WAYS OF WORKING · APPLIED NINO CHAVEZ
Every session with an AI agent is a transcript you usually close and forget. Kept and mined instead, the pile becomes a corpus — reusable prompts, your corrections as standing priors, and counted waste you can act on.
[ preference · i-prefer ]
"i want interactive html demos, not mp4s, for all use cases."
mined from a past session — and surfaced into this one, which is how these interactive decks got built instead of a video.
→ or scroll to advance · ← to go back
02 THE PROBLEM
A correction you gave the agent, a prompt that finally worked, a waste you noticed and meant to fix — all of it lives in the transcript, and the transcript ends. The next session starts blank, and you re-derive what you already learned. The raw material is just sitting there, unsearched.
Those aren't a backup you'll never open. Mined, they're the one thing that lets an agent with no memory start where the last one left off.
03 THE MINE
A small tool walks the transcripts and pulls three kinds of thing back out: prompts worth reusing, corrections worth keeping as priors, and waste worth counting. SQLite and full-text search make the pile queryable.
$ recall-scan --since 30d # walk transcripts, classify entries ✓ 585 user messages · 2,034 sentences scanned ✓ 1,419 signals extracted → SQLite + FTS5 $ recall "how did we set up turnstile" → 3 past sessions · the working incantation, ranked by recency
The point isn't the tool — it's the move: your own past sessions are the highest-signal reference you have, and they're already written. You just have to read them.
04 CORRECTIONS AS PRIORS
The richest vein is your corrections. Extracted and distilled into a character stack, they load as context so the next agent vets an idea the way you would — recognizing the shapes you routinely reject before you have to reject them again.
"A character stack extracted from prior sessions — the same red lines, rationale, and taste. Treat it as a lens, not a script: recognize the question shapes Nino routinely overrides with 'go', and skip asking them." ~/.claude/poe/stack.md — 1,419 signals across 74 projects, regenerated from the corpus
This deck is a live example: the "interactive, not video" prior on the title slide, and the bias to act rather than ask permission, both came from the stack — and both shaped what got built here, before you had to say either out loud.
05 WASTE AUDITS
"The agent keeps leaving stray servers running" is a shrug. A hundred-plus times, across the corpus is a mandate. Mining converts a vague irritation into a number big enough to justify building the fix — which is exactly where the wrapper on the guardrails deck came from.
$ recall-scan --waste pkill # stray http.server processes killed by hand, counted across sessions → 122 # the annoyance, quantified — build `preview` and stop chasing them
One discipline carries over from the verification demo: re-derive the number on use. A counted figure in an old note is a stale snapshot; mine it fresh before you quote it, or it becomes the next thing that lies.
06 THE HONEST LEDGER
A corpus that only remembers wins lies by omission. The valuable ledger grades what the agent built against what survived — the features quietly deleted, the abstractions reverted, the tool rebuilt until it finally held.
The confession is the signal
The mining tool's own honest entry is that it was rebuilt several times before it was made durable. That failure is worth more than any success it logged — it tells you which shape actually lasts.
Survival, not activity
Count what's still in the codebase months later, not what got committed. The corpus makes "did this last?" answerable — and answering it honestly is what keeps the priors trustworthy.A prior earns its place by surviving, not by being said.
07 HONEST LIMITS
The stack is a snapshot
It reflects what was true when it was generated. When the current conversation contradicts a mined signal, the live conversation wins — a prior is a prior, not a law. Preferences move; the corpus lags.
Regex misses nuance
Signals pulled by pattern-match from CLI prompts capture the shape of a preference, not its full texture. The stack is explicitly a lens for terminal cadence, not a voice for prose — treat missing guidance as missing, not as permission.
Local, and only yours
The corpus is machine-local and personal — it encodes one operator's taste, not a team standard. That's the point and the ceiling: it makes your agent sharper, and it doesn't transfer without re-mining.Trust the conversation over the corpus every time they disagree.
08 USE IT
If you never touch code
Before starting something you've done before, go find the last time you did it. Your old threads, emails, and notes are a corpus too — the reusable version of you is already written down, if you search it instead of redoing it.
If you're technical
Keep your agent transcripts and make them searchable. Grep for one counted waste number before you build anything; save the corrections where the next session will read them — a prompt file, a memory, a loaded stack.
If you build systems
Treat the session log as a data source, not exhaust. Mine prompts, priors, and survival — and re-derive the numbers on use. The corpus is only worth trusting if it's cheaper to search than the work is to repeat.