AI content quality gates: what Kompozy checks before autopilot ships a post

Published 2026-04-24 · Updated 2026-10-23

Most writing about AI quality gates describes an ideal setup. This page describes the one Kompozy runs today, read from its generation and autopilot code, including the checks it does not have.

Quality control happens at different points for different formats. Some of it is instruction, meaning rules written into the prompt. Some is checking, where code or a second model reads the finished output, and the rest is plumbing at publish time. Knowing which is which tells you where you can lean on autopilot and where you still need to read the post yourself.

Why does it matter whether a check is a prompt rule or an output check?

A prompt rule tells the model what to do. An output check looks at what the model did. Rules shape most outputs and still slip on some, because the model weighs each rule against everything else in its context. A check reads the finished text and either passes it or sends it back.

Kompozy uses both. Prompt rules are cheap, so every text format gets them. Output checks cost an extra model call or some code, so they run on the formats that have failed in recognizable ways before. Here is where each layer sits:

What do the prompt rules cover on every format?

Each text generation prompt is assembled from blocks. An admin compliance matrix decides which blocks each format gets, separately for first generation and for regeneration. By default every format gets these:

• Persona Brief. The brief you write in prompting settings goes in as its own section. If the brief is empty, the section is left out and generation still runs, so an empty brief gets you a default voice rather than a blocked render.
• Voice profile. Tone, personality, point of view, complexity, formatting preference, emoji rule, and your Prohibited Keywords field, which the model sees as a list of prohibited words and phrases.
• Anti-AI-tell rules. A fixed list that bans em dashes, stock openers and words like delve and leverage. It also tells the model not to invent dollar amounts, figures, dates, places or anonymous sources that are not in the source material.
• Brand identity, the format prompt, and rotations for angle, content type and hook. The topic pool and script scaffold are on by default only for scripted video.

All of these are instructions. Nothing in the text pipeline scans a finished post for your prohibited keywords or compares its numbers with the source. If you are about to turn on autopilot for text formats, that is the thing to know.

What extra checks does the email newsletter get?

Email Newsletter gets more checking than any other format. Three layers run inside the generation worker:

1. Topic refusal. When a topic sits too far from your brand and audience, the model can return a refusal instead of a draft. The worker refunds the credit and marks the row failed with the reason, so you see a topic mismatch instead of a stretched newsletter.
2. Structural validator. Code checks the draft for the first-name merge tag in the salutation, a P.S. that opens with an action verb, a body link where the archetype needs one, enough named examples in framework issues, a bridge section for tangential topics, a word ceiling, and a protagonist who fits the topic. A failing draft gets one rewrite with the failures listed. The rewrite is kept only if it passes.
3. LLM grader. A separate model call, with no view of how the draft was made, scores it from 0 to 100 against a newsletter rubric. Below 85, the worker regenerates once with the grader feedback and keeps whichever draft scored higher. If the grader call fails, the current draft ships.

The grader judges structure and style. It checks that before-and-after numbers use matching time windows, but it does not check numbers against your source.

What checks run on persona video scripts?

Persona Short (HeyGen and Kling) and Persona Frames run a critic pass once the script is written. A second model call scores the script from 0 to 10 on hook, pace, specificity and AI tells, where tells include stub sentences, meta-labels and invented dollar figures. If any score is under 6 and the critic sent back a revision, the revision replaces the draft. This happens once. If the critic call fails, the original draft ships.

Before the avatar voice reads the script, a code sanitizer strips bracketed section labels, stage directions, speaker tags and line-opening labels like "Hook:" so none of them get spoken.

What does autopilot check before it schedules and publishes?

Autopilot is two switches per workspace. Auto-Approve Raw Inbound sends new source items to generation. Auto-Publish Generated approves finished rows and puts them on the calendar. A scheduled job runs every minute and processes each workspace that has a switch on.

With Auto-Publish on, each generated row is approved and given a date and time from your posting days, per-day caps and preferred times. This step does not review content. If a format has no posting days or a posts-per-day limit of zero, the row stays in Generated with a note explaining why. Batch-generate output is never auto-approved; it waits for you in the pipeline.

At publish time the worker trims text to each platform character limit, caps how many images a platform gets, removes the characters YouTube rejects, and fails a video that runs longer than the platform allows, with a message saying so. Posts that need media and have none are not scheduled.

What happens when a check fails?

• A refused newsletter topic, or a generation every provider failed on, marks the row failed and refunds the credit.
• A newsletter that fails the validator or scores under 85 gets one more attempt, and the better version ships. It does not go to a review queue.
• A persona script that fails the critic is replaced by the critic revision, once.
• A video over a platform duration limit fails that platform post with an explanation of the limit.

Kompozy has no retry-three-times-then-review loop. If you want to see every output before it publishes, leave Auto-Publish Generated off and approve rows in the pipeline yourself.

What do these checks not catch?

• A prohibited keyword the model used anyway, on any format. The field is a prompt instruction, not an output filter.
• An invented number or quote in a text post, carousel, blog or newsletter. The anti-fabrication rules lower the odds, and nothing compares the output with the source.
• Tone that drifts from the brief in ways the rules do not name.
• Strategic fit: the right voice with the wrong message for the week.

This is why the ramp matters. Run the sources you plan to automate with Auto-Publish off, read what comes out, and tighten the Persona Brief and prohibited keywords until you would ship most outputs untouched.

What does this look like in Kompozy?

A Quote Graphic generated in the Kompozy workspace: an SVG quote card, a short-form format shaped by prompt rules and platform limits rather than an output check.

Who is this for, and who should skip it?

Best for:

• Teams about to turn on auto-publish for a workspace
• Operators who want to know which formats get checked before they publish
• Anyone testing whether a tool's autopilot claim is backed by real checks

Not ideal for:

• Regulated content, where no automated check replaces compliance review
• Anyone expecting automated checks to judge tone, timing or strategy
• Readers who need a guarantee that a banned word or invented stat never ships

Frequently asked questions

How do we know this?

We checked every product claim on this page against Kompozy's generation, autopilot and publishing code on 2026-10-23, and plan details against kompozy.io/pricing. Descriptions of a fact-anchor gate, a word-matching brand-safety gate, a Persona Brief block and retry loops were removed because the code has none of them.

Sources

1. Kompozy — Glossary: Quality gates (read 2026-10-23)
2. Kompozy — Pricing (plans, credits, workspaces, autopilot, HeyGen BYO) (read 2026-10-23)

Which other autopilot guides should you read next?

• AI content autopilot: what it actually means in 2026 — Definition, mechanism, failure modes, and the brutal honest assessment of which AI tools actually achieve true autonomy.
• The 14-day manual-to-autopilot ramp methodology — How to graduate from full manual review to fully autonomous shipping in 14 days without losing voice control.
• When NOT to use autopilot: regulated industries and compliance risk — Medical, financial, legal, and pharma require human review forever. Why autopilot is a productivity tool, not a compliance shield.

Which related topics connect to autopilot?

• AI Brand Voice & Persona — Without a Persona Brief, every AI output averages to the LLM default voice. This is the 5-section methodology that makes 100+ AI-generated posts feel like one human author wrote them.

← Back to Autonomous Content Creation overview · Get started →