Get them back: 10 AI skills for the list you already own
subject-lines
twenty written, one sent
How the two work together
Claude thinks it through. Paste the Claude prompt into Claude Code, or drop the folder into your skills folder. Claude does the judgement: what to look for, what is worth doing, what is right.
Codex gets it done. At the hand-off point Claude runs Codex on your machine with one command and passes it the Codex prompt. Codex does the mechanical part and hands the result back. Claude checks it before you see it.
No API key to set up: Claude calls the Codex you already have installed. If Codex is not installed, Claude does that half itself and tells you.
Prompt for Claude
---
name: subject-lines
description: Writes twenty subject lines for one email, scores them against published advertising rules and mailbox provider requirements, kills the ones that would breach, and hands you one to send with the reason. Use when you have an email written and no idea what to put in the subject box.
---
# Twenty subject lines written and scored, one sent, and the reason in writing
You give this the email you have already drafted and one sentence about who it goes to. You get back twenty subject lines in five different families, each scored, each with its preview text, the ones that would breach the advertising code struck out with the rule that kills them, and one recommendation with the reason and a runner-up. It never writes a claim you cannot evidence, and it never invents an offer that is not in the email.
## What it does
1. **Write down what the email actually says before writing a single subject line.** In one sentence: what the recipient is being told, what they are being asked to do, and what is genuinely true about timing. Then list every factual claim in the body: prices, dates, "new", "back by popular demand", "the only", "award-winning", numbers of any kind. The subject line has to be true against that list, because the ASA treats the subject line as part of the advertisement and rules on "the impression created by marketing communications as well as specific claims". A subject line written before the email is a subject line written about an email that does not exist yet.
2. **Generate twenty lines across five named families, four in each, so the set is genuinely different rather than twenty rewrites.** The families: **plain fact** ("The autumn menu starts Tuesday"), **named person** ("Sam's put the venison back on"), **specific detail** ("Three new dishes, one of them is the pie"), **direct question** ("Are you free on the 20th?"), and **useful information** ("Parking is free after six from Monday"). Four per family forces range. Twenty is the right number because the first six are always the obvious ones and the interesting one usually arrives somewhere in the teens. Keep each line under about fifty characters where you can, because a phone truncates, and write the preview text for each one at the same time rather than afterwards.
3. **Score every line out of twelve on six things, two points each, and show the scores.** True against the body of the email. Specific to this venue rather than generic to any venue. Clear on its own, without the email open. Readable at phone width without being cut mid-word. Different from the last three subject lines this list received. Says something, rather than teasing that something exists. A line scoring two on all six is rare; a line scoring zero on the first is disqualified regardless of the other five, because a subject line that is not true is not a subject line, it is a complaint waiting to be typed.
4. **Kill every line that makes an objective claim you cannot evidence today.** CAP Code rule 3.7 is the one that removes most of them: "Before distributing or submitting a marketing communication for publication, marketers must hold documentary evidence to prove claims that consumers are likely to regard as objective and that are capable of objective substantiation. The ASA may regard claims as misleading in the absence of adequate substantiation." Best in town, award-winning, the most popular, voted number one, locally sourced, the original: each needs a document in a drawer before it goes in the subject box. Rule 3.2 keeps obvious puffery lawful, so "the best pie in the universe" is fine and "the best pie in Hackney" is not, because the second is checkable and somebody will check it.
5. **Kill every line whose overall impression is wrong even though every word in it is true.** This is the failure the ASA rules on most often with emails, and there is a clean example. In its ruling on Maki Online Services Ltd, published 26 February 2025, the ASA considered an email with the subject line "Your BRP Card expires next month". The card did expire. The ASA upheld the complaint anyway: "Because we considered that the overall impression of the ad was that the immigration status of BRP card holders would change when the card expired, whereas in the majority of cases this was not likely to be the case, we concluded the ad was misleading." Applied to a restaurant: "Your table is at risk" when the booking is fine, "Your account needs attention" when it does not, and "Final notice" on a newsletter are all technically-true-sounding lines that create a false impression. Cut them.
6. **Kill every line that uses the mechanics the mailbox providers name as deceptive.** Google's sender guidelines say "Email message subject, headers, display names, and other message elements should accurately represent the sender identity and message content, and shouldn't be misleading. For example, don't send messages with subject lines starting with Re: or Fwd: unless the messages are actual replies or forwards". The same page says "Don't use emojis or other non-standard characters to imitate graphic elements in messages, with the intent to deceive or influence recipients." So: no fake Re:, no fake Fwd:, no fake forwarded thread, no emoji pretending to be a verification tick or a badge, no all-caps shouting, and no display name that is not the business. These do not merely risk a complaint, they risk the message being filtered, which takes the whole send with it.
7. **Kill every line with a deadline, a scarcity claim or a "free" that the email cannot support.** Rule 8.22: "Promoters must not claim that consumers must respond by a specified date or within a specified time if they need not." Rule 3.30 prohibits falsely stating that a product or its terms "will be available only for a limited time to deprive consumers of the time or opportunity to make an informed choice". Rule 3.23 bars describing something as "free" if the consumer has to pay anything beyond the unavoidable cost of responding and collecting. And rule 8.21.1 prohibits falsely claiming or implying "that the consumer has already won, will win or will on doing a particular act win a prize". "Last chance", "ends tonight", "only 3 tables left" and "you've won" are all fine when they are true and are all breaches when they are not, and the subject box is where they are most often written without anyone checking.
8. **Read the surviving lines for tone and offence before you read them for cleverness.** A clever subject line can be a breach on its own. In its ruling on Koi Footwear Ltd, published 18 June 2025, the ASA considered an email whose subject heading was "Take two and call us in the morning" with two pill emojis, and concluded that the subject line "referenced medical dosage instructions in a way that would be familiar to consumers" and, with the email's imagery, "condoned and encouraged drug use, and was therefore irresponsible, and likely to cause serious and widespread offence". The advertiser's explanation that it was surreal and tongue-in-cheek did not save it. For a venue: no drink-driving jokes, no hangover-as-achievement, no weight or body comments, no "you deserve this" about alcohol, and nothing about a news event. The room your subject line lands in includes people you did not picture.
9. **Pick one, name the runner-up, and log what was sent.** Output a table of all twenty with score, family, preview text, and for the struck ones the exact rule that killed them. Under it: the recommendation in one line, why it beat the runner-up in one line, and the runner-up itself so the owner can overrule you in three seconds. Then append a single line to a running log: date, subject sent, list segment, and the last three subjects sent to that segment. The log is what makes criterion five in step 3 possible next month, and without it every list eventually receives four variations of the same sentence in a row.
## Then it checks
1. Exactly twenty lines were produced, four in each of the five named families, and no two lines differ only by a word.
2. Every line is true against the body of the email as drafted, and every factual claim in a surviving line is matched to a line in the email that supports it.
3. Every objective claim in a surviving line names the document that evidences it, or the line is struck under rule 3.7 rather than kept with a caveat.
4. No surviving line uses Re:, Fwd:, a fake reply or forward, an emoji imitating a verification mark or graphic element, or a display name that is not the business.
5. No surviving line states a deadline, a limit, a "free" or a win that the email and the venue cannot support, and each one that does state a deadline names the real reason for it.
6. Every struck line carries the specific rule or ruling that struck it, by number, and the recommendation carries a runner-up and one line of reasoning.
Any check fails: name it, redo that step once. Failed twice: say what is wrong and stop.
## Rules
- Public information only.
- Never invent a fact, a number or a quote.
- Anything sent in someone's name says whose name it is. A subject line written as though the chef typed it goes to the chef first.
- Never state an open rate, a click rate, a benchmark or a "subject lines with X perform better" claim. There is no published figure for a single independent UK venue's list, the numbers that circulate come from vendors selling sending software, and an owner who sets a target from a borrowed number will conclude their own list is failing when it is not.
- Never write a claim into the subject line that is not also in the body. A subject line is the part of the advertisement most people read and the part least often checked, and the ASA rules on the impression it creates whether or not the body corrects it.
- Never reuse a subject line that was struck. Struck lines have a way of reappearing next month because they read well, which is exactly why they were dangerous.
- Never test a subject line by sending two versions to a small list and declaring a winner. A few hundred subscribers cannot separate a real difference from noise, and acting on the result is worse than picking the one you can defend.
- This output is a working document prepared for the owner to check against their own menu, prices and offers before sending, and for their solicitor or trade body to check where it makes a comparative or superlative claim. It applies published advertising rules to draft copy; it is not a ruling that the copy complies.
## Built from
- Committee of Advertising Practice, CAP Code (Edition 12), Section 3 "Misleading advertising", https://www.asa.org.uk/type/non_broadcast/code_section/03.html, no publication date shown on the page, read 14 September 2026: the substantiation rule 3.7 in step 4, the puffery allowance at 3.2, the false-limited-time prohibition at 3.30 and the "free" rules at 3.23, plus the background statement that the ASA "will take into account the impression created by marketing communications as well as specific claims".
- Advertising Standards Authority, "ASA Ruling on Maki Online Services Ltd", complaint ref A24-1270475, https://www.asa.org.uk/rulings/maki-online-services-ltd-a24-1270475-maki-online-services-ltd.html, published 26 February 2025, read 14 September 2026: the worked example of a true subject line creating a misleading overall impression, which is the whole of step 5.
- Advertising Standards Authority, "ASA Ruling on Koi Footwear Ltd", complaint ref A25-1286163, https://www.asa.org.uk/rulings/koi-footwear-ltd-a25-1286163-koi-footwear-ltd.html, published 18 June 2025, read 14 September 2026: a ruling in which the email subject heading itself was the offending element, which is why step 8 reads for offence before cleverness.
- Google, "Email sender guidelines", Gmail Help, https://support.google.com/a/answer/81126, no publication date shown on the page, read 14 September 2026: the accuracy requirement for subject lines, headers and display names, the Re:/Fwd: prohibition and the emoji-as-graphic prohibition, all in step 6.
- Committee of Advertising Practice, CAP Code (Edition 12), Section 8 "Promotional marketing", https://www.asa.org.uk/type/non_broadcast/code_section/08.html, no publication date shown on the page, read 14 September 2026: rule 8.22 on false deadlines and rule 8.21.1 on false win claims, in step 7.
Prompt for Codex
# subject-lines ## You are given The finished body copy of one marketing email for a UK hospitality business, exactly as it will go out. The twenty subject lines already written, four in each of the five families `plain fact`, `named person`, `specific detail`, `direct question` and `useful information`, each with its preview text. One sentence saying which list or segment the email goes to. Any evidence documents behind claims in the copy, such as an award certificate, a supplier invoice, a menu or a booking report. The running subject log from previous sends, if one exists. And, once the judgement has been made, the chosen recommendation, the one line reason and the runner-up. ## Produce Write into a `./subject-lines-output/` folder: 1. `claims-in-the-email.csv` with these columns in this order: `claim_ref`, `claim_text_verbatim`, `where_in_the_body`, `claim_type`, `evidence_supplied`, `evidence_source_file`, `evidence_date`, `substantiated`. `claim_ref` is C001 upward. `claim_type` is exactly one of `price`, `date`, `superlative`, `award or accolade`, `provenance`, `availability or scarcity`, `free or no-cost`, `number`, `puffery`. `substantiated` is `yes` or `no`. 2. `subject-lines.csv` with columns: `line_no`, `family`, `subject_line_verbatim`, `characters`, `truncates_at_41_characters`, `truncates_at_60_characters`, `cut_mid_word`, `preview_text_verbatim`, `preview_characters`, `score_true`, `score_specific`, `score_clear`, `score_readable`, `score_different`, `score_says_something`, `total_score`, `matched_claim_ref`, `status`, `rule_or_ruling_cited`. Each score column is 0, 1 or 2 and `total_score` is their sum out of twelve. `family` is one of the five exactly as listed above. `status` is exactly `kept` or `struck`. Every `struck` row carries a non-empty `rule_or_ruling_cited`. 3. `mechanical-kills.csv` with columns: `line_no`, `subject_line_verbatim`, `trigger`, `what_was_found`, `rule_cited`. `trigger` is exactly one of `Re: or Fwd: prefix on a message that is not a reply or forward`, `emoji imitating a graphic element or verification mark`, `all capitals`, `display name is not the business`, `claim not present in the body of the email`, `claim present but not substantiated`, `repeats a subject sent to this segment in the last three sends`, `differs from another line by one word only`. 4. `set-composition.csv` with columns: `family`, `lines_supplied`, `lines_kept`, `lines_struck`. Five rows, one per family, plus a `TOTAL` row. Any family not holding exactly four supplied lines is listed in `gaps.md`. 5. `recommendation.md` - exactly three lines: the recommended subject line with its preview text, the one line reason as supplied, and the runner-up with its preview text. All three copied verbatim from the input. Written empty if no recommendation was supplied. 6. `subject-log.csv` with columns: `date`, `list_or_segment`, `subject_sent`, `preview_text`, `previous_1`, `previous_2`, `previous_3`. Appended to the existing log, never overwritten. The three previous columns are read from the log itself. 7. `gaps.md` - a numbered list of: any family without exactly four lines; any line with no preview text; any claim in a kept line with no matching `claim_ref`; any claim in the body with no evidence document; any line longer than the supplied character budget; and any struck line whose rule citation is missing. ## Rules - Codex measures, scores, records and logs. Never write a new subject line, never reword, shorten, correct or tidy a supplied one, and never replace a struck line with a suggestion. This is copy the owner sends in their own name. - Never score `score_true` above 0 for a line making a claim that has no matching row in `claims-in-the-email.csv`. A line scoring 0 on `score_true` is `struck` whatever its other scores are. - Never mark a claim substantiated without a supplied evidence document named in the row with its date. - Never write an open rate, a click rate, a benchmark, an industry average or a claim that one kind of subject line performs better. No such figure exists for a single independent UK venue, and none is to be written into any file. - Never reinstate a struck line, and never carry a struck line into `recommendation.md` or `subject-log.csv`. - Never send anything, never connect to a sending platform, and never run a split test. - Never write the recommendation yourself. It is supplied. Where it is missing, `recommendation.md` is written empty and the fact is listed in `gaps.md`. - Character counts are counted, not estimated, and include spaces and punctuation. Truncation columns are computed at the stated widths and say where the cut falls. - Every quoted line is copied exactly, including its punctuation, capitalisation and any emoji. - Use British English, £ and DD Month YYYY dates. No em dashes in any file you write, and any supplied line containing one is recorded verbatim and flagged in `gaps.md`. - Every file ends with this line: this is a working document prepared for the owner to check against their own menu, prices and offers before sending, and for their solicitor or trade body to check any comparative or superlative claim. It applies supplied rules to draft copy and is not a ruling that the copy complies. ## Return The absolute path of each file written, the row count of each CSV, the number of lines supplied and the number in each family, the number kept and struck with the trigger breakdown, the highest and lowest total scores, every claim in the body with no evidence document, the recommended line and the runner-up exactly as recorded, the three previous subjects read from the log, and the `gaps.md` item count.
Built from the best public work on this
Sources for subject-lines
Everything below was opened and read on 14 September 2026. Nothing is cited that could not be loaded.
1. Committee of Advertising Practice, CAP Code (Edition 12), Section 3 "Misleading advertising"
https://www.asa.org.uk/type/non_broadcast/code_section/03.html, no publication date shown on the page; its background notes that the ASA "will take the Unfair Commercial Practices (UCP) provisions in Chapter 1 of Part 4 of the Digital Markets, Competition and Consumers Act 2024 into account when it rules on complaints", read 14 September 2026.
Section 3 is the rule set a subject line is actually judged against in the UK, and it is short enough to read in a sitting. Four rules do most of the work in this skill.
Rule 3.7 is the killer: "Before distributing or submitting a marketing communication for publication, marketers must hold documentary evidence to prove claims that consumers are likely to regard as objective and that are capable of objective substantiation. The ASA may regard claims as misleading in the absence of adequate substantiation." Note the sequence: the evidence must exist before the thing is sent, not be findable afterwards. That is step 4, and it removes more subject lines than any other single rule, because superlatives are what people reach for when they cannot think of a specific.
Rule 3.2 is the counterweight that keeps the skill from being humourless: "Obvious exaggerations ('puffery') and claims that the average consumer who sees the marketing communication is unlikely to take literally are allowed provided they do not materially mislead." The dividing line the skill uses, between an unfalsifiable boast and a checkable local claim, comes from reading 3.2 and 3.7 together.
Rule 3.30 and rule 3.23 handle the two most common subject-line tricks. 3.30 says marketing communications "must not falsely state that a product, or the terms on which it is offered, will be available only for a limited time to deprive consumers of the time or opportunity to make an informed choice", and it is asterisked, which the section's background explains means it "reflects a prohibited practice" requiring no transactional decision test. 3.23 bars calling something "free" where the consumer must pay "anything other than the unavoidable cost of responding and collecting or paying for delivery of the item". Together they are step 7.
The background paragraph is quoted in step 1 because it settles a question owners ask: whether the subject line can be judged separately from the email. It cannot, and it is not: "The ASA will take into account the impression created by marketing communications as well as specific claims. It will rule on the basis of the likely effect on consumers, not the marketer's intentions."
Where the skill departs: Section 3 also carries a long comparisons regime for claims against identifiable competitors. The skill does not teach it, it simply advises against naming a competitor in a subject line at all. For an independent venue the upside of "better than the place across the road" is small and the downside is a ruling published under the venue's own name.
2. Advertising Standards Authority, "ASA Ruling on Maki Online Services Ltd"
https://www.asa.org.uk/rulings/maki-online-services-ltd-a24-1270475-maki-online-services-ltd.html, published 26 February 2025, complaint reference A24-1270475, read 14 September 2026.
A published ruling on an email, upheld on both issues, and the clearest available demonstration of the point step 5 makes. The ad is described in the ruling itself: "An email from Nino Apply, sent on 29 November 2024, with the subject line 'Your BRP Card expires next month' stated, 'ACTION REQUIRED: Don't Lose Your Right to Stay in the UK!'"
The subject line was factually accurate. Biometric Residence Permits did expire. The ASA's assessment is the part worth putting in front of an owner: "Because we considered that the overall impression of the ad was that the immigration status of BRP card holders would change when the card expired, whereas in the majority of cases this was not likely to be the case, we concluded the ad was misleading." The ad breached rule 3.1, and on the second issue rules 3.1 and 3.3 for omitting that the replacement could be obtained free from gov.uk.
This is why the skill has a kill step that operates after the truth check rather than being the truth check. A subject line can pass "is every word true" and fail "what does this make the reader believe", and the second is the test the regulator actually applies. The hospitality translations in step 5 are ours, not the ASA's, but the mechanism is theirs: manufacture a sense that something of the reader's is at risk, and the accuracy of the individual words will not save you.
Where the skill departs: this ruling is about immigration services, a high-harm sector where the ASA is predictably firm, and a pub's "last chance for Sunday lunch" is not in the same territory. The skill therefore uses the ruling for the test rather than for the severity, and does not imply a venue is at comparable risk.
3. Advertising Standards Authority, "ASA Ruling on Koi Footwear Ltd"
https://www.asa.org.uk/rulings/koi-footwear-ltd-a25-1286163-koi-footwear-ltd.html, published 18 June 2025, complaint reference A25-1286163, read 14 September 2026.
Upheld on a single complaint. The ruling describes "An email from Koi Footwear, received on 2 April 2025 featured the subject heading 'Take two and call us in the morning' [pill emoji] [pill emoji]", with a syringe and a branded vial in the body. The ASA found that the subject line "referenced medical dosage instructions in a way that would be familiar to consumers" and, taken with the imagery, "condoned and encouraged drug use, and was therefore irresponsible, and likely to cause serious and widespread offence", breaching rules 1.3 and 4.1.
Two things make this the right second ruling to cite. First, the offending element named first in the assessment is the subject line, which is exactly the part of an email that gets written last and reviewed least. Second, the advertiser's defence is the defence every clever subject line has: they said it was "surreal and tongue-in-cheek", inspired by Japanese subcultures and the Silent Hill franchise, and "not intended to reference real drug use". The ASA's response, that "recipients of the email would interpret the syringe and language literally", is the reason step 8 reads the surviving lines for tone before reading them for wit. Intention is not the test.
Where the skill departs: this is a fashion brand addressing a young audience, and a hospitality subject line is unlikely to reach this territory by accident. The skill therefore lists the specific hospitality hazards, drink-driving, hangovers as achievement, body comments, alcohol as deserved reward, news events, rather than restating the ruling's own subject matter, which would not help anyone writing about a Sunday roast. One complaint was enough to produce this ruling, which is worth an owner knowing.
4. Google, "Email sender guidelines", Gmail Help
https://support.google.com/a/answer/81126, no publication date shown on the page; its own updates table lists a requirement added "Dec. 2023", read 14 September 2026.
Cited here for a narrow reason: it is the only source in the set that says what a subject line must not do in order to be *delivered*, as distinct from what it must not do in order to be lawful. Under message formatting it states: "Email message subject, headers, display names, and other message elements should accurately represent the sender identity and message content, and shouldn't be misleading. For example, don't send messages with subject lines starting with Re: or Fwd: unless the messages are actual replies or forwards, and include only actual senders and recipients in From: and To: headers." It adds: "Don't use emojis or other non-standard characters to imitate graphic elements in messages, with the intent to deceive or influence recipients. For example, don't use emojis or images next to display names or brand names to imply the name has been verified in some way." And: "Don't use HTML and CSS to hide content in your messages."
That is step 6. The distinction matters commercially. An ASA breach costs a ruling and an apology; a filtering decision costs the owner the send, and then the next send, and it happens silently with no complaint and no notification.
Where the skill departs: most of this page is about authentication and volume, and the skill deliberately quotes none of its spam-rate percentages at the owner. A venue with a small list cannot measure a 0.1% rate meaningfully, and the risk of quoting it is that three complaints from eight hundred sends reads as a disaster when it is one person clicking the wrong button.
Best public prompt we found for this job
The most relevant public artefact is again the `email-sequence` skill in Anthropic's `knowledge-work-plugins`, raw source at https://raw.githubusercontent.com/anthropics/knowledge-work-plugins/main/marketing/skills/email-sequence/SKILL.md. The repository has 24,016 stars, read from api.github.com on 14 September 2026. Its subject line instruction is compact and the good half is worth copying:
Vary approaches: curiosity, benefit-driven, urgency, personalization, question-based
That is the idea behind the five families in step 2: force range before you force quality, because an unconstrained generator produces twenty variations of whatever it thought of first. We changed the families themselves. "Urgency" is dropped entirely, because for a small venue urgency is nearly always manufactured and rule 8.22 and rule 3.30 are waiting for it. "Personalization" is dropped because a merge field is not a family, it is a formatting choice, and a first name in the subject box of an email from a pub the reader visits in person reads as software rather than as warmth. In their place: plain fact, named person, specific detail and useful information, all four of which are things a venue genuinely has and a competitor cannot copy.
Two other things we did not copy. It asks for "2-3 options per email"; we ask for twenty, because three options from one generator are three flavours of one idea and the discipline of twenty is what produces the line nobody would have written first. And it recommends A/B testing subject lines and measuring a winner. On a list of a few hundred that is measuring noise, and the skill says so in its rules rather than leaving the owner to find out by acting on a result that was never real.
Want this running in your business?
I optimise how businesses run — your sales, your visibility, your social media — and build bespoke software where nothing off the shelf fits. The first conversation is free. Work starts from £150 a day.
Foxera