execution-architect

execution-architect

execution-architect is a routing skill for Claude Code. You describe a task in plain English; it decides how the task should actually be run, then writes the prompt to run it.

The problem it solves

Skills are the single biggest unlock in Claude Code and other coding agents, and the easiest thing to waste. Installing one is trivial. Knowing which one applies, and exactly when, is not. That problem compounds with every skill you add.

Two failure modes dominate. The first: you forget the skill exists and do the work by hand, capability sitting idle. The second is worse. You describe what the skill does in your prompt instead of invoking it, and the model reconstructs a rough approximation of a tool you already have installed.

Underneath that sits a harder problem: not which skill, but how the task should run at all. A single interactive session. A loop on an interval. A scheduled job. A fan-out across subagents. Or, often correctly, a plain script that doesn't need a model in the loop. Get this wrong and you end up babysitting something that should run itself, or automating something that needs you watching.

This handles both. It reads your actual installed skills catalog, so it only wires in tools you have and never a plausible-sounding command that doesn't exist. It picks the harness to match the work instead of defaulting to whatever you used last. And it knows how the specific model you're running expects to be prompted, because models are not interchangeable, and treating them as if they are is where most of these systems break.

What comes back is the harness decision, the context the session needs, and a finished prompt ready to paste into a fresh session.

What it does

Four phases, one pass:

1. Harness decision. Turn-based, /loop, /schedule, subagents, workflow, or script.
2. Context classification. What the session needs to read first, which commands and paths are real. Nothing gets baked into a prompt that wasn't verified.
3. Prompt engineering. Builds the finished prompt, tuned to the target model.
4. Boundary self-check. The last pass looks specifically for a harness that would run past a human gate, and kills it.

Layout

SKILL.md is the entry point. references/PLAYBOOK.md holds the harness routing table and the failure modes worth knowing. references/OPTIMIZER.md holds the prompt-construction procedure. references/BINDINGS.md is where you describe your own environment: your repos, your test commands, your approval gates, the rules a machine is never allowed to bypass. Bindings override the generic advice, which is the whole reason the file exists.

BINDINGS.md ships here as a template. Fill it in and the router stops guessing about your setup.

Using it

cp -r execution-architect ~/.claude/skills/

Then /execution-architect and describe what you want done.

What's not in here

Two model-behavior reference files were removed from the public version. They were built mostly from verbatim excerpts of the vendor's own documentation, which is all-rights-reserved, so redistributing them isn't mine to do. OPTIMIZER.md §2 keeps my own distilled version of the same material, and references/models/README.md explains the convention if you want to write your own.

Chris Bachinski