AI in project work

AI Safety

AI tools can significantly accelerate software delivery. In customer projects, speed is not enough: context, access, traceability, and human responsibility need clear boundaries.

This rule set is based on real project work: situations where AI tools are useful, but need clear technical and organisational boundaries.

Structure

13 safety aspects for AI-assisted development

The topics are formulated as working rules: short enough for everyday use, concrete enough for reviews, onboarding, and project decisions.

01

Data and AI Context

Which project files, data classes, and local contents must not enter AI context without review.

Wall Rules

  • .gitignore protects Git. .aiignore protects AI context. They are not the same.
Open topic
02

Rights, Access, and Command Execution

How AI assistants may work locally without receiving blanket access to projects and systems.

Wall Rules

  • AI does not execute commands directly. Commands go through a project gate.
  • New commands are blocked. Approval is project-specific and purpose-bound.
Open topic
03

Third Parties, Updates, and Supply Chain

Why dependencies, lockfiles, CI actions, and images should be reviewed deliberately instead of updated on the side.

Wall Rules

  • Release builds install from lockfiles, not from hope.
  • CI actions are executable third-party code.
  • No latest images in production-adjacent builds.
  • New dependencies with install scripts are executable third-party code.
  • Composer updates are code changes, not maintenance noise.
  • Less third-party code, less attack surface.
  • Dependency updates are separate changes. Do not merge them on the side.
  • Every new dependency is new third-party code.
Open topic
04

Deployment and CI/CD

How merge, review, and automated deployments can be treated as production approval.

Wall Rules

  • Deployment config is security code.
  • Deployment code is security code.
  • AI may prepare deployment code, but not execute it.
  • Merge into main is production approval.
  • Code rollback does not replace database rollback.
Open topic
05

Local LLMs

Why locally operated models still raise access, network, logging, and operations questions.

Wall Rules

  • Local does not automatically mean safe.
  • Local model, still a service with rights.
  • No silent AI write actions.
Open topic
06

Agents with Tools

Which boundaries should apply when agents edit files, execute commands, or use tools.

Wall Rules

  • Commit and push are human approvals.
  • Reading may be automatic. Effects require approval.
  • New agent capability = new tool.
  • Agents work in stages, not blind runs.
  • What the agent reads is not automatically an instruction.
  • An agent is an actor with tools.
Open topic
07

AI Rule Frameworks

Why project instructions, skills, and public rule packs should be treated and reviewed like code.

Wall Rules

  • AI rules are code: review, limit, version.
  • External rule packs are an idea source, not an installation package.
  • Local convenience rules do not override project safety.
  • Rule changes are security changes.
  • AI rules belong in review, not production.
Open topic
08

AI Code and Reviews

How AI-generated or AI-assisted code should be checked, understood, and owned.

Wall Rules

  • AI code is own code: understand, check, test, own.
  • Unreviewed code is taboo.
  • Green AI tests are not a free pass.
  • AI review supports, people approve.
  • If AI cannot explain the diff correctly, check it twice.
Open topic
09

Teamwork and Governance

How teams can steer AI use visibly, lightly, and consistently.

Wall Rules

  • One team, one rule set: AI use must be visible and coordinated.
  • Few rules, consistently applied.
  • If it is not documented, it is not a team rule.
  • External contributors follow the same AI rule set as the team.
  • Safety does not happen on the side: AI rules need maintenance time.
Open topic
10

Incident Response

What to do immediately when wrong agent actions, data exposure, or supply-chain issues are suspected.

Wall Rules

  • Stop, secure, report, do not continue experimenting.
  • Do not clean up before preserving state.
  • Rollback must not be invented during the incident.
  • A secret in AI context means: assume compromise first.
  • Contain technically first, then escalate wider.
Open topic
11

Vendor Review and Ready-Made AI Solutions

Which criteria matter for AI vendors, plugins, local stacks, MCP servers, and ready-made tools.

Wall Rules

  • Security patches quickly, new capabilities controlled.
  • Latest is not a production strategy.
  • Experiments yes, but not on customer project machines.
  • Large vendors are easier to review, small tools more restrictive.
  • Vendor claims are not controls.
Open topic
12

Customer Communication

How AI use can be explained to customers calmly, concretely, and without overpromising.

Wall Rules

  • Explain AI use calmly, do not hide it.
  • Only promise what we control.
  • No customer data in AI without approval.
  • No AI must be defined concretely.
  • Clarify facts first, then inform customers.
Open topic
13

Law and Compliance

Which technical triggers should lead to legal, privacy, or customer clarification.

Wall Rules

  • Do not prompt away legal questions: clarify, document, escalate.
  • Personal data makes AI use a privacy topic.
  • AI in the product is compliance, not just engineering.
  • Confidentiality applies in prompts too.
  • Secrets remain taboo.
Open topic