How We Use AI in Development#

AI tools are powerful but they require discipline. These are our expectations for how engineers at Axelerant work with AI.

Human in the loop#

AI augments your capability. It does not replace your judgment. Always review AI-generated code — it’s rare not to find something to fix or improve.

Be aware of cognitive biases that come with AI assistance:

• Automation bias — over-relying on AI suggestions because they look confident
• Anchoring — difficulty thinking beyond the AI’s initial suggestion
• Framing effect — accepting the AI’s phrasing or approach at face value
• Sunk cost fallacy — reluctance to discard AI-generated code you’ve already invested time reviewing

If a solution from the AI isn’t quickly helping or you feel you’re losing control, revert the changes and try a different approach. Sometimes writing the code yourself is faster.

Session discipline#

• Use /clear between tasks. Stale conversation history wastes tokens and can confuse context.
• Separate implementation from review. If you ask Claude to build a feature and then review it in the same session, it will also review all the failed attempts and exploratory work. Instead, do a git diff and review in a separate session for objectivity.
• Use Plan Mode for complex work. For multi-file changes or architectural decisions, use Shift+Tab to enter Plan Mode before implementation. This prevents expensive rework.
• Use /effort for simple tasks. Deep reasoning on a single-file change wastes tokens. Lower the thinking budget when the task is straightforward.

Model selection#

Claude Code supports multiple models. Pick the right one for the task:

Token efficiency#

Every token costs money and context space. A few habits that help:

• Keep CLAUDE.md under 200 lines — it loads on every session
• Disconnect unused MCP servers (/mcp to manage)
• Use skills for repeated workflows — they load only when called
• Reference files by path instead of pasting content into the chat
• Use sub-agents for large tasks: Claude can spawn parallel agents instead of doing everything in one context window

Data protection#

These rules are non-negotiable:

• Use only company-approved tools and subscriptions. No free or personal accounts for work.
• Never share client code, proprietary data, or credentials with any AI tool.
• Sanitize sensitive data before sharing context, even with providers that claim data won’t be used for training. If you need to query sensitive logic, reformulate it to be generic and anonymized.
• Be transparent with clients about AI usage. Acknowledge their concerns and follow any client-specific policies.

Workflow integration#

AI works best when it’s connected to the systems we already use:

• Jira — always read the ticket (including comments) before starting work. Use the Atlassian MCP connector when available to pull ticket context directly.
• Slack — reference relevant Slack threads for decision context when discussing implementation approaches.
• Git — work on feature branches. Never push directly to main or master. Use separate sessions for PR creation and review.
• Sentry — check Sentry for error context before debugging production issues.

Open source contributions#

When contributing to open source projects, be especially thoughtful about AI-generated code. See our open source contribution guidelines for more on responsible use of automated tools in community contributions.

Sharing what you learn#

We improve our AI practices through shared experience. When you discover something that works well — a skill that saves time, a workflow pattern, a model selection insight — share it in #wg-ai-native-engineering on Slack. These learnings feed back into our project configurations, shared skills, and this handbook.