Tell AI who reads your work update, and jargon disappears.
What the Reel says
Tell AI who reads your work update, and jargon disappears.
I had AI write my boss an update.
Its first draft was full of jargon.
6
accounting and tech terms
So I added one line about my boss.
Then: "She isn't from finance… use everyday words."
6 → 0
dates, prices, decisions kept
Add one line on what your reader knows.
Sources and assumptions
- Anthropic: Prompting best practices (Claude Platform docs), 'Be clear and direct': Verbatim: 'Think of Claude as a brilliant but new employee who lacks context on your norms and workflows. The more precisely you explain what you want, the better the result.' and the key test: 'Show your prompt to a colleague with minimal context on the task and ask them to follow it. If they'd be confused, Claude will be too.' (checked 2026-11-01)
- Anthropic: Prompting best practices (Claude Platform docs), 'Add context to improve performance': Verbatim: 'Providing context or motivation behind your instructions, such as explaining to Claude why such behavior is important, can give Claude more information about your goals and help it deliver more targeted responses.' This page has no sentence specifically about describing the audience, so slides must not claim the docs say 'describe your reader'. They may say the docs advise giving context because the model lacks your context. (checked 2026-11-01)
- Our test: GPT-5.5 via the OpenAI API (engine/ai-test.js), runs 'before', 'after', 'after-plain': The three outputs quoted on slides 3, 6 and 7 (record.tests, ran 2026-11-02 UTC = evening of 2026-11-01 Pacific). Model name GPT-5.5 as returned by the API run. (checked 2026-11-01)
Assumptions:
- All sample material is fictional: 'Billbright' is a made-up invoicing tool, Dana is a made-up manager, and the figures (120 invoices a month, 38 open invoices, $45/month for 3 seats, 2.9% + $0.30 card fee, Nov 15/Nov 24/Dec 1 dates) are invented for the test.
- Jargon count = these accounting or technical terms from the notes appearing in the output: GL, API, re-keying, ACH, AR aging, Net-30. Before: all 6. After: AR aging, ACH, Net-30 (3). After-plain: 0. The word 'seats' (before and after) became 'users' in after-plain; it's not counted.
- Each prompt was run once on GPT-5.5 via the OpenAI API; another run or another tool can word things differently. Slides say 'in our test' and name the model and month (GPT-5.5 (OpenAI API), Nov 2026).
- The queue idea named Claude, but the Claude runs in this session failed (claude -p command error, output null in tests.json), so no Claude result exists and the post must not name Claude as tested.
- Facts kept in after-plain, checked line by line against the notes: Dec 1 start, automatic reminders at 7/14/30 days, bank transfer or card, 30-day terms unchanged, $45/month for 3 users on the annual plan, 2.9% + $0.30 per card payment, both decisions due Nov 15, 38 invoices moved the week of Nov 24, spreadsheet read-only Dec 1. Dropped: the accounting-system sync and the end of manual re-keying, which Dana said she doesn't care about. Not mentioned: tying the change to late payments being her biggest headache.


