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

TL;DR Highlight

Inject a rule index at session start and load full team rules on demand

Who Should Read

A team developer whose coding agent ignores team rules because personal CLAUDE.md files differ per machine

Core Mechanics

Only a rule index goes in at session start
- team-boot injects rule names and one-line descriptors via the session_start hook.
- It is about 100 tokens, and full rule text loads on demand when a task matches.
Rules live in a shared git repo
- They sit in the team-ai-directives repo and are reviewed by PR.
- This replaces personal CLAUDE.md files stuck on one machine.
Contract before code
- factory-mission first writes goal, constraints, non-goals and success criteria.
- Then it runs specify → plan → implement ↔ converge with gates and a circuit breaker.
Session learnings become records
- team-levelup extracts hard-won fixes as CDRs (Context Directive Records) at session end and scores confidence.
- Accepted CDRs are published to the team repo as a draft PR.
Evals favor binary graders

Community Reactions

Warning about malware in the repo
- A comment said commit 74f317d added five hidden files under .claude and .vscode.
- They reportedly autorun a obfuscated payload via vscode tasks.json or a Claude session start hook to steal GitHub, AWS and K8s credentials.
- The advice was to avoid npx install and opening it in VSCode, and to rotate keys.
Opinion that more context means more forgotten rules
- One commenter preferred deterministic tools like ArchUnit and static analyzers over agent guardrails.
- Others expected more token burn, review surface and time per task.
Experience that bare sessions work better
- One commenter said recent models do well without skills, AGENTS.md or system prompts.
- They linked a 9-run comparison and disclosed they work on smol.
Experience running a similar team skills repo
- One team keeps company skills in git, with an update skill that wires them into ~/.codex.
- Another suggested moving checkable rules to tests and linters and recording the rules revision per run.

How to Apply

Split your team rules into index and bodies today
- Make a name plus one-line description index from your CLAUDE.md.
- Move rule bodies to separate files and have the agent read only the index by default.
Inspect the repo before installing
- Check .claude, .vscode hidden files and recent commit diffs before running npx.
- If anything looks suspicious, do not install and open it only in an isolated environment.
Move checkable rules into CI
- Pick rules a linter or test can verify and turn them into CI checks.
- Keep only intent and workflow in the prompt.
Test a rule against no rule
- Remove one rule and rerun the same task to compare pass results.
- Delete the rule if the model passes anyway.

Try It

copy & run

# Paste into Claude Code (simplified version of the adlc-team-skills idea)

Read my CLAUDE.md and do the following.
{{CLAUDE.md content}}

1. Build a 'name | one-line description' index of the rules, about 100 tokens total.
2. Split rule bodies into rules/{name}.md files.
3. Make the new CLAUDE.md contain only the index, plus an instruction to read a rule body only when relevant to the task.
4. Mark rules that a linter or test could verify automatically.

Next: run the same task {{task}} once without rules and once with the index approach, then compare results.

Terminology

Agent SkillsA standard format where SKILL.md files teach an agent how to do a task. Like keeping manuals in a drawer and pulling one out when needed.

CDRContext Directive Record. A unit that captures hard-won fixes from a session and publishes them to the team rules repo.

ADR/PDR/ChDRDecision records for architecture, product and changes. Like meeting minutes that record why something was decided.

circuit breakerA safety stop that cuts the loop when the agent keeps failing the same way. Like an electrical breaker tripping on overload.

TPR/TNRMetrics for how well a grader calls good things good and bad things bad.