The numbers survived the LinkedIn rewrite. The caveat didn't.

Sources and assumptions

• Anthropic, Prompting best practices (Claude docs), 'Ground responses in quotes' under long-context prompting: 'For long document tasks, ask Claude to quote relevant parts of the documents first before carrying out its task. This helps Claude focus on the relevant content and ignore the rest of the document.' The old long-context-tips URL (docs.claude.com/.../long-context-tips) now redirects here. The docs aim this at long documents and at Claude; our test used a short post and GPT-5.5, so the post may only say the idea is adapted from this guidance, not that the docs recommend it for this case. (checked 2026-10-22)
• Our own test, record.tests ids before-gpt and after-gpt (GPT-5.5 via the OpenAI API, run 2026-10-23 UTC / 2026-10-22 local): All output quotes and the number-by-number comparison. The outputs come from the recorded runs, not from this URL. The URL is only the model family's documentation. (checked 2026-10-22)

Assumptions:

• The sample blog post (Northgate Cycles) is fictional, written for this test, about 800 words, with no real business or personal data.
• Each prompt was run once in GPT-5.5 via the OpenAI API (label 'GPT-5.5 (OpenAI API)', not the ChatGPT app). Another run may keep or drop different details.
• A Claude Sonnet 5.5 run of the vague prompt was attempted, but the local claude CLI failed in this session (record.tests id 'before', output null). The post must not claim any Claude result or say the prompt was tested in Claude.
• Number check of the before output: every figure used (2017-era 'seven years', three mechanics, about 40 calls, March 2025, 15-minute, 23 emergencies, 6→2 days, about 22→9, 71%, $74→$86, 31%→39%, 4.6→4.8, two weeks/two months) matches the blog, except '6 days', which became 'about 6 days'. The labor-rate caveat on the $74→$86 rise was dropped. The before output left out 412, $49, $300, 196 and the 8 slots (left out, not changed).
• Number check of the after output: every number it used matches the blog exactly, and the $12 caveat is kept. Two drifts remain: 'Half of our before numbers are estimates' became 'many of our before numbers are estimates', and the invented line 'Not zero stress. Not zero chaos.' reuses the blog's 'Not zero'.
• The after output was 13,862 characters, mostly the <numbers> list (about 60 entries, including words like 'one', 'first', 'second').
• Using your own blog post is low risk. If the post draws on client or customer details, remove names and other personal details first, and check your employer's approved tools before pasting.