Skip to content
tishad.net
SearchCtrl K

Lessons / Part 10. Coding agents, commands and skills

64 A tracker and written decisions before any build

Four planning files and a throwaway spike before the agent writes any feature code.

What people usually ask

Build me an app to log my lab's sample inventory.

What comes back

Files arrive one at a time, each guessing how the others connect, and the first screen works until a second person opens it.

Ask this instead

First
Write no app code yet. In the repository, create:
1. tracker.md: the features from the attached paragraph as rows, each with a status (not started, in progress, done).
2. next_steps.md: the next five steps, one line each.
3. decisions.md: every design decision (database, login, how a sample id is formed) with the reason and a link to an outside reference, not only your own opinion. At the top, a scope changes list with a date on each line, kept whenever the scope moves.
4. security.md: a checklist against the OWASP Top Ten, each item marked done, not needed with a reason, or open.
Stop and show me the four files.

Then
Build a throwaway spike in spike/ that proves the stack: one page that writes one sample row to the database and reads it back. No styling, no login. Touch nothing outside spike/. Stop when it runs and show me the command that proves it.

Attach: A paragraph on what the app must do and who uses it.

Why it works: The plan sits in files that both the agent and the reader can open, and the spike shows the stack works before any feature rests on it.

Drafted with Fable 5.1 from my notes; chosen and edited by me.