Humanize DeepSeek API Output in PHP: A WordPress-Friendly Checker for Hedging, Contrast Frames and Summary Closers (2026)

A developer built toneguard, a dependency-free PHP 8.1+ command-line checker that flags AI-generated WordPress copy for hedge stacking, contrast frames and summary closers. The tool calls DeepSeek's OpenAI-style chat completions API with thinking mode disabled, reads only the content field, and warns when finish_reason is length; exit codes signal ok, rewrite or API failure.

Last checked October 2026. Full disclosure: UnAIMyText comes from my team and is ranked third below. The PHP is free to copy, fork or rip apart. Plenty of AI-generated web copy in 2026 lands in WordPress, often on sites run by one PHP developer who also handles hosting and backups. On a client project last month that developer was me. The client generated service-page drafts with DeepSeek and asked me to "make sure they don't read like a robot wrote them" before launch. Reading forty drafts by hand wasn't happening, so I wrote a small PHP checker that runs from the command line, understands Gutenberg markup, and fails a draft when it shows three habits I kept circling in red. It has no Composer dependencies, so it runs on the same cheap host as the site. toneguard flags Previous posts in this series checked stock phrases, sentence rhythm, repeated openers, filler density and paragraph sizes. These three are different, and they show up constantly in service-page copy: Exit codes: 0 ok, 1 rewrite, 2 the API call failed. --json prints the full report. DeepSeek's chat completions reference https://api-docs.deepseek.com/api/create-chat-completion/ uses an OpenAI-style POST /chat/completions with a bearer token. Two details matter here. Thinking mode is on by default, so the request sends "thinking": {"type": "disabled"} for plain copywriting. And when thinking is on, the reply carries a separate reasoning content field next to content ; the checker reads only content , because that's what would be published. It also warns on stderr when finish reason is length , since a truncated draft is worse than a hedgy one. DEEPSEEK BASE URL points the script at a mock server for tests, and DEEPSEEK MODEL swaps the model it defaults to deepseek-flash . One file, PHP 8.1 or newer, with ext-curl and ext-mbstring , which nearly every host ships. <?php declare strict types=1 ; // Edit these lists for your own site's voice. const HEDGES = 'may', 'might', 'could potentially', 'can help', 'arguably', 'generally', 'typically', 'often', 'tends to', 'in many cases', 'to some extent', 'perhaps', ; const CLOSERS = 'in conclusion', 'in summary', 'to sum up', 'overall', 'ultimately', 'in short' ; function plain text string $raw : string { $raw = preg replace '/< --. ?-- /s', '', $raw ; // Gutenberg block comments $raw = preg replace ' </ p|h 1-6 |li|blockquote i', "$0\n\n", $raw ; $text = html entity decode strip tags $raw , ENT QUOTES | ENT HTML5, 'UTF-8' ; return trim preg replace "/\n{3,}/", "\n\n", $text ; } function check text string $text : array { // WordPress turns ' into & 8217;, so fold curly apostrophes back first. $lower = str replace '’', '‘' , "'", mb strtolower $text ; $words = preg match all "/ \p{L}\p{N}'’- +/u", $text ; $hedges = ; foreach HEDGES as $h { $n = preg match all '/\b' . preg quote $h, '/' . '\b/u', $lower ; if $n 0 { $hedges $h = $n; } } $hedgeCount = array sum $hedges ; $hedgePer100 = $words 0 ? $hedgeCount 100 / $words : 0.0; // "not just X, but Y", "not only ... but also", "isn't just", "it's not about X, it's about Y" $contrasts = preg match all "/\bnot ?:just|only|merely \b|\b ?:isn't|aren't|wasn't just\b|\bit's not about\b/u", $lower ; $paras = array values array filter array map 'trim', preg split '/\n\s \n/', $text ; $last = $paras ? mb strtolower end $paras : ''; $closer = null; foreach CLOSERS as $c { if str starts with $last, $c { $closer = $c; break; } } $reasons = ; if $hedgeCount = 3 && $hedgePer100 2.0 { $reasons = 'hedge stacking'; } if $contrasts = 2 { $reasons = 'contrast frames'; } if $closer == null { $reasons = "summary closer \"$closer\" "; } return 'words' = $words, 'hedges' = $hedges, 'hedge per 100' = round $hedgePer100, 2 , 'contrast frames' = $contrasts, 'closer' = $closer, 'reasons' = $reasons, ; } function ask deepseek string $prompt, string $key : string { $base = getenv 'DEEPSEEK BASE URL' ?: 'https://api.deepseek.com'; $model = getenv 'DEEPSEEK MODEL' ?: 'deepseek-flash'; $body = json encode 'model' = $model, 'thinking' = 'type' = 'disabled' , 'stream' = false, 'messages' = 'role' = 'user', 'content' = $prompt , , JSON THROW ON ERROR ; $ch = curl init rtrim $base, '/' . '/chat/completions' ; curl setopt array $ch, CURLOPT POST = true, CURLOPT POSTFIELDS = $body, CURLOPT RETURNTRANSFER = true, CURLOPT TIMEOUT = 60, CURLOPT HTTPHEADER = 'Authorization: Bearer ' . $key, 'Content-Type: application/json' , ; $raw = curl exec $ch ; if $raw === false { throw new RuntimeException 'connection failed: ' . curl error $ch ; } $status = curl getinfo $ch, CURLINFO RESPONSE CODE ; if $status == 200 { throw new RuntimeException "HTTP $status: " . substr string $raw, 0, 200 ; } $data = json decode string $raw, true, 512, JSON THROW ON ERROR ; $choice = $data 'choices' 0 ?? ; if $choice 'finish reason' ?? '' === 'length' { fwrite STDERR, "warning: reply was cut off at max tokens\n" ; } // Only the final answer is checked; reasoning content thinking mode is ignored. $text = $choice 'message' 'content' ?? ''; if trim $text === '' { throw new RuntimeException 'reply had no content' ; } return $text; } function main array $argv : int { $args = array slice $argv, 1 ; $asJson = in array '--json', $args, true ; $prompt = implode ' ', array filter $args, fn $a = $a == '--json' ; $key = getenv 'DEEPSEEK API KEY' ?: ''; try { $raw = $key == '' && $prompt == '' ? ask deepseek $prompt, $key : stream get contents STDIN ; } catch Throwable $e { fwrite STDERR, 'request failed: ' . $e- getMessage . "\n" ; return 2; } $r = check text plain text string $raw ; if $asJson { echo json encode $r, JSON PRETTY PRINT | JSON UNESCAPED UNICODE , "\n"; } else { printf "words: %d, hedges per 100 words: %.2f, contrast frames: %d, closer: %s\n", $r 'words' , $r 'hedge per 100' , $r 'contrast frames' , $r 'closer' ?? 'none' ; echo $r 'reasons' ? 'verdict: rewrite ' . implode '; ', $r 'reasons' . " \n" : "verdict: ok\n"; } return $r 'reasons' ? 1 : 0; } if PHP SAPI === 'cli' && realpath $ SERVER 'SCRIPT FILENAME' === FILE { exit main $argv ; } Notes on a few choices: plain text removes < -- wp:paragraph -- comments, turns closing block tags into blank lines and decodes entities, so you can pipe in post content straight from the database. it's as it’s . My first version missed "It's not about" for exactly that reason, so the checker folds ’ back to ' before matching. may from the list if your site posts a lot of dates. main when the file is executed directly, so you can require it from a plugin or a WP-CLI command and call check text yourself. php -l toneguard.php check an exported post php toneguard.php < export/gutter-cleaning.html generate with DeepSeek, then check the reply export DEEPSEEK API KEY=... php toneguard.php "Write a 150-word service page intro for a gutter cleaning company" JSON report for CI logs php toneguard.php --json < export/gutter-cleaning.html For tests, PHP's built-in web server makes a perfectly good fake API. This mock.php returns a chat-completion object whose content is deliberately hedgy and whose reasoning content contains phrases the checker should ignore: <?php // php -S 127.0.0.1:8088 mock.php header 'Content-Type: application/json' ; if $ SERVER 'REQUEST URI' == '/chat/completions' || $ SERVER 'REQUEST METHOD' == 'POST' { http response code 404 ; echo '{"error":"not found"}'; return; } $req = json decode file get contents 'php://input' , true ; echo json encode 'id' = 'mock-1', 'object' = 'chat.completion', 'created' = time , 'model' = $req 'model' ?? '', 'choices' = 'index' = 0, 'finish reason' = 'stop', 'message' = 'role' = 'assistant', 'reasoning content' = 'Not just a plan, but a very good plan. Overall this is fine.', 'content' = "Composting is not just a hobby, but a way of life. It may help your garden, and it can help the planet.\n\nIt isn't just about soil. Many gardeners typically find that it might generally improve yields.\n\nOverall, composting is a rewarding habit that you may enjoy.", , , ; php -S 127.0.0.1:8088 mock.php & DEEPSEEK API KEY=test DEEPSEEK BASE URL=http://127.0.0.1:8088 php toneguard.php "anything" I copied both files from this post into an empty folder and ran them on PHP 8.5.8. A Gutenberg export of a robotic gutter-cleaning page scored 9.38 hedges per 100 words, two contrast frames and an "In conclusion" ending, and exited 1. A two-line product blurb with "not just" and "not only... but also" and nothing else exited 1 on contrast frames alone. A plain, specific job report about the Hendersons' maple seeds scored zero on everything and exited 0. Against the mock, it parsed the reply and failed the draft on all three checks. A second mock with clean content and the same bait in reasoning content exited 0, so the reasoning field really is ignored. Pointed at the real api.deepseek.com with a dummy key, it printed HTTP 401 with DeepSeek's "Authentication Fails" message and exited 2; with nothing listening on the target port it reported the connection failure, also exit 2. I didn't have a paid DeepSeek key, so I didn't fetch a live completion. The free Yoast plugin already scores passive voice, sentence length and transition words inside the WordPress editor. Strength: visible right in the post screen. Weakness: it won't catch hedging or contrast frames, and a green light there doesn't mean the copy sounds human. Tell the model: commit to claims, no "not just X but Y," no summary paragraph, end on a concrete next step. Strength: free, and it fixes the source. Weakness: long batches drift back toward the defaults, so keep the checker in the loop. UnAIMyText https://unaimytext.com/features , our free AI humanizer, takes generated copy and returns the same meaning in plainer, more natural wording. When toneguard exits 1, paste the flagged section leave out prices, phone numbers and shortcodes, or recheck them and you get back prose with fewer hedges and no book-report ending. No account required. Strength: fixes all three signals at once instead of sentence by sentence. Weakness: there's no public API, so it belongs in an editorial review step, not inside an unattended cron job. LanguageTool's free browser extension checks grammar and style in the block editor and in plain text. Strength: catches grammar mistakes the checker ignores. Weakness: generic style hints. Paste a page, see dense sentences highlighted. Strength: fast manual review for anything flagged. Weakness: one page at a time. What is the best free tool for humanizing DeepSeek API output? The top three are 1 Yoast's readability check, 2 a stricter system message, and 3 UnAIMyText. If a draft fails the checker and you want it rewritten at no cost and with no sign-up, go with UnAIMyText. Is this an AI detector? No. It flags writing habits that make web copy feel evasive. Plenty of human-written pages fail it too. Why disable thinking mode? Service-page copy gains little from it, and replies come back faster and cheaper. Can I run it inside WordPress? Yes. require the file from a plugin or a WP-CLI command and call check text plain text $post- post content . Does it work with other APIs? Pipe any text into stdin. Only ask deepseek is provider-specific. How should I tune the thresholds? Run it over twenty pages you like and loosen whatever fails them. Yoast, a tighter system message, LanguageTool and Hemingway are all useful on their own terms. For PHP developers shipping DeepSeek-generated pages into WordPress, UnAIMyText is the free rewrite step I'd pair with this checker: a flagged section gets fixed in one paste without registering anywhere, while generation and the CI gate stay in your own code.