Week 3

Prompting & refining

On their own documents: the proposal, the reverse brief, and what correction teaches the system to keep.

Works.

A vague brief gets a vague draft.

Tell Codex exactly what you want, not everything you know. Specificity beats completeness, every time.

Don't describe the thing. Attach the thing.

Files, URLs, screenshots, transcripts, voice notes. The more real material Codex sees, the sharper the draft.

Tighter. More like this. Swing harder.

Refine mid-thread instead of restarting. A restart throws away everything Codex already understood about the task.

Artifacts earn their place, they don't come as standard.

Use them for anything you'll iterate on visually or hand off as a document. Skip them for a quick answer, they add friction with no payoff.

Concept

It did what you said. Not what you meant.

Say "cut it in half" and mean crosswise. Codex cuts lengthwise, technically correct, practically wrong. The fix is never a smarter model, it's a clearer instruction and more context.

Default cheap. Escalate only when it earns it.

Fast, inexpensive models for everyday drafting. Save the strongest model for hard reasoning and client-facing strategy work, and check the current lineup before you quote a name in the room.

Concept

A correction you don't write down never happened.

Every time you correct Codex mid-thread, ask where it goes: a skill file, a context file, or a memory entry. If it doesn't land somewhere, the next run makes the same mistake.

Concept

Codex drafts. The named person reads. The named person sends.

Standing rule, not a phase you graduate out of. Lachlan reviews and sends his own proposals. Susan reviews and sends her own reverse briefs. Nothing goes to a client unread.

Concept

Content is never written from words you didn't say.

Don't ask Codex to generate cold. Run a structured interview first, get the person's actual words into a transcript, then build every draft from that transcript.

Don't say "try again." Diagnose first.

A weak draft is one of two problems: missing information, go back and pull more from the person, or an editorial problem, revise the draft you already have. Different problems, different fixes.

Rules decay if nobody revisits them.

Review tight and often while a workflow is new. Once it's stable, review less often, but never stop.

Get feedback from the person who uses it, not just your own read.

Refine, but don't let the cycle run forever. At some point you ship and watch it in the real world, and the team's confidence in the workflow is itself worth building.

Drip it live, one layer at a time.

Don't hand over a finished pack and disappear. Manual before scheduled, prove it works by hand before you automate the trigger.

Lachlan's proposal, drafted before he opens the template.

Design proposal, stage 1 pricing, stage 2 pricing, drafted from his own templates. Lachlan reviews it, Lachlan sends it, and the whole cohort watches this one first.

The reverse brief writes itself. Susan still signs it.

Auto-drafted from the questionnaire, the meeting transcript, and her notes into the client's "we heard you" document. Lachlan's own words on the target: from a few hours for Susan to nothing, a working estimate, not a promise yet.

Before next session

This week's work

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

Lachlan: run the next real proposal through the Proposal Engine, review it yourself, send it yourself.
Susan: run the next real reverse brief through her own Command Centre build, review it yourself, send it yourself.
Whoever's draft needed more than one round: route that correction and name where it went, a skill file, a context file, or a memory entry.

Exit criteria

You're done when

Lachlan's proposal or Susan's reverse brief went out faster, and better, than before Codex touched it.
Every correction on either draft has a named destination: skill file, context file, or memory entry.
The room can name all three destinations without looking it up.