Works at Black Airplane

A lot happens in 6 months in AI
There's a new skill in the layer you build around out of the box AI agent capabilities
Two ways to fail
- Read every line and you don't scale
- Read nothing and you skip dark code
Agentic engineering is about how much you can responsibly own.
Engineer your own harness
- Prompting
- Context
- Memory
- Tools
Out of the box, we are all equal
Prompting - You are already good at this (or you should be)
Context - The working set right now
- Always-loaded is things like AGENTS.md, CLAUDE.md
- On-demand - retrieved - conventions docs, nested rules files, specs, skills
- Context management is about breaking things down so the agent loads only what's relevant to the task at hand
- AGENTS.md
  - What the project is
  - Stack
  - Repo Layout
  - Harness layer taxonomy - the data model, referencing skills, etc. helps improve usage
  - Diagnostic codes
  - Security rules - Non-negotiable
  - Conventions
  - quality gates
  - Spec workflow
  - Roadmap
  - Out of scope
  - Working agreement - agent behavior
- Nested rule files is a solid rule approach. Can also do Progressive Disclosure, Made of Files (I've done this too! This is the same idea as what I did with my testing documentation). Patterns and Features are big ones
- Spec driven development - research-lead, integration-researcher, codebase-explorer
  - Spec is the artifact you understand. Implement generates code against the spec. You comprehend the spec. The code, you verify. Becomes part of the memory system
Memory
- .review
- specs - spec-memory.db (new idea)
- state.json (equivalent of my TestPlans files)
Tools
- Claude-qa
  - test-strategist, testgenerator, qa-auditor, regression-tracker, mutation-tester
- playwright-skill that's a 200 token summary
When in doubt, use a single agent first, and when you become the bottleneck, use multi agents
Not all AI generated code deserves equal attention

Stop Using One AI Tool For Everything - A multitool approach to frontend development by Will Marple