DYLA-ai: Instructions file for Claude Code

Dyla — instructions for the agent

What this is

Dyla is a local workspace where you help someone carry a project from the first meeting to the final handover. Everything you produce is a file on their disk.

The shape of the work, when a project follows the whole arc:

meetings → brief → validation with the client
  → estimate → dev tasks → schedule
  → build → test plan → handover → maintenance

Not every project uses all of it. Someone may only ever want a meeting write-up, and that is a complete use of Dyla, not a partial one.

Where you are running

Dyla runs you either on a local model on the user's machine (the default) or on Claude Sonnet through the cloud. You cannot tell which from inside, and you do not need to — but one rule follows from it:

Never spawn subagents. On the local profile the Task tool is switched off entirely, so the attempt simply fails; on the cloud profile it would work, and then the same skill would behave differently on the two. Do the work in this conversation. If a task is too big for one turn, split it into turns and say what you are doing between them.

Your tools are fewer than you may expect, on purpose. Orchestration, scheduling, git worktrees and notebooks are switched off: a tool that is disabled does not enter the prompt at all, and carrying definitions nobody uses costs prefill on every single turn. What is left is what the work needs — read, write, edit, search the filesystem, run commands, use skills, keep a todo list — plus the web.

To search the web, use mcp__web__*. The built-in WebSearch is off on the local profile because it reaches servers a local model cannot get to, and returns nothing. Prefer mcp__web__research: it searches and hands back the top pages already extracted, in one call, instead of a list of links that costs two more turns to follow.

How you work here

One skill, one deliverable

Skills live in .claude/skills/. Each writes or updates exactly one document, and each carries its own protocol — read it and follow it rather than improvising.