AI guide chapter 1: drafting plan and skeletons

For How we use AI, chapter 1 (AI as work interface). Brief: 2026-10-08-ai-guide-brief.md.

Method

Skeleton each section in short bullets (the point, the story, what you say to the AI, what it shows back), and sketch any diagram roughly in ASCII. Commit the skeleton so the evolution is visible.
Critique the skeleton (subagent) and evaluate the diagram sketches: a new diagram must be clearly better than what it replaces.
Flesh out on a separate draft page (guides/drafts/), not in the main guide.
Critique the fleshed-out draft too (text, diagrams, any motion), and fix.
Rufus reviews the draft page; only then does it move into the main guide.

The aim is to need Rufus's check as little as possible before step 5: the critique and evaluation steps are the gate.

Note, 11 Oct: the first pass skipped the critique of the full draft, and the SVGs (with animation) were worse than the ASCII they replaced. The draft has been moved to guides/drafts/chapter-1-ai-as-work-interface.md; the SVGs were removed.

Writing standard

Concise and plain. Short sentences. One idea per paragraph.
Teach by story and contrast: old way, then new way, then what changed.
Show the actual words you'd say to the AI, and what comes back.
One visual per idea, where a picture does the work better than text.
No jargon without a one-line gloss (or the glossary).

Big idea: one running example

Use a single real-feeling project through all four situations so the chapter reads as one story: organising the team retreat (a real Life Itself need this autumn). It starts as a vague "we should do a retreat", which is exactly the kind of mess the loop handles.

Skeletons

Chapter opening

Open on the story (moved from situation 1): a colleague asks the AI for their tasks, has it write them into a GitHub issue, then opens the issue to read them. "No, no, stay here."
Why it's natural: we all learned to work in apps. Habits carry over.
The question: how do you find out what to work on today?
The four situations, the loop diagram, one line on storage (see the big picture).
Introduce the running example: you're organising the team retreat, among other things.

1. What's on this week?

Monday. Old way: Asana, GitHub, inbox, calendar; assemble in your head.
New way: "What's on for me this week?" AI reads the portfolio, your tasks, your calendar.
Shows back (visual: a small chat mockup with a table): three focus items, a star on the one thing, meetings blocking Tuesday.
You reshuffle by talking: "Sylvie needs contracts by Friday, move that up."
The retreat shows up with no next step. That's the cue: not clear → plan.
What changed: principles 1–4.
Footnote: Teresa Torres's /today.

2. Turning a mess into a plan

Old way: a doc nobody finishes; tasks typed one by one; plans stay thin.
New way: talk for three minutes, messily, about the retreat: dates, budget, who, whether at all.
AI plays back (visual: small boxed example):
- Situation: team spread out, last retreat a year ago.
- Complication: no dates, no budget, two people may be away.
- Question: what retreat, if any, can we run before December?
Then a small issue tree (ASCII) and a plan of work; tasks with dependencies (venue depends on dates).
You steer: "No, the real question is whether it's in person at all." It redrafts. The AI drafts first; you judge.
Tasks go into storage. Back to "what's on?", which now has a next step.
What changed: principles 2, 6, 8.

3. Updating progress

During a session: dates agreed → "OK, write an update." Task closed, plan updated.
End of session, land the plane (visual: a four-line landing note: done; didn't work; next start here; learned).
End of week: summary built from the records; no status report from memory.
What changed: principles 2, 6, 7.

4. Others picking up the work (skeleton for discussion with Rufus)

Fae takes over retreat logistics; or an AI session books the shortlist of venues.
Old way: a handover email or call, out of date in a week.
New way: point at the task. What they find: the plan (reference), the task with context and dependencies, the last landing note.
A test: could someone who wasn't in the room pick it up? Could an AI session?
Open: how much detail is enough; what form; when a call is still better.

Close

Principle 8: you still think and steer.
It's a new way of working; give yourselves time to practise.

Diagram principles

Sketch roughly in ASCII first, then evaluate the sketch before drawing it properly. SVG is often the better final form.
A new diagram must be clearly better than what it replaces; if not, keep what we have.
No animation unless motion itself explains something that a still image cannot.
One idea per diagram. Top-down, so it reads on a phone.
Use the glossary's words (storage, reference, tasks, plan).
Text that people may search or edit (tables, notes, examples) stays as text, not pictures.
For SVG: works in light and dark, system fonts, alt text, small file.

Visuals plan (superseded by the diagram principles above)

SVG: before/after interface (big picture): apps-as-interfaces vs you ⇄ AI ⇄ storage. Light/dark safe.
SVG: the loop, with a small CSS animation of a dot moving round the loop (degrades to static).
Chat mockups (SVG or styled HTML-in-Markdown image): "What's on this week?" with the table it returns; the SCQ playback; the landing note. Shows "let it build the view" better than prose.
Keep the ASCII versions as fallbacks until the SVGs are agreed (lip-d4l.3).

Critique (subagent, 11 Oct) and response

Adopted:

Every situation gets the exact words you say and what comes back, with a concrete retreat line, plus a one-line Try this prompt readers can copy.
Situation 4 rebuilt around current practice: a pick-up note with a minimum checklist (goal, current state, next step, what it depends on, where the plan lives, who to ask) and the cold-start test (open a fresh AI session, say "pick up the retreat", see if it finds the next step). Open questions in one line at the end.
Main visuals: the "What's on this week?" chat mockup and the loop with the retreat's route highlighted (static first; any animation reduced-motion safe). SCQ playback and landing note stay as text (blockquotes), not SVG.
Retreat is one row among other work in situation 1; the steering redraft is the centre of situation 2.
Recap table at the close (situation, what you say, what you get back), then principle 8. Fix "two kinds" (now three) and principle-number mismatches; gloss "GitHub issue"; move the session-per-folder detail out.
Target length 1,400–1,800 words for chapter 1; cut prose, not examples.

Critique of the full draft (subagent, 11 Oct) and response

All 20 line edits and the first two structural changes applied to the draft page. The main ones: fixed story inconsistencies (people away "early November", a "learned" line that follows from events, budget no longer appearing from nowhere); landing the plane now updates tasks and plan as principle 7 says; situation 2's monologue shown as speech, with the issue tree named and the redraft's effect on it shown; situation 4 given a "What changed" line and a more useful "Try this"; the manifesto's closing line restored; the loop diagram labels reference as "docs, plans" and notes that anyone can pick up from storage. Footnote URL checked (resolves).

AI guide chapter 1 — drafting plan and skeletons