How we work with Claude
Our agents move fast but never decide on their own. In every project the rules, skills and task status live in markdown files; the agent reads them, follows them, and stops for approval at every critical step.
The rules are written down
Three files tell the agent what to do
The rules sit next to the code, under version control. When the team changes a line, the agent follows it on the very next step.
# Working rules
## Approval
- Show the draft and get approval before any Jira comment, status change or assignment
- Reading the cloud is fine; writing only on explicit instruction
- Production accounts: describe, get and list only
## Secrets
- Access keys and tokens are never read, written or shown
## Reporting
- Separate what was done from what was not verified; no “done” without evidence
- Show the formula and source for every cost figure
The first file the agent reads in every session. What is allowed and what waits for approval is written here.
# OPS-102 · decommission test-api
The agent ticks off every step. A step that hits a rule stops in red and goes nowhere until approval arrives.
## Steps
1. Read-only checks and the saving
2. Written approval from the owner
3. stop → terminate, confirm the disks are gone
4. Jira comment, worklog and staged closure
When the same job comes back, the steps proven by hand become a skill and run in the same order next time.
The loop
Every task goes through the same four steps
AI agents speed the work up; we set the limits.
Read
We study the process and the data read-only. Nothing is written to any system at this stage.
Prototype
With AI agents we ship a working first version quickly and try it on your real data with you.
Verify
Tests, approval steps and an evidence-based report. We never hand over an unverified result as done.
Package
We package the recurring work as a skill or agent and hand it to your team with documentation.
Principles
Your data stays with you
Personal products run on-device; encryption and backups stay with the user.
No writes without approval
Deleting, shutting down or changing records happens only with explicit approval.
Every number has a source
Each cost calculation shows its formula and source, and keeps estimates apart from actuals.
Tools
We work with Claude Code, ChatGPT and Codex. AGENTS.md and SKILL.md files are tool-agnostic, so the same setup works with whichever AI tool your team prefers.
A free 30-minute assessment
Let’s look at your cloud costs and inventory together: where money is wasted, what the inventory is missing, and where automation should start.