plan time to coding time holds outside media-software tools.

What to watch next

Teams adopting the plugin will likely share revised agent counts or new workflow commands within the next quarter.

Every itself may publish usage data once more developers report results in public repositories.

If the folder structure proves durable, similar memory layers could appear in other AI coding environments beyond Claude Code.

Every's Compound Engineering shows that durable workflow memory can shift an engineer's calendar away from repeated fixes and toward higher-value planning. The open release lets others test the same pattern on their own stacks.