What happened
The GitHub repository describes product-engineering as a collection of 45 skills for product managers, founders, designers and engineers building products with coding agents. It is free under the MIT License. The skills are designed for Claude Code, Codex, Cursor, Gemini CLI, OpenCode, GitHub Copilot and other agents that read SKILL.md folders. The six packs cover product decisions, AI shipping, agent teams, agent harnesses, application stack and work humanization. The README says 28 skills were learned through work on three live products; 14 use established methods, while others draw on vendor documentation or testing. The author’s examples include AskOne, Staygreen and AFK Pilot.
The process starts by asking an agent to identify the current state of the work and the next step. For proposals, the skills call for research into alternatives and the target segment, then a single artifact containing sourced claims, key screens and open decisions. Complex changes can receive a plan with assumptions, phases, files, tests and acceptance criteria. Small changes that can be checked quickly can skip the planning stage and go directly to implementation and checking.
The build workflow separates creation from review. One session can coordinate and another build; the session that wrote the code does not review its own work. A fresh reviewer examines each completed phase and each push that deploys, and repeats review after fixes. The described release gate keeps checking until the change introduces nothing above Low severity and there are no High or Critical findings of any age. Before release, the process also proposes deeper checks of documentation, tests, security and performance. A person chooses which checks to run, clicks through the product and explicitly approves the release; silence is not treated as approval.
Installation varies by agent. Claude Code users can install packs for a project or for themselves, and the README recommends enabling marketplace auto-updates. Codex uses its own plugin installation and requires a new session after setup or marketplace updates. Other compatible agents can use the skills CLI or copied skill folders. The repository also documents update caveats, including checking worktree records after project-scoped Claude Code updates. It reports that descriptions for the 41 skills that start automatically total about 6,200 characters—about 2,700 tokens in Claude Code—and that a skill’s full text loads only when it runs. The repository provides a walkthrough, case studies and instructions for applying the skills to existing code.
What the archive adds
New Runtime’s earlier coverage includes ScrapeCreators’ 13 open-source research skills (2026-10-22) and Theo Browne’s nine-skill tstack collection (2026-10-26). This is a new release in a recurring category, but with a broader product-engineering workflow and explicit review and release practices. An earlier note also covered Blender workflow skills; that is a related domain-specific collection rather than the same event.
Trend signal
This release supports skills becoming a portable capability layer: the repository targets several agent environments and offers skills that can be installed individually or as packs. Its use of sourced claims, fresh review and explicit release gates also makes provenance and execution safeguards part of the workflow, rather than assuming that portable instructions are safe by default. Portability still depends on each agent loading the skill correctly and on teams retaining control of permissions and review.
Why it matters
A skill collection can make product work more repeatable, but the more useful lesson here is how the workflow assigns responsibility. The agent gathers evidence and proposes; a person decides what to build and approves release; a separate session checks the implementation. That structure can help teams use coding agents without expecting a non-engineer to inspect every line or trusting the generating session to certify itself.
Teams should treat the repository as a menu and a documented method, not as proof that following its steps guarantees a safe product. Its instructions include concrete points worth evaluating locally: keep research and decisions together, test fixes with a previously failing test, require independent review, and make release approval explicit. Those checkpoints can be adopted even if a team uses a different stack or writes its own skills.
Limits and uncertainty
The repository’s account of three live products and 28 practice-derived skills is the author’s description; it is not an independent evaluation of product outcomes or of the collection’s effectiveness. The page describes the intended review threshold and workflow, but does not establish how consistently every user will apply them or whether they prevent defects in practice. Skills depend on agent-specific installation and loading behavior, and the repository notes that platform facts change. A team should verify current commands and inspect permissions, scripts and instructions before using a skill in a production project.
Skill to build now
Choose one low-risk change in an existing product and try the collection’s separation of roles: ask one agent to make the change, then use a fresh session that did not see the build to review it. Give the reviewer the acceptance criteria and require it to identify tests and any release-blocking findings. Keep the final decision with a human, and compare the result with your team’s normal review process before expanding the experiment.
Reader takeaway
The release is useful both as 45 reusable skills and as a workflow for keeping product decisions, implementation, independent review and release approval distinct. Start with one skill and one explicit checkpoint; adopt the full process only after checking how it fits your tools, permissions and team responsibilities.