Week 5

Skills as playbooks

One of your own builds finishes today, as a skill Codex runs the same way every time.

Works.

Works already ships eight skills before any custom work starts. Yours is the ninth, not the first.

Review, Debrief, Research, Pre-call Prep, Voice Check, ICP Score, Win/Loss Log, Quote Capture, each a clean input, process, output.

Concept

A prompt says what to do right now. A skill is the same method, running correctly every time the task recurs.

Rachel Woods's frame: a skill is an SOP written so Codex can execute it instead of a human reading it.

What do you deliver every week, on a schedule, without fail. That's your first skill.

Kat's weekly cost-code report is exactly that kind of output, the one this session builds from.

Concept

A skill has seven parts. Miss one and the method stops being repeatable.

Name, description, trigger, inputs, steps, output, examples, each filled with a real example from the build in front of you.

Most skills fail because they're written abstractly. Specificity beats abstraction, every time.

Write the exact step, not the general instinct.

Fix the skill, not the prompt.

Kat's cost-code tool is already a working estimate of about 80 percent there. The gap gets fixed in the skill file, not by writing a smarter prompt next time.

Two disciplines before you build anything non-trivial.

Write a half-page plan and get sign-off before you build. Validate against a throwaway example before you ever touch real data.

Concept

Three questions, before anything runs unattended.

Answer no to any one of them, redesign it, don't push ahead on enthusiasm.

A skill only gets the connectors and permissions its workflow actually needs, nothing more.

We read from Jack and Databuild once the read-only exports land. We never write into either.

Concept

Access gets Codex into the work. Skills teach it how the work gets done.

Most companies think they're further along than they are, because they've only done the first rung.

Public skill libraries will mostly be generic. Your own judgment isn't.

Your escalation process, your qualification lens, your review standard, a competitor can't download those.

A skill raises the floor. It doesn't cap the ceiling.

The skill carries the repeatable work. The person stays in the loop for the genuinely novel calls, the first and last mile of the job stay human.

Kat already built this herself. In this room, she finishes it as a skill.

Cost-code tool, v1 shipped solo, a working estimate of about 80 percent there, finished live on Codex.

Concept

Pull context, execute, write back what happened. That's the shape of every skill, not a swarm.

The write-back feeds the next run's context, decisions, blockers, next actions.

Tool subscriptions are rent. The skills you build are the asset.

Rent stops the day you stop paying. Skills compound, and they stay when the vendor changes.

Before next session · built and demoed on Codex

This week's guided projects

Susan: the reverse brief, formalised into a repeatable skill.
Michael: the QA Console's shared checklist state, scoped around the review he keeps, not replacing it.
Mollie / Grace: the Project Lead Command Centre, one live view fed from one capture point instead of three.
Melony (Mel): the Pack Generator, only once a team win is already visible. Her track may not reach a full self-build by this week, and that's acceptable.
James: Quote Control, scoped against the three suppliers already agreed to trial. The takeoff half stays hard to win and stays unsold as anything more.
Trent: the image library and render-compliance question, a small build honestly sized as small.

Exit criteria

You're done when

One skill, used three times, demonstrably better on run three than run one.
You can explain exactly what changed between runs and why.
The skill has a written plan and passed at least one fixture run before it ever touched real data.
Any workflow set aside has a stated reason, not just "didn't get to it."