Content Quality Gate

Content Quality Gate

Weak commercial content is rarely badly written. It is usually well written and addressed to nobody: a competent paragraph about a category rather than a person, opening on the firm rather than the reader, resting on a number whose origin nobody can name. It gets approved because there is nothing wrong with any individual sentence, and it produces no conversations, which is then explained as a channel problem.

Two costs follow. The first is quiet: a publishing calendar that consumes a day a week and generates nothing, defended for a year because nobody can say which piece failed or why. The second is sudden. A prospect checks a statistic, finds it unsupported, and the credibility that took eighteen months to build is spent on a post that took forty minutes. The gate exists because both costs are avoidable by asking nine questions before the piece leaves the building, and because "looks good to me" has never caught either one.

When to use this, and when not to

Use it on anything a buyer, a prospect, a client or a public audience will read: a post, an email sequence, a newsletter issue, a blog article, a landing page, a one-pager, a case study, a conference abstract, a proposal narrative, the text on a deck. Use it in drafting mode before writing, and in review mode on a finished draft. Use it especially when another skill produced the draft, because a fluent draft is the one least likely to be challenged.

Do not use it as a writing skill. It decides; it does not compose. The format skills write, and this gate judges what they produced. Do not use it on internal documents that are not selling anything: a status note, a runbook, an internal memo. Do not use it to settle a disagreement about strategy, because a gate that is asked to adjudicate whether the belief is the right belief will always return a fail and never say why.

Adjacent cases belong elsewhere. human-voice-editor does the line-level repair that check 8 only diagnoses. breaktalk-brand holds the visual identity rules this gate collapses into a single pass or fail on branded assets. linkedin-post-writer and newsletter-post-writer produce the drafts this gate judges. Where the piece is a client proposal, proposal-writer carries the stricter standard and this gate runs afterwards on the narrative sections only.

What you need before starting

The draft, complete. A gate run on a half-draft produces fails that were going to be fixed anyway and teaches the writer to ignore the gate. Missing: run in drafting mode instead, which resolves checks 1, 3, 4 and 7 before a word is written, and note that the piece has not yet been gated.

The standard file for the organisation. Personas, beliefs, the named method, terminology rules, approved claims, voice rules. Without it, checks 1, 3, 4, 6 and 7 can only be run in outline. Missing: build it in the same session, in about ninety minutes, from existing sales decks and the last five approved pieces, and mark the run as provisional until someone with authority confirms it.

The piece's job and its funnel stage. Awareness, nurture, or conversion. This decides what check 9 will accept as a legitimate ask, and it is the input most often absent. Missing: infer it from where the piece will be published, state the inference at the top of the verdict table, and let the writer correct it.

The intended reader, by name or by persona. Not a segment, a person you could picture. Missing: this is check 1 failing before the gate starts. Return it to the writer with that finding rather than guessing a persona and grading against your own invention.

The source for every number in the draft. A workbook cell, a client-approved case study, a published report with a date. Missing: mark each unsourced figure and route it to whoever can approve it. Do not approximate, do not round to something safer, and do not delete it silently, because a quiet deletion teaches the writer nothing.

The name of whoever can approve an exception. Some fails cannot be fixed by the writer, for instance a claim needing a client's sign-off. Missing: assume no exception authority exists, which makes an unsourced number a hold rather than a ship.

The method

1. Set the mode and record the piece's job. Drafting mode resolves checks 1, 3, 4 and 7 before composition, because those four are premise questions and answering them afterwards means rewriting rather than editing. Review mode runs all nine on a finished draft. The judgement call is which mode a request implies. Rule: if a draft exists, review mode, even when the request is phrased as "help me write this", because the draft already contains the premise and it is cheaper to test it than to discuss it.

2. Read the whole piece once without marking anything. The purpose is to form the impression a real reader forms, which you can only do once. Write one sentence afterwards: what a reader would say this piece is about. Where that sentence and the writer's intent differ, note it, because the gap usually predicts which check fails.

3. Load the standard. Open the standard file and keep the never-say list and the approved claims list visible. Judgement call: what to do when the standard is silent on something the draft asserts. Rule: silence is not permission. Treat an unlisted claim as unsourced and an unlisted term as ordinary language that must be defined on first use.

4. Run checks 1 to 5 in order, and stop at the first premise-level fail. These five are about what the piece is, not how it is written. A fail on 1, 3 or 4 means the remaining checks are grading a piece that should not exist in this form. Rule: on a premise-level fail, stop the pass, report that one check with the fix, and return the piece. Do not deliver nine annotated fails, because a writer who receives nine fails fixes the easy four and re-submits.

5. Run checks 6 to 9. These are mechanical and cheap. Terminology and proof are binary and admit no judgement; voice and next step admit a little. Run them even when 1 to 5 passed comfortably, because the most dangerous piece in any pipeline is the one that is strategically correct and carries one invented figure.

6. Classify every fail by fix level: line, section or premise. A line fix is a sentence rewritten in five minutes. A section fix rewrites a block and changes the shape of the argument. A premise fix returns to the brief. The classification converts nine yes or no answers into an ordered piece of work.

7. Assign the verdict. Ship when all nine pass. Rewrite when any check fails and the fix is within the writer's control. Hold when the only outstanding fail requires a decision or an approval the writer cannot give, typically an unsourced number or a client case awaiting permission. The distinction between rewrite and hold matters because a piece marked rewrite goes back to the writer and a piece marked hold goes to whoever can unblock it, and confusing the two is how drafts sit for three weeks.

8. Write the table with quoted evidence in every row. Every pass names the sentence that earns it. Every fail names the sentence that breaks it and the specific replacement. Judgement call: how specific a fix has to be. Rule: specific enough that the writer could paste it. "Sharpen the opening" is not a fix. "Open on the moment the operations lead finds the variance, not on our founding year" is.

9. Order the rewrites by fix level, premise first. Never by check number. A writer working the list top to bottom should never do work that a later item makes redundant, and that is the single most common waste this gate produces when the list is unordered.

10. After the rewrite, re-run the affected checks plus 1, 2 and 8 always. A rewrite changes the opening, the emphasis and usually the rhythm, and those three are the checks a rewrite silently breaks. A full nine-check re-run is warranted only after a premise fix.

The nine checks

1. Written for a named buyer. One persona, one segment, one trigger event. The problem is stated in words the reader would recognise as their own sentence, not as a category label. Pass test: name the persona, the segment and the trigger, and quote the problem as they would say it. Fail signature: the word "organisations" in the first two sentences, or a problem that would be true of any company in any industry. Fix level: premise.

2. Problem before authority. The reader is the protagonist and the firm is the guide. Pass test: cover the firm's name in the opening paragraph. Does the reader still see themselves? Fail signature: the first specific noun in the piece belongs to the seller, not the buyer. Fix level: section.

3. Anchored in a core belief. Every piece advances one of the two or three positions the organisation actually argues for, and the reader should feel it rather than find it in a closing line. Pass test: name the belief and quote the sentence carrying it. Fail signature: the piece is true, useful, and could have been published by any competitor. Fix level: premise.

4. Anchored in something the organisation actually does. The piece has to rest on one specific, repeatable practice rather than on the discipline in general, and a reader has to be able to see which one.

Where the organisation has a named method, the standard file lists its steps or components in their real order and the piece maps to one of them, using those names exactly as defined and in the defined order. A claim about a stalled deal maps to the step that handles a stalled deal; a claim about maturity maps to the component that measures it.

Where there is no named method, which is the ordinary case and not a deficiency, the anchor is the practice itself: the check the team runs before a proposal goes out, the questions asked in a first meeting and in what order, the threshold at which a deal is disqualified, the thing that is done differently here from everywhere else. Write the two or three of those into the standard file as the anchor list, one line each, and gate against that. An organisation that cannot name any is not failing this check yet, it is failing to have a position, and the finding goes to whoever owns positioning rather than to the writer.

Pass test: name the step, component or practice this piece supports, and say where it is written down. If the answer is the discipline in general, or a practice nobody in the organisation could state in one sentence, it fails. Fail signature: the piece recommends a category of activity, "invest in enablement", rather than a way of doing it. Fix level: premise. Where the piece is not selling anything, strike this check and say so, as the edge cases set out.

5. Shows the shift. Buyers move when standing still becomes intolerable, not when a solution sounds appealing. The old way is drawn against the new way and, where approved claims allow, the cost of the old way is sized. Pass test: point to the before and the after. Fail signature: an informative piece with no tension in it, which readers finish and do nothing about. Fix level: section.

6. Terminology and product fidelity. Proprietary terms carry their exact definition, capitalisation and category. The never-say list is checked word by word, including plurals and hyphenated variants. Pass test: scan the never-say list against the draft. One hit is a rewrite of that sentence, not a discussion. Fix level: line, unless the wrong term is load-bearing, in which case section.

7. Proof with integrity. Only claims on the approved list, attributed where they belong to a named client, with the date the figure was measured. Anything else is marked as needing a source and escalated. Pass test: open the source for every number. Fail signature: a figure that reads as though it were rounded for rhythm, such as a suspiciously clean forty percent. Fix level: line to hold.

Where there is no approved claims register, which is the common case, the absence is not a pass and it is not a reason to skip the check. Build the minimum version in the same session, from the piece in front of you: one row per number the draft uses, carrying the figure as written, the document and the location inside it where the figure comes from, the date it was measured, whether a client is named and whether that client has agreed to be, who approved it, and when it expires. Ten rows takes about twenty minutes and that table is the register from then on. Until a number has a row, it is unsourced, and the three options are the same as always: cut it, replace it with the mechanism rather than the magnitude, or mark it held and route it to whoever can source it. A figure that has appeared in an earlier published piece is not sourced by having been published, and a figure somebody remembers is not sourced by being remembered confidently.

8. Human voice. Specific, confident, unpadded, written at professional register. No generic business vocabulary on autopilot, no long dashes standing in for commas and colons, no emoji, correct typeface and palette on any branded asset. Pass test: read two paragraphs aloud. Does it sound like a person who has done the work, or like a summary of a category? Fix level: line, and hand to human-voice-editor when the count of tells exceeds roughly one per hundred words.

9. One clear next step. Exactly one primary ask, matched to the funnel stage: read or follow at awareness, subscribe or download at nurture, a conversation at conversion. Pass test: what is the single thing the reader should do, and is it unmissable? Fail signature: three asks in the final paragraph, which produce none. Fix level: line.

The filter in one line: right buyer, real problem, one belief, a real method, a visible shift, exact language, honest proof, human voice, clear next step.

The standard file

The gate is generic. The standard file is what makes it specific to an organisation, and it is written once, by whoever owns the positioning, then edited perhaps quarterly. Keep it as standard.md beside this skill. A reader could write theirs this afternoon from three sales decks, the last five approved pieces, and one conversation with whoever sells.

# Content standard: [organisation name]
Last reviewed: [date]   Owner: [name]   Exception authority: [name]

## Personas
For each, three to six lines:
- Title and segment: e.g. Head of Operations, mid-market distribution, 200 to 800 staff
- Trigger events: the moments that make them look for help, dated where possible
- Pain, in their words: a sentence quoted from a real call, not a paraphrase
- What they are measured on
- What they will not admit in a first conversation

## Beliefs
Two or three, each one sentence, each contestable. A belief a competitor
would also sign is not a belief, it is a platitude. Under each, one line on
what follows if it is true.

## The named method, or the anchor list where there is none
The framework the content must reflect. List the steps or components in
their real order with their exact names, one line each on what that step
produces. Note which steps the organisation is known for and which it is not.
Where there is no named method, list instead the two or three specific,
repeatable practices the content must anchor to, one line each, concrete
enough that a reader could do one on Monday. Check 4 gates against whichever
of the two this section contains.

## Terminology
| Term | Exact definition | Capitalisation | It is not |
Never say: a plain list of words and phrases, including plurals and
hyphenated variants. Typical entries: competitor names, an internal
codename, a claim of being the only or the first, a regulated word the
organisation is not entitled to use.

## Approved claims
| Claim as written | Figure | Source (document, cell, page) | Date measured | Client named? | Approved by | Expires |
Every number that may appear in external content lives here. Nothing else
may appear. An expired row is treated as unsourced. Where this table does
not yet exist, start it from the numbers in the first piece you gate; ten
rows is enough to be useful and the table grows one piece at a time.

## Voice
Register, sentence length range, person (first plural or first singular),
punctuation rules, the words banned for being autopilot vocabulary, and
whether long dashes are permitted. State the default call to action per
funnel stage.

Two rules keep the file honest. Every approved claim carries an expiry, because a client result from four years ago is a different claim today. And the never-say list grows only through incidents, which keeps it short enough to be scanned. Where no standard file exists and none can be produced, say so in the output header, mark checks 1, 3, 4, 6 and 7 as run in outline, and never present an outline run as a full pass.

Worked example

Situation. Harbourline Analytics sells inventory forecasting software to mid-market distributors. Their marketing lead drafted a 780-word blog article titled "Why demand forecasting matters in 2026" and asked for a review at 16:00 on a Tuesday, to publish Wednesday morning ahead of a trade event where three prospects were expected. The standard file existed and was six months old. The draft opened with two paragraphs on volatility in global supply chains, moved to Harbourline's founding and its 2023 platform rebuild, and closed with three calls to action: book a demo, download the buyer's guide, and follow the company page.

Task. Return a verdict inside an hour with a rewrite list the marketing lead could work overnight, or a clear statement that the piece should not publish. Good meant the writer knew exactly what to change and in what order, and that nothing shipped that a prospect at the event could challenge.

Action. The first attempt started with check 8, because the voice problems were the most visible: eleven instances of autopilot vocabulary, four long dashes, and a closing paragraph in the register of a press release. Forty minutes went into line edits. Then check 1 was run properly and failed at the premise: the article addressed "distribution businesses", never a person, and the pain quoted was a paraphrase of a category. The persona in the standard file was an operations director at a distributor of 200 to 800 staff whose trigger was the first stock-out that reached a named customer. None of that was in the draft. The forty minutes of line editing was discarded, because the rewrite changed every sentence that had been edited.

That wrong turn produced the ordering rule now in the method: run checks 1 to 5 first and stop at the first premise fail. Line-level work on a piece with a premise fail is always wasted.

The full run, done in order, returned four fails. Check 1 failed at premise level. Check 2 failed at section level: the first specific noun in the piece was Harbourline, in sentence three. Check 5 failed at section level, since the article described forecasting without contrasting the current way of working against anything. Check 9 failed at line level on three competing asks. Check 7 passed narrowly but produced one hold item: the draft claimed customers cut stock-outs "by up to 40 percent", and the approved claims list held a single row supporting it, a result of 31 percent measured at one client in March 2024, approved with the client named. The rounding to 40 with "up to" attached was not on the list.

The verdict was rewrite with one held item. Rewrites were ordered premise first: rebuild around the operations director and the stock-out trigger; open on the Monday morning the variance is discovered; contrast a spreadsheet reforecast with a rolling one; make the buyer's guide the single ask, since the article sat at nurture stage; and replace the 40 percent claim with the sourced 31 percent and its attribution, or cut the sentence if the client would not be named by 09:00.

Result. The marketing lead rewrote in about three hours that evening. The re-run covered checks 1, 2, 5 and 9 plus the mandatory 1, 2 and 8; check 8 caught two new long dashes introduced in the rewrite, fixed in four minutes. The client agreed to attribution at 08:40, so the 31 percent figure ran with the client named and the measurement date, and the article published at 10:15.

The commercial outcome is not fully attributable: two of the three expected prospects mentioned the article at the event, one downloaded the guide. What is certain is that the 40 percent claim did not go out, and the standard file gained a terminology entry forbidding "up to" before any approved figure, since the phrase turns one measured result into an implied ceiling.

A second scenario, where it goes differently

Three weeks later the same gate was run on a renewal one-pager for an existing client, Marlowe Freight, going to a procurement lead who had inherited the account. Same nine checks, different behaviour on four of them.

Check 1 tightened: the buyer was one real person with a name and a start date, so a generic persona would not have passed even in outline. Check 2 inverted, because Marlowe already knew who Harbourline was and opening on the reader's world without acknowledging an eighteen-month relationship read as evasive; the pass test became whether the first paragraph named the shared history in the reader's terms. Check 5 changed source: the shift was Marlowe's own stock-out rate before and after, not an industry contrast, so every figure came from the client's reporting rather than the approved claims list. That made check 7 stricter, not looser, since a client's own number quoted back to them gets checked against their system and any discrepancy reads as carelessness about their account. Check 9 was unchanged: one ask, a thirty-minute renewal conversation with two proposed dates.

The verdict was hold rather than rewrite, for a reason the first scenario never produced: two of the four client figures came from a workbook last refreshed in July, and the account team could not confirm they matched Marlowe's current reporting. The piece sat two days until they were reconciled, which was correct, because a renewal document that misstates a client's own data does more damage than one sent late.

What changed: the reader knew the sender, so the proof standard moved from approved for external use to reconciled against the reader's own system, and the failure mode moved from overclaiming to being wrong about facts the reader already holds.

Output

Always render the table filled in, with quoted evidence in every row and a specific fix beside every fail. Never leave a row saying only "pass".

Header block first:

CONTENT QUALITY GATE
Piece:         [title, format, word count]
Job:           [what it is for]      Funnel stage: [awareness / nurture / conversion]
Reader:        [persona or named person]
Standard file: [loaded / absent, checks run in outline]
Verdict:       SHIP / REWRITE / HOLD

Close with the ordered rewrite list, premise fixes first, each one line and each actionable by one person. Then the held items with the name of the person who can release them and what decision they owe. Then, on a re-run, the list of checks re-run and their results.

Failure modes

Starting with the voice. Recognise it when the first twenty minutes produce line edits and the premise has not been tested. Line work on a piece with a premise fail is discarded entirely. Run checks 1 to 5 first, always.

Answering a check with an adjective. "Strong opening" is not a verdict. Recognise it when a row contains no quoted text. Quote the sentence or mark the check unrun.

Nine fails delivered at once. A writer who receives nine fails fixes the four cheapest and re-submits. Recognise it in the re-submission, which will still fail check 1. Stop at the first premise fail and send one finding.

Approving a number because it appears in an older published piece. Prior publication is not a source. Recognise it when the trace ends at another marketing asset rather than at a measurement. Trace to the workbook, the report, or the client sign-off.

Rounding for rhythm. 31 becomes "roughly a third", then "up to 40 percent" two drafts later. Recognise it when a figure is prefixed by "up to", "as much as", or "over". Restore the measured number and the date.

Passing a piece because it is late. Recognise it in the sentence "this is fine for now". Deadline pressure changes the verdict from rewrite to hold and back again in the writer's head. It should change nothing. If a piece must go out unfinished, that is a decision by the exception authority, recorded, not a quiet pass.

Grading against your own taste. Recognise it when a fix rewrites a sentence into your voice rather than the standard's, or when the same check keeps failing for the same reason across unrelated pieces, which usually means the standard is stale rather than the writers careless. Fixes cite the standard; preferences do not, and a stale standard goes back to its owner.

Edge cases

No standard file and no time to build one. Run checks 2, 5, 8 and 9 fully, since they need no organisational specifics, and run 1, 3, 4, 6 and 7 in outline against whatever the writer can state verbally. Mark the run provisional in the header. Never let an outline run be recorded as a pass.

The piece is not selling anything. A recruitment post, a community note, a conference abstract. Checks 4 and 5 do not apply; strike them and say so rather than forcing an artificial method anchor. Check 1 still applies, because a recruitment post written to nobody fails for the same reason a blog does.

Two legitimate audiences. An announcement read by both customers and candidates, for instance. Pick the primary audience, gate against it, and check the secondary audience is not repelled. Where both need their own primary ask, that is two pieces.

Ghostwritten under someone else's name. Check 8 grades against that person's voice, not the house voice, so two or three samples of their real writing are needed before the gate runs. Without samples, mark check 8 unrun.

A number that is true but not approved. For example, a result the delivery team measured last week. It is not a fail of accuracy, it is a fail of process. Verdict is hold, routed to the exception authority, and if it clears, the claim is added to the approved list with its source and expiry so the next piece does not repeat the loop.

Legal or regulated claims. Anything about compliance, certification, safety or a regulated status leaves this gate entirely. The verdict is hold and the routing is to whoever carries that responsibility. Do not attempt a judgement on wording.

A piece already published. The gate still runs, but the verdict set becomes leave, amend or retract. Retract is reserved for an unsourced claim or a never-say term; everything else is amended in place, with a note where the change is material.

Quality bar

• Every check row quotes text from the piece; not one row is answered with an adjective.
• Every number in the piece traces to a source with a date, or appears in the held items list.
• Checks 1 to 5 were run before any line-level edit was made.
• Every fail carries a fix specific enough to paste, and the rewrite list is ordered premise first.
• The verdict is one of ship, rewrite or hold, and a hold names the person who can release it.
• The never-say list was scanned word by word, including plurals and variants.
• After a rewrite, checks 1, 2 and 8 were re-run and the results recorded.
• Where the standard file was absent, the header says so and the affected checks are marked outline.

Adapting this to your context

The nine checks are general. What makes them bite is the standard file, and that file assumes named buyer segments, a position the organisation argues for, and someone who can approve a claim. Several of those are commonly missing.

• The named method. Check 4 no longer assumes you have one. If you do, list its steps and use those names exactly. If you do not, write the two or three specific practices you actually anchor to and gate against those. Never invent a framework to satisfy a check.

• The approved claims register. Where none exists, build it from the piece you are gating, ten rows at a time. A grant-funded or public body usually has the opposite problem, published evaluation it may quote under rules, so the register records the permitted form of words alongside the figure.

• The nine checks as a set. Checks 4 and 5 are struck for anything not selling. A solo practice ghostwriting runs check 8 against that person's samples. Regulated claims leave the gate entirely.

• The exception authority. Decide who can release a held item before the first gate runs, or every hold becomes an indefinite pause.

• What not to change. Run checks 1 to 5 before any line edit, and never record an outline run as a pass.

Related skills

linkedin-post-writer, newsletter-post-writer and discovery-to-proposal-deck produce drafts this gate judges; run it last, after they finish. human-voice-editor performs the repair that check 8 only diagnoses, and returns tell counts you can record in the re-run. breaktalk-brand owns the visual identity rules that check 8 collapses into a single pass or fail on branded assets. proposal-writer and sow-and-scope carry their own stricter standards for client-facing commercial documents; gate their narrative sections only, and never let this gate override a contractual wording decision.