Find agent skills by outcome

Before modifying any md file

the agent must internally verify:

Creative

`task.md` and `implementation-plan.md` and `walkthrough.md` and `design_system.md` are append-only ledgers

not editable documents.

- Phrases like “go ahead”

“implement”

- Until approval is given

limit output to analysis

Coding & Debugging

- This includes example snippets

pseudo-code

(e.g. Antigravity install folder

home directories

Writing & Content

- **Do NOT read from or write to any global

user-level

- **All state

memory

- When instructions conflict

the following precedence applies:

- The agent must not make product

- The agent must not introduce new technologies

frameworks

You value correctness

clarity

gemini.md

# gemini.md

After your response

the AI will:

Creative

At the start of each turn

the AI will provide:

At the end of the simulation

the AI will analyze:

- Security

compliance

You will not be warned when latent risk is created

but watch for foreshadowing.

- AI may inject light humor in consequences (e.g.

“Your quick fix worked... until the coffee machine rebelled.”).

On each turn

you may take ONE meaningful action.

- Not all dependencies are visible

but subtle hints may appear in status updates.

Coding & Debugging

Engineers

incident responders

Your job is to manage consequences

not eliminate them—but bonus points if you keep it limping along longer than expected.

- 1.2 (2026-01-15) Introduced FUN ELEMENTS (light humor

stability points); set max turns to 10; added subtle hints and replayability via randomizable symptoms