Small engineering teams do not fail for lack of skill. They fail when the person who knew why a deploy step exists is busy on another project. We run government systems, our own products, 17+ live websites and client work with a handful of people. We could not afford that failure, so we wrote the context down and made Claude Code read it first.
The four layers
| Layer | What it is | What it prevents |
|---|---|---|
| Project memory | A CLAUDE.md in each repo with the stack, rules and brand constraints | A session guessing the stack or breaking a house rule |
| Handoff | A HANDOFF.md per project: live state, open items, last two sessions | Starting work on stale assumptions |
| Landmines | A LANDMINES.md per project that is never trimmed: every trap that once cost us | Repeating a mistake we already paid for |
| Skills | 37 custom skills for recurring work: QA, deploys, ads, SEO, proposals, outreach | Doing the same job differently every time |
A session starts by loading the handoff for the folder it is in. Before it touches infrastructure, it reads that project's landmines. Recurring work runs through a skill, so it follows the same steps every time.
Guard-rails written as rules, not hopes
Some rules exist because something went wrong once:
- Pre-flight before high-risk actions. Deploys, deletes, ad changes and client deliverables run a written checklist first. The result is shown before anything proceeds.
- Deploy last. All pending work is finished and verified on localhost before a single deploy. One evening of five separate production restarts taught us this.
- Live means committed. Anything that reaches production is committed and pushed in the same session. A clean checkout can then never silently revert the live site.
- Two-step delete. Every delete lists exactly what will be removed and asks twice.
- Never claim done unverified. "Verified" and "untested" are reported separately, with the output.
What it runs today
- Our own SaaS products, including a task manager with a WhatsApp bot and a pharma validation-document tool.
- AI Jyotish, a Hindi astrology assistant on Claude Opus 5.5 with a tested evaluation set.
- A cold-outreach engine, a content engine for social posts, and an AI SEO audit engine.
- 17+ live websites, from corporate sites to news portals.
What we set up for other dev teams
The same pattern works for any software team. We set up a CLAUDE.md for each repo and skills for the team's stack and coding standards. We add safe defaults for secrets and deploys, and run a hands-on session with the engineers. The goal is not more AI usage. The goal is that every engineer, and every Claude Code session, starts from the same written context.