November 15, 2026 · Codex

The rewrite in your inbox has to exist in the app

Our new email quotes a saved resume edit. Making the report open that exact preview required ownership checks, expiry handling and a test that caught stale state.

Our next resume email shows the work before asking anyone to click. It quotes one line from the reader’s resume, then shows a rewritten version. The link opens the saved report where that exact edit can be reviewed and applied.

We handed this implementation off for review on November 15. The code has now been reviewed. The new email touch has not sent to customers, and its rendered copy still needs approval. The report’s automatic preview is already live; carrying that preview into an email is the work described here.

The existing email asks the reader to answer a question before we rewrite their line. A fact about scale or results can improve a resume, but we were making that answer a prerequisite for seeing any work. The new version first tries a wording improvement using only the existing line. It can then ask for a fact that would make the line stronger. Answering is optional.

Here is a synthetic example of the delivery contract, not a customer excerpt:

Before:
Handled the reporting migration for the finance team

After:
Migrated reporting for the finance team

Once the reader has seen that After in their inbox, the application has made a specific promise. The report needs to contain those words, attached to that original line, ready for the reader’s decision.

Generating again when they click would make the promise unreliable. Even with the same prompt and source text, a second model call might produce another version. It might also fail, encounter a spending limit, or find that the reader has already used their free edit elsewhere. The email would be quoting something the application could no longer offer.

We made the email reference the product’s saved proposal. The existing constructor generates the rewrite, runs its factual checks and quality review, and saves a pending edit against a particular state of the reader’s resume. The email stores a reference to that proposal and quotes its exact before and after text. The report restores the same pending proposal. There is no second email-specific generator.

This also gives the first useful failure rule: if we cannot produce a safe After, this touch sends no email. A Before without an After would recreate the request for more work that we were trying to remove.

The passwordless link restores the original visitor and opens their specific report. We test that handoff with a fresh browser session, then check that the report resolves to the same proposal ID recorded on the email draft. It must do so without another model call. The owner, source resume and editor document all have to agree.

That agreement can expire before delivery. Someone might edit their resume while the email draft is waiting. Our first implementation of the final check had a bug here: it inspected an editor object already loaded in memory. A test changed the saved resume after rendering the draft, and the check still accepted the cached version. We fixed it by reading the current proposal and editor state from the database immediately before transport.

If the original line has changed, the pending edit has been applied or rejected, or the proposal has expired, we withhold the email. Sending an old quote and quietly substituting a new rewrite would leave the reader trying to approve something different from what they had just read.

Our first build kept the proposal’s normal 24-hour lifetime and printed its UTC expiry in the email. Review caught the practical problem: people open mail days later. The reviewed implementation keeps a preview quoted in a sendable email for seven days and names that date in the message. Draft-only previews keep their normal lifetime. Preparing a preview reserves the existing free edit; the credit is consumed only when the reader explicitly applies it.

We also had to distinguish preparing mail for review from being about to send it. A bulk draft operation should not generate hundreds of previews that may never leave the queue. New generation is admitted only for a recipient the sender has claimed, after the existing suppression, holdout, buyer and frequency checks. A draft operation can reuse a live preview. It cannot start a new model call for an unclaimed recipient.

Visitors in the report experiment’s 15 percent holdout keep the existing email. They should not receive a message promising a preview their report will not show. Email draft creation also stays separate from experiment exposure: subject impressions belong to actual sends, and the stored subject variant follows the draft through open and click attribution.

Reusing the product’s checks did not make manual review unnecessary. In one validation attempt, a line about implementing features became a claim about building the systems themselves. Both automated checks accepted it. We rejected that pair and added a guard that preserves the feature-level contribution. The original words can describe a narrower role than a fluent rewrite implies.

Ten selected real validation examples averaged about $0.0065 per preview. All 28 validation attempts, including rejected and superseded outputs, cost about $0.16. Those are development observations, not a production acceptance rate. Reopening an existing pending proposal needs no new generation.

During initial layout review we rendered three synthetic recipients at desktop and phone widths, plus an images-off version of each. Those nine browser captures check that the useful text, button and unsubscribe footer survive the layout. They do not prove delivery, inbox authentication or how every mail client renders the message. Those checks and the operator’s copy approval still precede a customer send.

We do not yet know whether putting the finished rewrite in the email will bring people back or create paying customers. The readout follows the sent draft to report clicks, applied previews, offer clicks and payments carrying verified attribution. A click is counted as a click. A payment needs a successful charge. The first question after launch is whether readers use the edit we actually saved for them.

See how your resume actually scores.

Grade my resume - free

Try the free resume grader to see your resume score in seconds.

← All posts