160 checks
Checklists
The work that gets skipped is the work nobody wrote down. Each of these is a list of things you can verify — not things you can agree with — with the reason each one is there and the tool that checks it.
Ticks persist in your browser. No account, nothing sent anywhere.
Shipping
- LLM Production Readiness ChecklistTwenty-eight verifiable checks to run before an LLM feature reaches real users: failure handling, cost ceilings, evaluation, privacy and rollback.
- RAG System Launch ChecklistChunking, embeddings, retrieval quality, grounding and cost: the checks that separate a RAG demo that impresses from a RAG system that answers correctly.
- Model Migration ChecklistWhat changes when you switch model or provider: tokenizer, context, prompt format, tool calling, stop reasons, refusal behaviour and cost — with a rollback plan.
Cost
Quality
- Prompt Review ChecklistA reviewable checklist for prompts, the way a code review checklist works: structure, ambiguity, examples, output contract, failure paths and cost.
- LLM Evaluation Setup ChecklistHow to build an evaluation set that catches regressions: case selection, expected outputs, graders, statistical significance and keeping it honest over time.
Safety
- AI Agent Safety ChecklistPermissions, loop limits, spend ceilings, irreversible actions and audit: what to verify before an autonomous agent can act on real systems.
- LLM Data Privacy ChecklistWhat to verify before user data reaches a model provider: legal basis, minimisation, retention, training opt-out, subprocessors, transfers and deletion.
Why a checklist and not a guide
A guide is read once, by someone learning. A checklist is opened every time the same work happens again — before a launch, during a review, at the point where somebody has to say yes.
That is why every item here is written as a statement you can be wrong about. “Consider rate limiting” cannot be ticked honestly. “A single user cannot exceed a defined request rate” can only be ticked if someone goes and looks.
Where to start
If something is going live soon, start with the production readiness list and do the six items marked critical. Those six prevent most of what goes wrong in a first month of real traffic.
If the bill is the problem, the cost control list is ordered by saving per hour of effort — the first five items are configuration changes rather than rewrites.
If a model is about to get tools and credentials, the agent safety list is the one to do before it has write access rather than after.
Using these in a team
The two useful habits: paste the outstanding items into the pull request that ships the feature, and add a line to the list every time something goes wrong that the list would not have caught. A checklist that never grows is a checklist that stopped describing your system.
Every page has a “copy what is left” button that produces markdown checkboxes for exactly that.