Using Claude Wisely: Context Engineering for Real Engineering Work

A large language model is a context engine: the quality of what comes out is bounded by the quality of what you put in. Most bad AI output is not a model failure, it is a context failure. Here is how we get real leverage from Claude on production engineering work.

Context engineering beats prompt tricks

Clever phrasing helps at the margins. Giving the model the actual map helps enormously. Point it at the real files, paste the real error, and state the real constraints instead of describing them.

• Share the failing test output verbatim, not a paraphrase

• Name the files and functions in play, with paths

• State the constraints up front: runtime, versions, what must not change

• Give one good example of the pattern you want followed

Make the project legible

A short project memory file pays for itself on every task. Keep it factual and current, it is the brief every session starts from.

# Project: payments-api

## Stack
- Go 1.22, chi router, sqlc, Postgres 16
- Deployed on EKS via Argo Rollouts (canary)

## Conventions
- Errors wrapped with %w; no naked panics in handlers
- Every endpoint has a table-driven test
- Never log PII; use the redact() helper

## Commands
- test: make test
- lint: golangci-lint run
- build: make build

A workflow that scales

1. Ask for a short plan before any code, and read it critically

2. Work in small diffs you can actually review

3. Run the tests and the linter after every change

4. Review the final diff yourself, you own what you merge

Treat the model like a fast, tireless pair, not an oracle. Ask it to show the diff and to explain the risky parts, then verify:

make test && golangci-lint run
git --no-pager diff --stat
git --no-pager diff # read it before you stage

When not to reach for it

• Novel architecture decisions with lasting blast radius

• Anything touching secrets, auth, or money without human sign-off

• Problems you do not yet understand well enough to check the answer

• Compliance-sensitive wording that needs a real owner

Never delegate understanding. Use AI to move faster through problems you can still verify yourself.

Great context plus tight verification turns Claude into a genuine force multiplier. Skip either half and you are just generating plausible text and hoping.