File Convention Over Data Passing
Day 19 · Week 4 · Skills — Advanced
Slack Message — copy & paste
🤖 Tip #19 — How do you pass data between AI skills in separate contexts? You don't. Use a file convention — each skill writes to a predictable location, the next knows where to look.
#19 Claude Code
The Design Pattern
- **The pattern** — Each skill writes its output to a predictable location (`.ai/specs/<ticket-id>-<slug>/`), and the next skill knows where to look. No APIs, no return values, no coupling.
- **Same principle as Unix pipes** — Small tools, predictable I/O. Skills find each other's output by convention, not by API contracts or shared memory.
- **Inspectable** — You can read any file to see exactly what happened. `raw-story.md` shows what was fetched, `explain.md` shows how it was interpreted, `implement.md` shows the plan.
- **Testable and debuggable** — Each skill is fully independent. If step 3 fails, check step 2's output file. Mock any input by writing the expected file manually.
Spec Directory Convention Decoupling
#19 Skills — Advanced
See It in Action
- 1 **Try the flow** — Run `/dx-req <id>` then look inside `.ai/specs/` — read the files to see exactly what the AI captured at each phase.
- 2 Understand the file flow
- › `/dx-req` writes: raw-story.md, explain.md, research.md
- › `/dx-plan` reads explain.md + research.md, writes: implement.md
- › `/dx-figma-extract` writes: figma-extract.md
- › Each skill is independent — test and run them in any order
- 3 **Resumable by design** — Skip a step and re-run just what you need. `/dx-plan` reads `explain.md` which is already there from `/dx-req`. No need to re-run the entire chain.
- 4 **Apply to custom skills** — When building your own skills, follow the same convention. Write output to `.ai/specs/<id>/` so other skills can chain from it.
Spec Directory Convention Decoupling
Your screenshot here Optional — add a screenshot from your own workflow