Your Coding Standards Are Becoming Agent Skills

A consultancy just tried to encode an entire software development lifecycle as agent skills — 28 of them, covering everything from team constitutions to eval pipelines — and the Hacker News response was a collective wince. Both halves of that sentence matter. Tikal's ADLC ("Agentic SDLC") repo is an early, overbuilt attempt at something that is genuinely coming: your team's coding standards living in version control as portable skills, loaded by whatever agent your teammates happen to run.

The format won before the frameworks did

The load-bearing fact here isn't Tikal's repo. It's that a skill is now a commodity artifact. Agent Skills — a folder with a SKILL.md file carrying name and description frontmatter, instructions, and optional scripts and reference files — started as an Anthropic feature in Claude and Claude Code last October. Anthropic then released the format as an open standard, and the adopter list has since become absurd: OpenAI's Codex, GitHub Copilot and VS Code, Cursor, Gemini CLI, Goose, OpenCode, Amp, JetBrains' Junie, and dozens more.

Codex's implementation is instructive because OpenAI didn't invent a rival format — it adopted its competitor's. Codex discovers skills from .agents/skills in your repo, ~/.agents/skills for personal ones, and /etc/codex/skills for fleet-wide defaults; you invoke one explicitly with $skill-name or let the model pick it up from its description. Claude Code does the same dance with .claude/skills and /slash-command invocation. Same folder, same frontmatter, both ends of the duopoly. That's the actual news environment Tikal's repo landed in: for the first time, a team can write its standards once and have them travel across vendors — something none of the earlier config dialects (CLAUDE.md, .cursorrules, AGENTS.md) managed.

The mechanism that makes this viable at team scale is progressive disclosure. At session start, an agent loads only each skill's name and description — Codex budgets roughly 8,000 characters for the whole metadata index — and reads the full SKILL.md only when a task matches. A team can keep fifty skills on the shelf and pay context for two.

What Tikal actually shipped

The ADLC repo is a maximalist bet on that mechanism. Its 28-plus skills are organized into lifecycle families — team-* for session boot and "team constitution" setup, product-* and architect-* for spec-driven feature work, evals-* for verification with PromptFoo or DeepEval, plus a mission-brief orchestrator that routes across skill repos from other authors. Skills write structured decision records (ADRs, PDRs, and Tikal's own "CDRs") into a .adlc/ directory, the idea being that standards, architecture decisions, and extracted patterns become reviewable git artifacts rather than folklore in someone's prompt stash. Installation is one command via the community skills CLI (npx skills add tikalk/adlc-team-skills -a claude -g), with a wrapper CLI that also wires up slash commands and session-start hooks for nine different agents.

Some of the repo's slogans are genuinely good engineering instincts. "Separate the maker from the checker" — never let the agent that wrote the code verify it — is advice most solo Claude Code users would benefit from. "Build to delete," pruning rules as models improve, is the correct posture toward any prompt artifact. And the core diagnosis is right: individual developers hoarding private prompt tricks produces inconsistent codebases and unreviewable AI-generated PRs.

The objection that's half right

HN's pushback clustered on two points, and they deserve to be separated. The first — one commenter likened the framework to "dumping the Clean Code book into context," predicting 100K tokens of rules the agent will ignore — is half wrong on cost. Progressive disclosure exists precisely so that 28 skills don't ride along on every request; only the metadata does. If Tikal's descriptions are tight, the standing tax is small.

But it's half right on the deeper issue: skills are suggestions, not enforcement. A SKILL.md is prose a probabilistic model may or may not follow, and adherence degrades as instruction volume grows — several commenters noted that stuffed context makes agents forget guidelines, not follow them. Another called the result "more of a vibe guideline." That's the honest framing. Nothing in the skills mechanism verifies that the agent applied your naming conventions or ran your checklist. If a standard matters, its enforcement still belongs in deterministic tooling — linters, CI gates, pre-commit hooks, PR checks — and the skill's job is to get the agent to write code that passes those gates on the first try instead of the third. Skills reduce iteration count; they don't replace verification. Tikal's own "maker vs. checker" pillar concedes this, which makes the framework's sheer surface area — a Twelve-Factor methodology, four pillars, three bespoke record formats — feel like process arriving before evidence.

What to actually do with this

If you lead a team using Claude Code or Codex, the move is smaller than Tikal's and worth making this quarter. Pick the three standards that cause the most PR churn — say, your migration conventions, your error-handling policy, your test structure — and write each as one narrow skill: a folder, a SKILL.md with a precise trigger description, maybe a script or a reference file. Commit them to the repo (.claude/skills/ and .agents/skills/ can coexist, or symlink one to the other), and let discovery do the rest. You'll know within a week whether agents pick them up at the right moments; a skill that never triggers has a bad description, not a bad idea.

Skip the governance superstructure until you've felt the need. The record-keeping instinct — decisions as reviewed git artifacts — is sound, but plain ADRs in a docs/ folder that a skill knows how to read gets you most of it without adopting anyone's methodology wholesale.

The direction, though, is settled. Team standards are leaving wikis and lint-adjacent prose and becoming executable-ish artifacts in a format every major agent reads. Tikal got the destination right and the vehicle overweight. The teams that win this transition will be the ones shipping five boring, tightly-scoped skills — not twenty-eight ambitious ones.