Agent skills that bring team coding standards to Claude Code and Codex

I'll pull the GitHub repo and HN thread so the paragraphs stick to real project details, not invented numbers or claims.Tikal published adlc-team-skills, an open set of agent skills that load a team’s coding standards into Claude Code, Codex, and other agents that implement the Agent Skills standard. On Hacker News the post drew 72 points and 39 comments. The pitch is narrow: coding agents start each session without team context, so people patch that gap with personal CLAUDE.md files that live on one machine, drift, and do not transfer across teammates or tools.

Mechanically, team-boot runs at session start and injects only an index of rules, personas, and decisions—names and one-line descriptors, roughly a hundred tokens—not the full rule text. When a task matches a rule, the agent pulls that full text on demand; a SQL task loads the SQL rule and nothing else. That index lives in a separate git repo (team-ai-directives), versioned and reviewed by PR. mission-brief forces a contract first—goal, constraints, non-goals, success criteria—then runs specify → plan → tasks → implement with a converge loop. evals skills build executable suites with PromptFoo or DeepEval, preferring binary Tier 1 graders over LLM judges. levelup-specify captures session fixes as Context Directive Records (CDRs) and commits them back so the next session starts with the same knowledge. product-* and architect-* skills capture PDRs and ADRs and compile them into PRD.md and AD.md. team-repair --build-to-delete re-runs evals without a rule and proposes deletion if the model still passes. Install is via npx skills add (or adlc-skills-cli for slash commands and session_start hooks). The design cites long-context degradation research (arXiv:2510.05381) as the reason for progressive disclosure rather than stuffing more rules into the prompt.

For engineers, this treats team standards as shared, versioned infrastructure instead of private prompt folklore. Conventions, deprecations, and “service X never calls service Y” become queryable artifacts with a publish path, not tribal knowledge that evaporates when a chat ends. The evals path matters for anyone who has watched agents invent APIs or skip constraints: checkable claims get code graders; nothing auto-merges—the pipeline ends at a human-reviewed PR. Build-to-delete is the counterweight to rule sprawl as base models improve and yesterday’s scaffolding becomes context noise.

In market terms, this sits on the Agent Skills standard and targets Claude Code, Codex, OpenCode, Cursor, Copilot, and peers rather than locking to one IDE. mission-brief scans installed skill directories and lets the model pick from complementary packs—mattpocock/skills, addyosmani/agent-skills, superpowers, spec-kit style frameworks—so it competes less as a closed stack and more as team-context and lifecycle skills that compose with others. Adjacent work still centers on personal skill packs and ad hoc markdown rules; Tikal’s differentiator is the shared directives repo plus CDR/PDR/ADR/evals lifecycles under a Twelve-Factor Agentic SDLC framing.

Practical takeaway: try team-boot and a small team-ai-directives repo before adopting the full product/architecture/evals surface—the progressive-disclosure index is the load-bearing idea. Watch whether teams actually PR and retire CDRs (especially via --build-to-delete), and whether multi-agent shops can keep one directives repo coherent across Claude Code and Codex. Also treat clones carefully: on 2026-07-27 a supply-chain worm briefly hit .claude/ and .vscode/ via a stolen maintainer token; the npx skills path did not ship those files, history was rewritten, and the stated lesson is to treat editor auto-run config as executable code.