Skip to content

Guide

Troubleshooting

Symptoms, in the words you would use to describe them, with the usual cause and where to read more. Most problems here are working-as-intended behaviour that looks like a fault.

Check these first

Three questions that resolve most reports

Which workspace are you in? Nothing is shared between workspaces. Work that has “disappeared” is usually in the other one.

What is your role here? Roles differ per workspace. You may genuinely be a viewer in this one.

Did the run actually finish? History shows every attempt, including the ones that failed and produced nothing.

Content and generation

SymptomUsual cause
A piece I generated is not in my libraryThe run failed, or it was a voice sample — samples never create a piece. Both appear in History.
The writing does not sound like usCheck the run used the voice you meant; a run with one-off instructions ignores voices entirely. If it did, the voice itself needs work — Voices.
Brand fit is low on every pieceThe voice describes writing you do not actually produce. Re-capture from a better source.
Brand fit is high on some topics, low on othersThe brand voice is over-fitted to one subject. Broaden it, or split it in two.
The piece is the wrong lengthWord count is a target, not a contract. Set it explicitly if the length matters.
It invented a factIt was not given the fact. Put real numbers, names and dates in the brief — it will not source them for you.
A run failedThe reason is recorded against the run. Most failures are transient and retrying works; a brief that fails repeatedly needs rewording. See Generating.

Brand voices

SymptomUsual cause
A brand capture failedThe source could not be read — commonly a site that blocks automated readers, or an address that does not resolve. Try a specific article rather than the homepage.
The captured brand voice describes us badlyIt read sales copy rather than your writing. Add real examples as extra material and capture again.
I cannot edit a brand voiceEditing voices needs a manager or admin role — Team and access.

Delivery

SymptomUsual cause
Blocked before sendingThe address resolves somewhere BrandKind will not send — a private network, or an unusual port. Use a public address.
The endpoint returns 401Usually no secret is saved. A destination set to bearer or custom header with no secret sends no credential at all.
Nothing arrives, and nothing failedThe destination is paused. Attempts are recorded as failures with the reason stated — check the delivery log.
Deliveries sit at queuedA delivery budget was reached. It resumes automatically — quotas.
The same content arrives twiceYour endpoint is not deduplicating on the delivery id. Retries are expected — retries.
Images break the next dayThe image link expires after 24 hours. Receivers must download the image, not store the link — the image object.
The credential is missing on arrivalYour infrastructure is redirecting across origins, which drops it. Register the final address directly.

Access and people

SymptomUsual cause
A control is there but refuses meYour role does not allow it. The message says who can — that is deliberate, so you know whom to ask.
An invitation link says it has expiredInvitations time out. An admin can reissue it.
I can do something in one workspace but not anotherRoles are per workspace. This is working correctly.
Someone left and their work is unreachableTheir private pieces stay private. Ask people to share what the team needs before they go.

Schedules

SymptomUsual cause
A job stopped runningIt ran out of topics and paused itself, showing that as the reason. Add topics and re-activate it.
It ran at the wrong timeThe time is in the job’s own time zone, not the reader’s. Check the job’s setting.
Credits are draining unexpectedlyA forgotten schedule. A daily job is roughly thirty runs a month, plus delivery if configured — Scheduling.

Scores that look wrong

  • A piece passes the publish check but sits in the middle band in the library. The two use different cut points on purpose — the check judges one piece, the filter groups many. Neither is a mistake; see the thresholds.
  • The verdict is bad but every score looks fine. The verdict is the worst result among the checks, not an average. One failing check fails the piece.
  • A high Human score on writing that reads badly. That score measures how strongly the text reads as machine-generated, not quality. See Human score.
  • Reading age looks wrong for our audience. It is shown for information and never affects the verdict. Decide for yourself whether it suits your reader.
  • Run counts and piece counts disagree. Failed runs and voice samples produce no piece. Both are counted as runs.

Still stuck

Contact RoodtSquared at info@roodtsquared.co.uk. The details that make a report quick to answer are the workspace, roughly when it happened, and what you expected instead. You do not need to reproduce the problem — the run record is already there.