Pattern of Interaction with an LLM
This document captures a conversation about how to interact with an LLM.
The conversation compared old patterns of human work with LLM work.
The analogies are strong.
The Core Idea
The unit of work is a work item, not a conversation.
A conversation is ephemeral.
A work item is durable: goal, constraints, deliverable, and acceptance.
The human writes work items.
The LLM executes them.
The human reviews the results.
This is the email pattern applied to machines.
Human Documents versus Prompts
Human documents persuade, justify, and record.
A human reader needs motivation, history, and defense of a decision.
A prompt constrains a literal-minded entity with a context window and no memory.
A prompt is a work order, not a whitepaper.
A work order states the goal, the constraints, the deliverable, and the definition of done.
The document habit that carries over is version control, single source of truth, and review cycles.
The document habit that does not carry over is narrative persuasion.
Keep the human-shaped document as durable memory.
Let prompts reference the document instead of restating it.
Episodes
An episode is a bounded unit of strongly related work with one goal and one acceptance.
Group strongly related items into one episode.
One episode pays dispatch overhead once.
Dispatch overhead is prompt setup, context loading, and human attention per handoff.
One episode also preserves coherence.
A model that handles the whole episode keeps the episode’s context.
A model that handles fragments loses context between fragments.
The queue holds episodes.
Sub-tasks live inside the episode.
Sub-tasks are the execution plan, drafted by the model and ratified by the human.
Sub-tasks are not decision points for the human.
Size an episode so one competent reviewer can review its complete diff in one sitting.
Risk Items First
Put risk items first.
Phrase each risk item as a question.
A question is a hypothesis with a checkable answer.
Verify the risky assumption before building on it.
The model can draft the risk questions.
The human orders them and judges the answers.
The TODO Habit
The TODO list is a queue of episodes.
Update it near daily.
Check off items done.
Cross out items deliberately dropped.
Carry forward items meant to be done soon.
Keep speculative items unchecked.
Groom the queue every few weeks.
The model can draft items.
The human decides.
Noise
Watching a model think is a waste of human attention.
The model thinking is where the value is produced.
The problem is attention, not thinking.
Ask for artifacts, not transcripts.
An artifact is a summary, a diff, or test results.
Work asynchronously so there is nothing to watch.
Async Interaction
Send off a work item, then proceed with other work.
Check for results later.
Git is the message channel.
The model commits on its branch.
The human merges when the episode settles.
The merge is the check-the-mailbox moment.
Review
The review surface makes async trust possible.
Review the artifact, not the process.
The human is the synthesis point and the final authority.
The model proposes; the human disposes.
A model reviewing its own output in the same context is weak.
A fresh-context review pass finds misses, like fresh human eyes.
Consensus Ritual
The old ritual: walk around, gather perspectives, synthesize, run past the group.
The LLM version: fan out several perspectives, synthesize yourself, review with fresh context.
Fan out can mean several agents with different expert angles.
Synthesis stays human.
The review pass uses fresh context.
Isolation
Branches do not isolate working trees.
Git worktrees do.
Run the LLM in its own worktree on its own branch.
See 01-async-worktree.md for the worked example.
Worktrees versus Records
Worktrees and checkpoint records solve different concurrency problems.
Git worktrees isolate concurrent streams of work.
They suit general development: parallel features, independent files, several LLM sessions at once.
Checkpoint records suit serial work against one shared live system.
Hardware and operations work is single-threaded by nature; concurrent changes against physical devices are usually bad.
The homelab exercise ran that way: every change touched the same switches, hosts, and cables, so the pattern that worked was a checkpoint or handoff record per session, not parallel worktrees.
Choose by the work, not by habit:
- Parallel, file-isolated development: worktree per episode.
- Serial work against one live system: checkpoint and handoff records per session.
- Both remain valid; the same project may use both for different threads.
Intent and Record
A work order is intent.
Intent lives in prompts/ before dispatch.
Results are the record.
The record is the diff, the decisions, and the outcome.
The record lives in records/ after completion.
The repository history is the conversation’s product, not the transcript.
Notebooks
Notebooks are a thinking medium, not a work-item medium.
Notebooks have hidden state, non-linear execution, and noisy diffs.
Use notebooks for exploration.
Keep work items as version-controlled text.
Open Question: the DELTA Protocol
The DELTA correction protocol was suggested in a prior LLM conversation.
It is defined in prompts/01-contract.md and prompts/common/02-universal-rules.md.
Its merit is not yet established.
Whether to keep, change, or remove it is open.