Reading your mail with an AI harness

Your AI harness reads conversations once: receipts, travel, events, tasks, who is waiting. · 5 min read

Some things in your mail are only found by reading it. A Home Depot receipt says 885911470902 FLEXVOLT 20- and never “DeWalt”; “I’ll get back to you Monday” means you owe the next message even though you wrote last. No search finds those and no rule decides them.

So InboxQL keeps a short list of things worth knowing later — elements — and lets the AI you already use (Claude Code, Antigravity, anything that can run a command) read your conversations once and answer them. The answers are checked, kept with who gave them, and become as easy to ask about as from:.

The elements

Element	About	What it records
`receipts`	each message	merchant, date, order number, every item with its brand and price, total
`travel`	each message	flights, trains, hotels, cars: provider, confirmation, from, to, when
`events`	each message	things to put in your calendar — proposed, for you to approve
`tasks`	a conversation	what someone has to do, and who owes it — proposed as tickets
`waiting`	a conversation	whether someone is waiting: on you, on them, or no one
`importance`	a conversation	important, normal, low or ignorable
`sender`	a sender	person, organization or system
`topics`	a conversation	a few tags, chosen from the ones in use

Install them with iql elements install, or from Settings → AI Configuration → Starters. iql elements show receipts prints one in full: the instructions the reader is given, and its fields or levels.

Pointing your AI at it

Tell your AI to run:

iql work guide

That prints everything it needs. The loop is two commands:

iql --json work next --limit 10          conversations to read, newest first
iql --json work submit < answers.json    one answer per conversation

work next hands over each conversation readably — the email rather than its HTML, the text of its attachments, who is in it — with what to answer, and leases it for half an hour so two runs do not read the same thing. Narrow it with --only receipts,travel or --query "from:homedepot".

A prompt that works well: “Use iql. Run iql work guide, then answer the newest 100 conversations.”

Why you can trust the answers

Checked before they are kept. Levels must be the element’s levels, records must have its fields and types, and every judgement and record must quote words that really are in the conversation. One problem keeps nothing for that conversation and says what to fix.
Kept with who gave them. Every answer carries the model that gave it, a confidence and why. iql work status counts them by element and model.
Yours win. Right-click a label chip to rule on it; your ruling outranks the AI’s, now and after re-reading. Review in Annotators draws a random sample, and iql annotate score waiting --model claude-opus-5-5 says how often that model was right.
Undoable. iql work clear --model claude-opus-5-5 --since 1d removes a run’s answers, and those conversations are read again next time.
Proposals, not actions. Tasks arrive as proposed tickets on the Board and events as proposed calendar events; nothing happens until you accept them.

Measuring the loop

Intelligence → Work loop shows how well the reading is going: how often each kind of answer is right, how much effort it takes, and where it should read more carefully.

Guesses are kept apart from conclusions. work next --template fills every answer in advance with a guess, so the AI only edits what differs. An answer sent back exactly as guessed is stored as a guess: it is counted separately, a deeper read (iql work next --deep) asks it again, and it does not teach the next template anything about that sender. The template never guesses where the conversation looks like it holds the thing — a receipt-looking subject is never pre-filled with “no receipt” — and never adds mail labels; it only suggests them.

Checks. iql work audit next hands answered conversations to a second reader — ideally a different or stronger model — to read again from scratch, without seeing the first answers. Half of each batch is drawn at random, and only those measure accuracy; the other half goes where a wrong answer is likeliest and costliest. Where the two readers agree, the answer stands. Where they disagree, the conversation is flagged for a deep read and the pair waits for you under Waiting for you, as Answer A and Answer B — which was the first read and which the check is shown only after you choose (keys 1, 2, 3 for neither). Your choice becomes a ruling that outranks both.

Targets. Each element has a target (how often it must be right) and a cost (what a wrong answer does: a missed receipt costs more than a wrong topic). Where checks show an element missing its target, the template stops guessing it and it has to be read; costlier and less certain elements are checked more often. A sender whose answers passed ten checks in a row earns the cheap path back. Change targets on the page, or with iql work targets set receipts --target 0.99 --cost high.

What the page shows:

Right, on random checks — with the range it is probably in. A range that dips below the target means more checks are needed before trusting it.
Your reviews — your rulings from Review in Annotators.
What you did with its work — proposed cards and events you approved or rejected, labels you took off, conversations you moved back to the inbox: corrections you make without being asked.
Effort — conversations per day, refusals and why (rework), leases that ran out, time from lease to submit, and cost per conversation when the harness reports it (work submit --input-tokens … --output-tokens … --cost-usd …).
Confidence — whether answers that say they are 90% sure are right 90% of the time.

From a terminal: iql work metrics --since 30d, iql work audit list, and iql work audit resolve <id> original|checker|neither.

Asking about the answers

extract:receipts.items.brand=DeWalt          anything DeWalt you bought
extract:receipts.merchant="Home Depot" after:2025
extract:travel.kind=flight after:2026-01-01
awaiting:me | timeline                       conversations you owe a reply
in:contacts awaiting:me -kind:system         people waiting on you
extract:importance.level=important after:7d
topic:kitchen-remodel

A path through a list — receipts.items.brand — matches if any item does. unlabeled:waiting is what nobody has judged yet.

When something changes

A reply makes a conversation’s waiting, importance, tasks and topics pending again, but not the receipts in the messages before it. Changing an element’s instructions makes it a new question: earlier answers stay, marked as answers to the old one, and everything is read again.