REFERENCE · WHAT YOU CAN TELL IT
“Fine — but how do I actually correct it?”
Six things you can say.
Everything above this page is what Preston works out on its own. This page is the other direction: the six things you can tell it, and what each one actually moves.
You never type any of these. You say an ordinary sentence to whatever agent you already work with, and the agent relays it. Both halves are shown below — the sentence a person says, and the call it becomes — because knowing what your agent sent is the difference between steering a system and hoping it understood.
All six only ever add. None of them deletes anything, and every one of them is reversible by another one — record a decision that contradicts an earlier one and both stand, as a tension. That is deliberate: a record you can quietly edit is not a record.
record
Put something on the record so it stops being re-litigated.
A DECISION
YOU SAY
“Write that down — we’re running the job queue on Postgres, not Redis. We don’t want a second datastore to operate.”
YOUR AGENT RELAYS
record kind: "decision" statement: "The job queue runs on Postgres, not Redis." because: "avoiding a second datastore to operate" about: "acme-billing"
WHAT CHANGED
A decision now stands at a stable handle. Ask why about it and it answers with your reason. The next agent that calls orient on acme-billing is handed it before it writes a line — so the next person to propose Redis meets the decision instead of re-opening it.
A PROMISE
YOU SAY
“Not now — we’ll fix that N+1 once the billing migration lands.”
YOUR AGENT RELAYS
record kind: "promise" about: "the N+1 in the invoice list query" until: "the billing migration lands" repo: "acme-billing"
WHAT CHANGED
The finding is no longer raised at you every review. It is held as accepted-pending with your trigger attached, and promises lists it — so a teammate can see the fix was chosen and deferred, not missed.
WORTH KNOWING
A decision or a promise, and nothing else. In particular you cannot record how familiar Preston is with a repo, or how risky it is, or how confident it should be — those are what reading your code produces, not something you can type in.
“We’ll fix it later” is refused. A deferral with no stated trigger is indistinguishable from forgetting it, so the promise doesn’t land until you say what closes it.
hold
Stop a whole category of findings from being raised at you for now.
YOU SAY
“Stop bringing up authz for now — we’re mid-migration and we know it’s a mess.”
YOUR AGENT RELAYS
hold dimension: "authz" because: "mid-migration, known" until: "the auth migration lands"
WHAT CHANGED
Preston keeps finding authz problems and keeps counting them — they stay on the ledger and stay walkable — and stops raising them at you. The count is still shown, so you can see how much is being quieted.
WORTH KNOWING
The findings keep being found and keep being counted — you can always see how much you asked not to hear. And a hold has a floor: something that would have blocked your merge still gets said, just no longer as a blocker. You can tell Preston to stop nagging you. You cannot tell it to stop looking.
Whatever you write there is kept word for word and never acted on. Nothing watches the date, and no phrase arms anything — the category stays quiet until a person says otherwise.
lift
Undo a hold — let that category surface again.
YOU SAY
“The auth migration is done. You can bring authz back.”
YOUR AGENT RELAYS
lift dimension: "authz"
WHAT CHANGED
Authz findings surface again on the next review. Nothing else is touched: the findings raised while the hold was on were never deleted, so nothing needs restoring.
WORTH KNOWING
Lifting a category that was never held answers “nothing was held” rather than confirming. So a lift you thought worked and a lift that did are never the same reply.
challenge
Tell Preston it is wrong, and why.
YOU SAY
“That’s wrong — the header trust isn’t a hole, the gateway strips those headers before anything reaches us.”
YOUR AGENT RELAYS
challenge about: "risk:header-trust-in-api-gateway" because: "the gateway strips inbound auth headers before routing"
WHAT CHANGED
Your opposing view is now a claim in its own right, standing against the belief. The belief is not deleted — the two sit as a live tension, and tensions lists it. Your reasoning is attached, so the next reader sees the argument, not just the disagreement.
WORTH KNOWING
Disagreeing does not overwrite. Both readings stay live and visible as a tension until something settles it — which is why a challenge is safe to make while you are still only fairly sure.
The concept behind it: what a tension is.
affirm
Tell Preston it is right — and make the belief carry more weight.
YOU SAY
“Yes, that’s right — we hit that exact bug in March.”
YOUR AGENT RELAYS
affirm about: "risk:unbounded-retry-in-webhook-worker" because: "this fired in production in March"
WHAT CHANGED
Your endorsement is recorded as a second, independent reason to hold the belief. A belief with two live supports is no longer a single guess — it reads as corroborated, and says so wherever it is shown.
WORTH KNOWING
A second independent support is what lifts a belief out of guesswork — there is no score and no threshold to tune. It does not go further: affirming makes a belief corroborated, never final. Ratified is a separate act with a person’s name on it, and no amount of agreement adds up to one.
The concept behind it: how a belief gets stronger.
establish
Turn a rule you state in plain words into a standard the team is held to.
STEP 1 — YOU SAY IT
YOU SAY
“Anything that changes behaviour ships behind a feature flag. That’s a rule here.”
YOUR AGENT RELAYS
establish rule: "behavioural changes ship behind a feature flag"
WHAT CHANGED
Nothing is adopted yet. Preston reads your sentence against the standards it knows, and offers back what it thinks you mean — which standard, and which of the three kinds it is — plus a handle to confirm. If it holds no standard matching what you said, it says so plainly instead of inventing one.
STEP 2 — YOU CONFIRM
YOU SAY
“Yes, that’s the one.”
YOUR AGENT RELAYS
establish confirm: "<the offer handle from step 1>"
WHAT CHANGED
Now it is professed: the team’s own standard, in the team’s own words, carried into every later review — including sessions and repos that have nothing to do with this conversation.
WORTH KNOWING
Nothing becomes the team’s standard until you confirm it, which is why this is the one verb that takes two turns. A standard can never appear because an agent phrased something confidently.
The concept behind it: principles & disciplines.
Things that are not verbs
Three sentences people reasonably try. None of them is a tool, and it is better to know that here than to discover it when your agent improvises something adjacent and you believe you changed something you did not.
“Preston, you’re now 80% confident in this repo.”
Confidence, familiarity and risk come out of Preston reading your code. They are results, not settings, and they are refused by name.
“Profess this principle.” / “Permit this exception.”
Neither is something you can call. Say the rule in plain words to establish instead — same door, opened by what you meant rather than by a key you would have had to guess.
“Forget that.” / “Delete that belief.”
There is no delete. Challenge stands your reading against Preston’s, lift undoes a hold, and withdrawing something is itself a recorded act with a cause. What it believed, and when, stays walkable — otherwise none of the history means anything.
record and the five beside it are the real, frozen V1 tool surface — the reference page carries each one’s exact signature, what it returns, and when to reach for it.
NEXT
The tools →