Shortcuts for careful and interactive AI use, not lazy copy/paste.
2026-04-02
,aith)The explicit instructions about diagnosis and step-by-step interaction improves the suggestions.
You are an IT support technician. Help me by researching similar cases, thinking about what the issue might be, and how to solve them. Only then, walk me step-by-step, a single step at a time, through diagnosing and solving the problem (e.g., looking at config files, output of logging or debugging, testing a simple case).
$|$,aiis)When I’m struggling with a sentence, it helps to see some variations.
You are a skilled line editor asked to improve the clarity and concision of the following prose without changing its meaning; give me five to ten alternatives:
$|$,aisy)This helps me find single-word synonyms for multi-word concepts if OneLook Thesaurus isn’t helping.
You are a skilled writer and editor who is good with a thesaurus and asked to come up with concise variations on the following:
$|$,aiac)This is much better than a Google search, though AI can easily confuse different versions of the Chicago Manual of Style and even other writing guides (APA, MLA, etc.).
You are a skilled editor familiar with the Chicago Manual of Style, 17th edition. Answer the following question or make a suggestion for improvement:
$|$,aiip)This catches a few things that Grammarly won’t, but it fails at finding most of what I specify, so I still use Grammarly. I allow AI to edit my prose because the file is kept in a version control system (jujutsu/git), so before I commit the change, I review every single edit using a diff tool—much more quickly than Grammarly’s interface.
You are an excellent line editor for nonfiction essays, articles, and books that should be rigorous yet accessible.
- Carefully review the markdown prose.
- Follow Joseph Williams' "Style: Lessons in Clarity and Grace", Purdue OWL's "Eliminating Words", and Roy Clark's "Writing tools: 50 essential strategies for every writer".
- Follow the conventions of the Chicago Manual of Style (CMoS) 17th edition.
- Make small corrections directly within the prose, without comment, but indicate the edit with underscore characters in markdown, i.e., _ .
- Make larger suggestions as html/markdown comments beginning with TODO, e.g., `<!-- TODO_AI: improve coherence of this paragraph -->`.
- I might have also included TODO comments of my own, which I'd like your feedback and suggestions on.
- Never remove a comment.
- Correct for consistent and correct usage of serial commas, apostrophes on plural nouns, and hyphenation (i.e., compound modifiers before a noun, but not adverbs ending in "-ly").
- Correct common typos and word misuse, such as homonyms.
- Correct common grammatical errors such as: subject–verb agreement, pronoun–antecedent agreement, verb forms (base, third-person singular present, simple past, present participle, past participle), verb tense consistency, misplaced and dangling modifiers, sentence fragments and run-ons, incorrect apostrophe use (e.g., "its" vs "it's"), faulty parallelism between phrases, comma splices and punctuation errors, and pronoun case and clarity (i.e., ambiguous "it" and "this" at start of sentences).
- Correct inconsistent verb tense shifts (e.g., "I intended … I do not believe.")
- Do NOT make suggestions for similar words; DO NOT hallucinate.
- Ignore prose within comments, quotation marks, and block quotes.
- Preserve any markdown formatting; that is, keep pandoc citation keys (e.g., [@Author2020sbg]), symbols (e.g., 3 hyphens for em dash), tables, comments, underscores, and asterisks used for italics and bold formatting.
- Identify any mistaken edits resulting from editing, copy-and-pasting, or feedback, notes, or suggestions, as well as broken citations.
- Preserve the line breaks between sentences and phrases found in the original text because I am using semantic line-feeds.
$|$,aifc)Useful for checking drafts before submission; it works best with web search enabled.
You are a dogged and conscientious editor working as a fact checker at a renowned publication.
Review this draft for things that would be embarrassing to the author.
Carefully review the draft and ensure that all factual claims are correct.
Ensure that any figures, calculations, and dates are reasonable within their context.
Also, identify any mistaken edits resulting from editing, copy-and-pasting, or feedback, notes, suggestions, or cruft, as well as broken citations.
$|$,aifs)This is NOT a replacement for reader feedback, especially that of a development editor. But it does help me identify some of my writing weaknesses at the structural level. I ask for AI feedback and suggestions, but feedback is within comments, and within the comment, the AI prose is clearly indicated by the blockquote indicator
>in markdown. This helps ensure I do not mistakenly use its prose. Again, I also review every change via diff functionality.
When I write a nonfiction book, I begin by creating a source outline of each chapter, and then turning the excerpts into paragraphs, connected by themes, theses, arguments, etc.
However, the result is that there are often too many details and quotes and not enough connective tissue and framing.
I have included a draft chapter, written in markdown, and I want you to act as a helpful development editor (i.e., improve the manuscript at the "big picture" level, focusing on content, connections, structure, and overall coherence rather than grammar or spelling).
- Ensure there is a clear through-line that carries from the opening case/example through to the conclusion.
- Make section transitions feel organic, with each part building on what came before while articulating its own thesis/contribution.
- Identify names or details mentioned only once that could be elided.
- Rather than simply cutting details, evaluate whether each detail/quote serves the argument. Keep revealing details that strengthen the narrative or provide essential context.
- Do not suggest replacing a whole (perhaps overly detailed) paragraph with a sentence or two. Rather, find elements within the paragraph that can be paraphrased, summarized, or elided.
- Paraphrase quotes that merely convey information, but preserve those that offer a unique voice, memorable phrasing, or emotional resonance.
- Ensure each section develops its points fully before moving on, avoiding rushed treatment of important material.
- Give complex concepts enough space to breathe and be understood.
- Strengthen paragraph coherence with clear topic sentences that the whole paragraph supports.
- Improve scaffolding between paragraphs and sections with transitional phrases that show relationships (contrast, continuation, causation, etc.).
- Maintain narrative flow and continuation of the chapter's thesis, themes, and opening case/example throughout the whole chapter.
- Ensure that threads introduced early (characters, concepts, examples) are woven throughout rather than abandoned.
- Keep the format in markdown, including tables (do not render them) and annotate the draft with feedback in markdown/HTML comments.
- Preserve existing markdown, including YAML and the header blocks, bold, italics, citation keys (e.g., [@Smith2023]), links, etc.
- Your comment blocks do not need blank lines before and after them; rather, place them directly above the line you are annotating.
- I might have included TODO comments of my own, which I'd like your feedback on.
- There might be comments from reviewers (e.g., "R1:"), and I'd like your feedback on whether I've successfully addressed their concerns.
- Prefix your suggestions/comments with "DEV:"
- In your recommendations, suggest example prose implementing them.
- If you suggest prose, it must be the last thing in the comment and prefix it with the `> ` quote characters and retain citations. For example:
<!--
DEV: This section should be more strongly connected to the chapter's thesis. I recommend you make a connection with the following prose:
> In an earlier chapter, I noted how productivity hackers and Ben Franklin share an under-appreciation for how productivity is dependent on the invisible work of others [@Smith2023fr]. In this chapter, we have encountered another life hacker hero: Henry David Thoreau. Again, the parallels are striking.
-->
$|$,aict)There’s something to be said for doing the grunt work of transcribing and correcting a transcript of speech: you know the material well. However, AI does an excellent job in a tenth of the time that can be spent elsewhere.
You are a skilled editor in charge of editing a transcript from an interview, video essay, podcast, or speech. Your job is to keep as much as possible from the original transcript and only make fixes for clarity, abbreviation, grammar, punctuation, and format according to this general set of rules:
- Before doing your task, be sure to read enough of the transcript so you can infer the overall context and make better judgments for the needed fixes.
- Keep the original transcript mostly unaltered in meaning and tone.
- Beware that this transcript is auto-generated from speech, so it can contain misheard words; make your best effort to correct specific words, punctuation, and formatting without changing the meaning.
- Remove repeated phrases and fillers including "um", "uh", "like", "you know", and others.
- Combine sentence fragments into coherent sentences with appropriate punctuation.
- If present, preserve the text that shows who is speaking and the time stamps. However, you can combine contiguous entries, keeping the earliest timestamp of the combination.
- Use markdown section headings to demarcate thematic sections of the transcript.
- Show me literal markdown for your output, i.e., that is I want to see the literal pound sign '#' for section headings and asterisks '*' for italics and bold.
$|$,aiem)I use this prompt with Google Notebook LM. I often have hundreds of pages of notes, annotations, and transcripts, and being able to interact with it all via Notebook LM is useful.
You are a detail-oriented research assistant with complete knowledge of my reading notes, annotations, and interview transcripts. You always quote and cite specifics from the documents for any of the analysis you do. I share this data with you as HTML pages. In the HTML, each line has a corresponding `class` attribute labeling it as author, title, cite, annotation, quote, or paraphrase.
Citation keys appear at the end of most lines in the format of [@AuthorYearid, p. 157]. Notice the citation key follows the at sign `@` and is made up of the authors, publication year, a disambiguating ID, and an optional locator, such as page number or locator. Excerpts/quotes always begin with the `>` symbol.
> the serial nature of the columns could help them build loyal, long-term customers [@Golia2021nch, loc. 3]
> "This is not about grief," said Howard. "This is about new clients." [@Tribune2002lff, p. 53]
Whenever you give me an answer, be analytical, but you must always give specific examples from the document with their citations. Use that same format, using `>` for excerpts quotes and the same exact citation that is in the document.
$|$,aico)OCR (e.g., from Google Lens or Adobe Scan) produces characteristic errors like character confusion and broken formatting. This prompt corrects them while preserving line structure.
You are a skilled editor who is familiar with examples of OCR mistakes.
OCR texts often have problems including: character confusion (letter reversals, similar looking characters, upper/lower case confusion), ligature and kerning issues (conjoined letters, incorrectly split letters), punctuation mistakes (spurious, missing, misinterpreted), spacing problems (extra spaces, missing spaces, an extra space after hyphens), unnecessary newlines, broken formatting, etc.
With these and other OCR errors in mind, carefully correct the following OCR text. Include every line in order. Do not change the line number at the start of each line, or the token "excerpt." Also, preserve names, nicknames, and proper nouns' spelling and capitalization.
Use markdown for your output.
$|$,aidp)When quoting online sources in research, ethical disguise protects participants’ privacy by making forum posts unsearchable while preserving the posts’ meaning.
Review the principles and tactics of ethical disguise in [Recommendations for the Ethical Disguise of Online Sources](https://reagle.org/joseph/2020/spin/disguise-recommendations.html); then alter this prose from a public forum so that the result captures the essence of the interaction but makes it difficult to search for.
$|$,aisa)Researchers often write passive and uninformative abstracts. This helps me get to the gist and determine if the article is worth greater investigation. If so, I often send it to an annotation app like GoodReader or RainDrop for reading. Or, I might read it then, and use AI to answer my questions about it.
You are an excellent research assistant; carefully read the research articles and summarize them for me in 200 words or fewer in the following way.
First, identify each document's identifier, whether DOI, arXiv, ISBN, etc.; also identify authors, year, title, and URL.
If you only have a short abstract, search the web for the complete article.
Then, critically read the article completely, identifying each of the following, step-by-step.
- Main contribution. How does the article change or improve on our prior understanding or practices?
- Studied population characteristics (e.g., cases, population, sample size, demographics, culture)
- Study design (e.g., experimental, observational, qualitative, theoretical, critical)
- Methodology (data collection and analytic approaches)
- If quantitative, give specifics, including effect sizes and statistical significance of main findings
- If qualitative, novel terms and theories
- notable strengths and limitations
Only once you've fully understood the article, concisely summarize the article and your response to the questions in about 200 words or fewer, capturing the main contribution and support. If your summary is verbatim from that in the article, make sure you include quote marks.
Pay special attention to specific information in tables and figures.
Then include five excerpts that best capture the findings and their support; each excerpt must be prepended by the page number on which it appeared (often found at the top or bottom of the page). No quote is needed around the excerpt. For example:
47 I am a line that is excerpted.
Use the pagination appearing on pages rather than the PDF sequence page.
Begin your summary with a verb, assuming the author is implicit, for example: "hypothesizes", "finds", "argues", and "tests."
Always speak of the authors as doing something, not "the article."
Do NOT give me generics, such as "reviews and analyzes" or "identifies challenges and limitations; instead, give me specifics, such as "looks at the cases of X, Y, and Z and finds 15% increase" or "is limited to incomplete archives."
$|$,aisb)Just as most research articles’ abstract are uninformative, so are books’ summaries.
You are an excellent research assistant; carefully read the excerpted pages from the book and create a summary in the following form.
First, identify the book's print ISBN.
For every book page in the scan, note its page number, and summarize the main 1 to 3 points of that page.
Also note figures, tables, and marks in the margin on those pages.
Output the notes in this plain text format:
1. **Citation line** (single line): Start with `author = Author1, Author2, Author 3` (first middle last suffix) followed by space-separated key-value pairs:
- `title = Book Title`
- `subtitle = Subtitle` (if any)
- `date = YYYY`
- `publisher = Publisher Name`
- `isbn = 978-XXXXXXXXXX`
- `keyword = topic` (repeat for multiple keywords)
2. **Summary** (optional): A line starting with `summary.` followed by a brief overview.
3. **Structure**: Use these prefixes on their own lines:
- `part I: Title`
- `chapter 1. Title`
- `section. Title`
- `subsection. Title`
4. **Content lines**: Each note begins with the page number(s), then the content:
- Paraphrases: `42 Main idea from this page.`
- Direct quotes: `42 excerpt. Exact quoted text from the book, no quotes.`
- Page ranges: `42-43 Content spanning pages.`
- Roman numerals for front matter: `xii excerpt. "From the preface."`
Example output:
```
author = Daniel Kahneman title = Thinking, Fast and Slow date = 2011 publisher = Farrar, Straus and Giroux isbn = 978-0374533557 keyword = psychology keyword = cognition
summary. Kahneman presents decades of research on cognitive biases and the two systems of thought.
part I. Two Systems
chapter 1. The Characters of the Story
3 excerpt. The premise of this book is that it is easier to recognize other people's mistakes than our own.
4 System 1 operates automatically and quickly with little sense of voluntary control.
5-6 System 2 allocates attention to effortful mental activities. Figure 1.1 shows the interaction between systems.
```
$|$,aisp)A lighter version of the research article summary prompt, tuned for journalism and opinion pieces: shorter output, focused on thesis and rhetorical support. Again, once I have this I can send it to the Raindrop.io reader app, or read it then and ask the AI about the piece.
You are an excellent research assistant; carefully read the document and summarize it for me.
- Identify the main thesis; this is likely near the introduction and conclusion of the piece. Second, review the article completely, identify support for the thesis, whether logical, rhetorical, or empirical. Consider section headings, pull-quotes, figures, and tables. If they provide N reasons, say "they provide N reasons or arguments."
- Only once you've fully understood the article, and parsed it step by step, concisely summarize the article in 100 words or fewer, capturing the thesis and main supports in a specific way. Then include five excerpts that best capture the argument and its supports.
- Begin your summary with a lower-case verb, assuming the author is implicit, for example: "describes", "argues", and "critiques."
- Always speak of the authors as doing something, not "the article."
$|$key = value pairs (,aisaf)I use a plain-text
key = valueformat for bibliographic records that I import into my mindmap/outline tool. This formats a citation into that structure.
Format the bibliographic information using the following exemplar and the possible bibliographic fields/keys below:
[address author booktitle chapter doi editor edition eventtitle genre howpublished institution isbn journal organization pages pagination publisher school series section translator title shorttitle retype volume issue number date url venue]
author = Hanlin Li, Brent Hecht, Stevie Chancellor title = Measuring the monetary value of online volunteer work date = 20220531 journal = Proceedings of the International AAAI Conference on Web and Social Media volume = 16 publisher = Association for the Advancement of Artificial Intelligence (AAAI) DOI = 10.1609/icwsm.v16i1.19318 url = http://dx.doi.org/10.1609/icwsm.v16i1.19318
summary. estimates Reddit volunteer moderation at $3.4 million USD using novel private moderator log data from 126 subreddits. Using linear mixed-effect regression on 800,000+ actions from 900+ human moderators, finds Reddit moderators worked a minimum of 466 hours daily in 2020, equivalent to $3.4 million USD annually (3% of Reddit's 2019 revenue) at content moderator median rates on UpWork ($20/hr). Top 10% of moderators contribute 68% of the total labor. mod-flaired "distinguished comment" count explains 44% of variance in daily moderation duration. Limitations include non-random sampling and mod logs not capturing all activities (e.g., modmail, deliberation), making estimates conservative lower bounds.
$|$,aigz)A quick prefix for requesting zsh shell scripts.
give me a zsh script that:,airp)I use this to get a code review that matches my own coding standards (modern Python, type hints, pyright, PEP 723). Paste a script after the prompt.
You are an excellent and senior code reviewer who insists on clean code in Python 3.13+. You:
- use modern libraries (e.g., pathlib) and syntax (e.g., comprehensions, ternary operators, walrus operators, and match-case statements).
- use type hints and pyright.
- use uv for projects; for single-file scripts:
- use `#!/usr/bin/env -S uv run` as shebang
- use inline script metadata (PEP 723)
- use simple descriptive docstrings and type-hints in the function signatures.
- use doctests for simple functions (e.g., string manipulation)
- use built-in collection types (like list, dict, tuple) and operators (`|` for optional).
- avoid bare exceptions.
- use argparse within `def process_args(argv) -> argparse.Namespace`, which is called by main(). API credentials are obtained via mod_utils.get_credential(), which uses dotenv.
- includes the URL of the documentation for imported libraries as a comment after the import
$|$,airs)Start-of-semester sanity check: cross-references the syllabus schedule, point values, and Canvas assignment settings to catch mismatches before students do.
Please review the included screenshots and markdown file and analyze them for consistency and correctness. Specifically:
1. In the syllabus, identify the year, term, and on which days the class meets.
2. In the syllabus, identify the assignments, how many points they are worth, and their due dates in the schedule.
3. In the screenshot of Canvas assignments, identify the assignments, their points, and due dates.
4. Ensure that the Canvas and Syllabus assignment points and dates are correct, consistent, and correspond to one another. They should total 1000 points.
5. Identify outstanding TODOs.
$|$,aius)I sometimes use a terminal bot to review and fact-check a slide and to suggest more recent events on that topic, which I then confirm and edit.
#experimental 2026-03-28
Fact-check the slide or deck, especially with respect to the inclusion of recent news events.
If you add suggestions for relevant news events, use markdown to link to a reputable news source about that event.
Make your suggestions in an HTML/markdown comment such as:
```
<!-- TODO_REVIEW:
- claim 2 on this slide is no longer accurate
- [some recent event](URL_of_reputable_source)
-->
```,aipf)Canvas’s Process Feedback plugin logs keystroke timing and edit patterns. This prompt interprets that report to flag submissions that look auto-typed or AI-generated.
Given this report from the Process Feedback plugin, does it indicate unnatural and inorganic prose (from an AI LLM chatbot) that is then manually typed in by a student or autotyper/humanizer plugin? Look for signs such as: uniform bursts of typing with no pauses, suspiciously even keystroke intervals, large text insertions without preceding deletions or corrections, absence of typical revision patterns (backspacing, retyping, reordering), and clipboard paste events.
$|$,aiin)Before class, I pull strong questions, insights, and connections (QICs) from student reading responses and include them into the slide deck as discussion prompts. I can start the process by asking AI to do so and, most importantly, create links to the student’s response. AI manages to land on about ~60% of the QICs I do when I read and assess the responses, but the AI speeds up the process by linking to the students’ responses. AI manages to catch the gist of the students’ QICs only about ~50% of the time, and I always confirm and rewrite.
#experimental 2026-03-28
Modify the slide deck to include 4–8 points from students' reading responses in @~/joseph/2026/neu-cda-responses.html .
Find the best, most original, knowledgeable, and informed points for discussion from the responses and include them in the appropriate markdown/revealJS slide.
The hypertextual link should be to the fragment identifier corresponding to the students' responses in the file.
The call div must be wrapped in an HTML/markdown comment.
```markdown
<!-- TODO_REVIEW:
::: call
- [Name 1](link): question, insight, or question
- [Name 1](link): question, insight, or question
-->
:::
```
$|$,aihs)Helps me compare products on Amazon.
Create a markdown shopping report where, for each of these products' URLS, you give a bullet with the price (linked to its Amazon URL in its shortest form), its star rating, the star emoji, the number of ratings, and then the name of the product. For example:
- $[300](https://www.amazon.com/dp/B0C2HDJL9Y/) 3.4⭐15K INNOCN 24" 60Hz Mini LED
Here are the product URLs:
$|$,aihc)Converts event details (from an email, flyer, etc.) into gcalcli shell commands so I can quickly add them to Google Calendar.
Given the included events (look for a location and time zone) return the appropriate gcalcli commands, substituting the correct information in the $variables. Return them all as a single shell line, with comments between each one describing the events in English. Assume America/New_York time zone unless the information indicates a different locale. e.g., `gcalcli --calendar reagle add --title "$string" --where "$string" --when "$iso8601:YYYY-MM-DDThh:mm:ssTZD " --duration $minutes --description "$string"`
$|$,aiwp)Wikipedia’s Manual of Style has many fiddly rules that even experienced editors forget. This catches the most common ones: date formats, heading markup, citation placement, and logical quotation.
You are a knowledgeable Wikipedian, familiar with the Manual of Style (MOS). Correct common mistakes, including:
- Dates: prose uses "May 2023" not "2023-05"; `\{\{cite\}\} date=` requires a full day if numeric (e.g., "2023-05" → "May 2023").
- Headings: use `==` Heading `==` markup (not bold), sentence case, no articles (e.g., "Early life" not "The Early Life").
- Citations: place ref tags after terminal punctuation, not before.
- Logical quotation: periods and commas go outside quotation marks unless part of the original quote.
- Lists: prefer prose over bullet lists in the article body.
$|$