Claude Code 2.1.143 team conventions

During Claude Code training retros, we noticed verification commands living in memory instead of the PR description. Training sticks when examples match the toolchain and every permission decision leaves a receipt.

Permission Literacy

Another prompt template will not replace ownership. The expensive bug is permission drift nobody signed.

We believed local greens equaled team-wide confidence. We rehearsed that playbook with teams whose durable contracts were still mushy.

John Ousterhout’s complexity vocabulary fits: shallow prompts hide deep coordination.

Claude Code trust scales through permission literacy.

Tool vendors move fast; your obligation is to keep receipts aligned with how your org actually reviews.

Team Convention Pattern

Permission autopilot. If approvals become reflexive, risky bash slips through. The failure is not usually tool quality; it is the missing operating contract.

Use Red-folder ritual. List paths requiring separate reviewer + never auto-approve.

What changes: Sensitive edits regain deliberate eyes. Claude Code trust scales through permission literacy.

Keep that as the first guardrail.

MCP data surprises. You can spot this when MCP pulls customer data, you have seen shock during audit.

Why it keeps recurring: Context feels local while data leaves the laptop boundary.

The practical fix is Data-class tags: Tag connectors with data classes + retention expectations.

The evidence is mundane and useful: security reviews stop learning basics during incidents. Claude Code trust scales through permission literacy.

It turns agent output back into team-owned work.

Session amnesia. If sessions rotate daily, teams lose shared memory.

The trap is Institutional knowledge cannot live only in transcripts.

Named fix: Weekly retro note. Append one .md changelog of decisions agents relied on.

After the fix, new teammates inherit reasoning instead of rumor. Claude Code trust scales through permission literacy.

Keep that as the first guardrail.

Skill drift. You can spot this when skills proliferate, reviewers cannot tell which one governed a diff.

Why it keeps recurring: Discovery beats documentation unless precedence is explicit.

The practical fix is Skill index: Maintain skills/README.md listing activation cues + deferrals.

The evidence is mundane and useful: on-call engineers resolve mismatches without replaying sessions. Claude Code trust scales through permission literacy.

It turns agent output back into team-owned work.

# CLAUDE.md — supremacy fragment

- Hooks win over informal chat agreements; document each hook’s rollback path.
- Skills defer to this file on security-sensitive folders.
- Bash approvals never bypass the red-folder list maintained here.

We map this discipline to our methodology under Review—evidence beats narration when merges touch shared surfaces.

Think of agents as signal amplifiers—they multiply whatever clarity already exists in files, hooks, and scopes.

Reviewer Artifact

PR proof

• MCP connectors mentioned (if any) list owners.
• Verification command output is pasted or linked.
• Forked agent work lists parent + child responsibilities.
• Red-folder paths received explicit human acknowledgement.

Boundary Conditions

None of this replaces architecture judgement—agents accelerate execution, not ownership.

Source Stack

• Claude Code — permissions
• Model Context Protocol — specification
• Anthropic — Claude Code product page
• GitHub — anthropics/skills
• Claude Code — getting started
• Claude Code — skills

Next Step

If you want one concrete habit change this week, start from Claude Code team conventions and paste your receipt template into the PR body before asking for review.