This open-source layer gives coding agents team rules

AI coding assistants can write code quickly, but the repository behind ADLC Team Skills argues that speed is no longer the hardest engineering problem. Its focus is shared context: putting a team’s principles, product decisions, architecture rules, workflows, and evaluation criteria into version-controlled files that coding agents can use consistently.

The open-source project, tikalk/adlc-team-skills, is designed for Claude Code, Codex, OpenCode, Cursor, GitHub Copilot, and other tools supporting the Agent Skills standard. It is positioned as the team layer of the project’s “Twelve-Factor Agentic SDLC,” intended to replace isolated prompt customizations with a common engineering framework.

That approach fits into the broader shift toward reusable coding-agent instructions, including Claude’s reusable screen skills and other community skill repositories.

Installing the team layer

The project provides two installation paths. Its own CLI can install skills, generate slash commands, and connect session-start events:

''bash npx adlc-skills-cli add tikalk/adlc-team-skills -a opencode ''

Recommended reading

Cloud giants plan $595 billion for AI infrastructure

For skills alone, users can instead run:

''bash npx skills add tikalk/adlc-team-skills -a claude -g ''

The adlc-skills-cli wrapper supports slash commands and event hooks through .events.json. According to the repository, it wires session-start events for nine coding agents; repositories without an event file receive commands but not event hooks.

On an unconfigured project, the team-boot skill runs at session start and warns the user to run /team-setup. Setup can scaffold a new team-ai-directives repository or clone an existing one. The generated structure includes README.md, AGENTS.md, CDR.md, .skills.json, a constitution placeholder, and OKF index files, followed by a Git repository initialization.

Once configured, team-boot assembles the team constitution, Context Directive Record index, product and architecture decision indexes, and skills registry into the agent’s system prompt. The goal is to make team context available before an agent starts working rather than relying on each developer to supply it manually.

Four pillars for agentic development

ADLC organizes its skills around four areas:

• Strategy and team directives: team-boot, team-discover, team-constitution, and team-repair manage shared principles and context modules.
• Product and architecture governance: Product Decision Records become a unified PRD.md, while Architectural Decision Records are assembled into AD.md using Functional, Security, Deployment, and Performance viewpoints.
• Spec-driven execution: mission-brief turns a feature request into a contract containing goals, constraints, non-goals, and success criteria, then runs a specify → plan → tasks → implement → converge workflow.
• Verification and evaluation: evals skills create application-level tests with PromptFoo or DeepEval, combining fast checks with LLM judge subagents before human review.

The repository’s central rule is “Debug the Spec, Not the Code.” When an agent makes a mistake, the team is expected to add the missing constraint to the specification instead of merely patching the immediate output.

The governance layer also separates authorship from evaluation: “Never let the agent that wrote the code decide if the code is good.” The project calls this “Separate the Maker from the Checker,” with automated evaluation followed by a human “Great Filter” review.

“Speed is solved; Trust and Verification is the new bottleneck in AI engineering.”

Vendor-neutral orchestration

mission-brief is the project’s central orchestrator. At the start of a mission, it scans directories such as .claude/skills and .agents/skills, reads each SKILL.md file’s front matter, and builds an inventory of available skills.

Rather than using a fixed mapping, the system gives the full inventory to a subagent, which decides whether a skill applies to each phase. If no suitable skill is found, the subagent proceeds directly; if one matches, it invokes that skill and continues the pipeline.

The repository says this can combine ADLC skills with sources including Spec-Kit, OpenSpec, mattpocock/skills, addyosmani/agent-skills, and Superpowers. Its own example installs multiple sources before asking mission-brief to handle a request such as adding a user-profile API with JWT authentication.

The project does not provide benchmark results or an independent evaluation of whether these workflows improve code quality. Its contribution is instead a concrete, Git-based operating model for teams that want coding agents to follow shared rules rather than behave as disconnected prompt-driven assistants.