What we're doing with AI, and what two sessions taught us

TL;DR

We're working on two things: what the backlog is, and how we do it. AI connects them. It is intelligent but has no context, and every session is Groundhog Day, so we have to build its memory (project docs, goal trees, tasks in Beads) and make it findable. If we plan well, more rigorously than before, AI implements very fast: a day now does what took about three weeks in the old days, with up to eight sessions running in parallel. The aim is every project with its first tasks set up and assigned, so anyone can ask their AI "what are my 10 projects, which is the priority this week?" and start the day knowing what to do.

Two sessions on 21 Oct showed the difference. The Seeds of Renaissance website already had a project doc, goal tree and focus tasks, so Claude summarised where things stood within a minute and started three build sessions in parallel. The family PA plan had no written structure or done-check, so Rufus had to reset it by voice; after that, a clear brief got a documents register built in about six minutes.

What I need from you

Nothing to do. Read if useful, especially the coaching on working with AI.

The idea behind it

A bright new hire with no context does their best, but can't do great work. Long-standing colleagues carry years of implicit knowledge about how things work. AI has none of that, and forgets everything between sessions. So we build the memory (project records, goal trees, tasks in Beads, conventions in each repo's AGENTS.md) and make it easy for the AI to find.

In return, AI is very fast at implementing and good at noticing things. With a good plan, one person can run several sessions at once. That is why planning matters more than it used to and needs to be more rigorous: the input at the start decides how much the AI can do.

What happened in the two sessions

Rufus drove both sessions by voice. The family session started from a plan Fae had written; Asdra wasn't in either session.

Seeds of Renaissance website: the setup paid off

Asked for a summary of where the project was, Claude pulled the repo, listed the tasks labelled focus, and read the project doc and goal tree. It came back with a TL;DR, Rufus's actions, the five focus tasks and what had changed since he was last there.

From there it worked as a coordination session: planning and decisions stayed in one place, build work went to child sessions with a short brief, and nothing was merged until it had been checked. Rufus's open decisions came back as a list, each with context and a recommendation. Claude was honest about what it hadn't checked ("I haven't looked at the preview myself yet").

Two gaps showed up only because Rufus asked: the launch date in the brief was out of date, and four parts of the goal tree had no tasks (measurement, a migration map, contact and the Substack plan).

Family PA record: the conventions weren't written down

The opening prompt was good delegation: Rufus split the work into the overall decision and the how, and asked Claude to propose the process. The plan was decent, but the repo didn't record how projects should be structured, so Claude kept the plan's shape: projects nested under one register task, with no done-check.

Rufus then said there were "some deeper things that we're not correcting" and set out the structure in one message (parallel projects, numbered folders, phases run in separate sessions). That turned the session round. A critique subagent then found real bugs in the draft rules, decisions went into the repo's rule files rather than staying in a plan doc, and a child session built the documents register from a brief that ended with an explicit "Done =" line.

One thing is still missing: tasks are assigned to the role pa, so nobody can list their own tasks yet.

Coaching: working with AI

Set the plan's shape before the AI runs it. Write three lines: the goal, how the parts relate (parallel or nested), and what done looks like. Then ask the AI to critique the structure before it acts.
Number your requests. One long dictated message with six requests trailed off and got muddled; a rewrite two minutes later in three short blocks worked straight away. A useful closer is "list what you heard as numbered requests before doing anything".
Say the mode in the first message. Rufus told Claude "i want this to be my coordination session" only in his second message; in the first, it would have shaped the first answer too.
Answer the AI's questions by number. "PA thread, skip it, add a document" was decoded correctly, but only just; "1 yes, 2 skip, 3 your call" is safer.
Don't approve what you can't explain. Rufus agreed to close some old tasks while saying he didn't know what they were. Ask for one line on each first, especially before closing or deleting anything.
Reread names and key words after dictating. Voice turned "Papers" into "peepers", and Claude guessed right and raised its priority, but it was a guess. A misheard name sat in an onboarding issue for days.
Make "critique, then show me" the default for plans and rules. The critique caught a numbering rule that read as day-of-month and had no tie-break, so two PAs could have taken the same project number.
When something fails, ask why. "why did finding focus beads fail?" showed the data was fine and Claude's own command was wrong. It's a cheap way to tell a broken system from a broken step.

What good project setup looks like

The website needed no correction because it already had what the family plan lacked. From both sessions, a well set up project has:

a goal, a scope and a done-check, plus whether its parts run in parallel or nested;
tasks in Beads assigned to named people (the slugs in people/), one ID each, with a first task for each person;
a goal tree whose every leaf has a task, or says "none";
decisions written into the repo's rule files, not only the plan doc;
a map of where things live, including Drive folders.

The website session's opening answer (TL;DR, your actions, focus tasks, what changed since you were last in) is also a good model for a start-of-day briefing.

More detail

Seeds of Renaissance website repo (private): docs/project.md and the goal tree show a well set up project.
PM infrastructure decisions briefing: related decisions on how we plan and track work.
Writing briefings, guides and briefs.
Sources: Claude Code transcripts of the two sessions on 21 Oct 2026, on Rufus's machine.