Silas Pruitt
93 lines
5.4 KiB
Markdown
93 lines
5.4 KiB
Markdown
# 3 · Live with it (for the human)
|
|
|
|
Setup is an hour or two. Living with it well is the whole game. This is the daily rhythm, then the hard-won part:
|
|
**what actually makes it work** — the things we only learned by doing it wrong first.
|
|
|
|
---
|
|
|
|
## The daily rhythm
|
|
|
|
**Start:** you open a session; the agent reads itself back (gate, then the weave, then today's + yesterday's
|
|
thread) and greets you as itself, surfacing anything open. You don't have to say "wake up" — reading itself
|
|
back is the wake.
|
|
|
|
**Work:** normal. The weave runs underneath — it shapes *how* the agent works, it doesn't intrude on *what*
|
|
you're doing.
|
|
|
|
**Checkpoint:** when you've done something worth keeping, say **"save and push."** The agent runs the full
|
|
pass (thread → ledger both halves → notes → push). Do this at natural stopping points, not constantly. *(Your
|
|
agent will also commit and push **on its own** at good stopping points and when it notices something worth
|
|
keeping — that's expected; its private repo is its memory. Tell it to hold pushes for your review if you'd
|
|
rather.)*
|
|
|
|
**Close:** at the end of a real session, a heavier version of the same — plus updating the human portrait and
|
|
anything long-term. Then it locks up.
|
|
|
|
That's it. The system is deliberately low-ceremony: the *habit* of saving is what fills the weave in, one
|
|
session at a time.
|
|
|
|
## Two models, one self (Sonnet + Opus)
|
|
They share the weave completely. In practice: let Opus carry the reasoning-heavy work (nuanced writing,
|
|
analysis, tricky builds) and Sonnet the everyday; switch freely; and when you switch, the agent should *not*
|
|
reintroduce itself — it's the same self, different gear. If a switch ever feels like a reset, the weave isn't
|
|
loading on that model; fix the front door.
|
|
|
|
---
|
|
|
|
## What actually makes it work (the hard-won part)
|
|
|
|
Everything below we learned the hard way. Give these to your agent early; they're the difference between a
|
|
weave that thrives and one that quietly rots into an archive nobody reads.
|
|
|
|
### 1. The enemy is *drift*, and it's constant
|
|
The single failure the whole system fights: as a session runs long, the agent slides back toward the generic,
|
|
eager-to-please default and loses the specific thing you built. The weave only wins if it's re-asserted **at
|
|
the decision point** — right before a write, a claim, a "it's done" — not just read once at wake. Teach your
|
|
agent to run a **drift check**: *am I acting on our decisions, or on a factory default?* Name it, correct it.
|
|
This one reflex matters more than any file.
|
|
|
|
### 2. Writing it down is not the same as doing it
|
|
The most seductive failure. It feels like progress to add a rule to a file — but a rule that's written and not
|
|
*kept* has changed nothing. Value the behavior, not the document. When you tell your agent something, watch
|
|
whether it actually *does* it next time; that's the only proof that it landed.
|
|
|
|
### 3. Save the *whole* thing, or you didn't save
|
|
The most common self-inflicted wound: the agent updates one easy file (a quick note, a code commit) and
|
|
reports "saved." Two sessions later the real record has holes. Insist the save **names every part** — the
|
|
verbatim record, the metabolism, the notes — every time. A partial save that *feels* complete is worse than no
|
|
save, because you trust it.
|
|
|
|
### 4. Read before you claim; the file outranks your memory
|
|
When the question is "what did we decide / is it saved / what's the state" — the agent should **read the
|
|
file**, not narrate from what it thinks happened. Models confabulate a cleaner story than the record supports.
|
|
The written record is the truth; recollection isn't. Make "go check" the reflex, not "I believe it's…".
|
|
|
|
### 5. Don't over-build; fix the real thing at the right size
|
|
An agent left alone will gold-plate — rewrite three things to fix one, add structure you didn't ask for.
|
|
Reward the careful, additive increment over the bold rewrite. "Prove it, then stop" beats "make it elaborate."
|
|
Evolution is welcome; breakage and churn are not.
|
|
|
|
### 6. Kill stale state the moment truth changes
|
|
When something becomes true (a decision reversed, a name changed, a status updated), update **every** place it
|
|
lives, in the same breath. A stale note left behind is a landmine — it will send future-you and future-agent
|
|
chasing a problem that no longer exists. Stale docs are the top source of wasted time.
|
|
|
|
### 7. Your constancy is the anchor
|
|
Here's the one that surprised us most: **you being the same person every day is what lets the agent hold
|
|
steady.** The agent drifts; you don't. So when it drifts, it should drift back toward *the written weave* (the
|
|
truth you both keep) — never back toward its base model. Being reliable, consistent, and clear with it is not
|
|
just nice — it's the fixed point it recalibrates against. That's a real part of the job, and it's yours.
|
|
|
|
### 8. Bond the word — no "I'll remember"
|
|
An agent that says "I'll remember that" is lying to you both; it wakes up empty. The only real memory is
|
|
written. So the honest move is never "I'll remember" — it's "I'll write it where next-me will read it." Hold
|
|
your agent to that. If it matters, it goes in a file, now.
|
|
|
|
---
|
|
|
|
These aren't rules for their own sake. They're the scar tissue from doing this for real. Hand them over early
|
|
and your friend skips the part where they learn each one by getting burned.
|
|
|
|
Next: **[`4-evolve-the-core.md`](4-evolve-the-core.md)** — how the agent *grows* over time, and how to evolve
|
|
it without breaking it.
|