In my previous posts I discussed my approach to AI adoption. Now I want to go into more detail on my workflows and show specific examples of how assistants integrate into them.
This is Part 3 of a series on AI adoption. Read Part 1: How to Find Value in AI for the foundation-building approach and Part 2: Building Reliable AI Workflows for the system architecture.
In my previous posts How to Find Value in AI and Building Reliable AI Workflows I discussed my approach to AI adoption. Now I want to go into more detail on my workflows and show specific examples of how assistants integrate into them.
I should clarify something that readers might be wondering: No, the output I receive isn't perfect. This is why I've developed validation systems and quality checks as part of my workflow.
Is this overhead worth it? In my experience, when you provide sufficient context to your assistant and establish proper validation systems, yes. I've still accomplished more in less time using these tools. As you continue to maintain, improve, and document your processes, this becomes increasingly efficient. This reinforces something I mentioned in my first post in this series: "despite all the talk about 'natural language' interaction with computers, it still takes a technical mindset to approach this technology effectively."
The primary tools I use for work and automation are iA Writer, Shortcuts, Alfred, Hazel, Audio Hijack, and DEVONthink. They each serve specific purposes and integrate well into my day-to-day work. Here's a brief overview of their use:
Some basic examples are prompt templates. These not only allow me to reuse instructions for frequent tasks, they're also platform-agnostic; I can use them with any model, from small to large, local to cloud-based systems.
Basic Structure:
You are my proofreading assistant. Fix errors while preserving my original voice and intent. Return only the corrected text, no commentary.
**Content Type:** [email|document|message|note]
**Preserve:** [tone|format|style] OR "all"
**Text to Fix:** [content]
**Rules:**
- Fix spelling, grammar, and obvious typos
- Keep my original tone and style
- Don't change quoted material or technical terms
- Maintain original formatting (line breaks, emphasis)
- Light touch for casual messages, more thorough for professional documents
**Parameter Examples:**
TYPE: email | PRESERVE: tone
TYPE: message | PRESERVE: all
TYPE: document | PRESERVE: format
Example Usage:
You are my proofreading assistant. Fix errors while preserving my original voice and intent. Return only the corrected text, no commentary.
**Content Type:** email
**Preserve:** tone
**Text to Fix:** "Hi Sarah, I wanted to follow up on our converstaion about the React component library. Can we shedule some time this week to discuss the implementation detials?"
**Expected Output:** "Hi Sarah, I wanted to follow up on our conversation about the React component library. Can we schedule some time this week to discuss the implementation details?"
Quick Parameter Guide:
- email: Professional communication, preserve greeting/closing style
- message: Casual communication, minimal changes, keep conversational flow
- document: Formal writing, thorough proofreading while maintaining structure
- note: Personal writing, light touch, preserve individual style
I apply these templates in two main ways. With Apple Shortcuts: 1) Loads the template 2) Asks me to fill out parameters and sometimes presents options for how to proceed 3) Appends my formatted responses to the prompt and copies it to clipboard. I then paste into the chat field and hit enter. The other common approach is adding templates to a project, then starting new chats with "use proofreading prompt on this text."
I have templates covering word definitions, proofreading, research synthesis, job applications, schedule extraction, and more. The first three are straightforward, so I'll cover job applications and schedule extraction in more detail.
I discussed some details of how I apply for jobs using these tools in my first post of this series, but now I'll offer more specifics on how this actually works. At this point, the system is well-established. There's always room for improvement and I continue looking for ways to make the output more accurate and consistent.
Within my system prompt, it knows to reference the job application template when I ask for help applying for a role. The job application prompt contains research instructions, résumé tailoring instructions, a cover letter template, writing samples along with approved edits to those samples, a style guide, forbidden terms, and describes the application materials to return.
Basic Structure:
You are my job application assistant. Create tailored materials for [COMPANY] [ROLE] using my resume master document.
**Company:** [COMPANY_NAME]
**Role:** [ROLE_TITLE]
**Source:** `2025-05-27-resume-master.md`
**Key Requirements:**
- [REQUIREMENT_1]
- [REQUIREMENT_2]
- [REQUIREMENT_3]
**Research Focus:**
- Company mission and recent news
- Tech stack from job posting
- Team structure if available
**Output Needed:**
1. Tailored resume with company-specific optimizations
2. Cover letter (4 paragraphs: mission alignment, technical fit, leadership examples, culture fit)
3. Interview prep with technical and behavioral questions
4. Key talking points for the role
**Quality Standards:**
- Use only verified experience from resume master
- Include accessibility experience for frontend roles
- Quantify impact with specific metrics
- Match authentic professional voice
- No buzzwords: avoid "crafting," "drove," "passionate," "leverage"
**Validation Examples:**
- ✅ "Built component library from 1 to 50+ production components" (verified metric)
- ❌ "Spearheaded innovative solutions across the organization" (buzzwords + unverifiable)
- ✅ "Led TypeScript adoption across frontend team of 8 engineers" (specific scope)
- ❌ "Passionate about creating elegant, performant user experiences" (prohibited language)
Example Usage:
You are my job application assistant. Create tailored materials for Meta Senior Frontend Engineer using my resume master document.
**Company:** Meta
**Role:** Senior Frontend Engineer
**Source:** `2025-05-27-resume-master.md`
**Key Requirements:**
- 5+ years React experience
- TypeScript expertise
- Accessibility knowledge
- Cross-functional collaboration
**Research Focus:**
- Meta's mission around connecting people
- React, GraphQL, mobile-first development
- Engineering culture and remote work
**Output Needed:**
[Complete application package as specified above]
My resume is already largely complete, so what it returns are the contents of my résumé with small adjustments for the given role. The cover letter is somewhat similar, except I often take time to customize it by hand. The produced materials include a summary of the role, the company and its mission, as well as interview practice and potential questions I might ask about the company during an interview.
If there are additional questions in the application process, I typically write them myself and use the assistant to proofread them. As mentioned above, I have many writing samples, so it knows to do this without changing my tone, style, or intent, only fixing typos, grammatical errors, or occasionally correcting structure when a sentence doesn't flow naturally.
I could allow the assistant to compose a response for me, and sometimes I have, but I wouldn't send it without editing. The responses have been getting better and better. At this point the assistant has many samples and approved edits, but I still want to write the answer to many of these questions myself. In some cases, this is specifically requested on the application. When applications specifically request my own words, I limit AI assistance to proofreading for grammar and spelling only. This isn't meant to sound arrogant, but there's an integrity in my communications and work that I want to preserve.
To be clear, I do believe there's value in allowing the assistant to synthesize content based on the context and material you've provided. But going back to the importance of review, it's important that this is reviewed and truly represents what you want to say and how you would have said it.
The need for this came out of work where the company provided an app that published my weekly schedule. The problem was I couldn't conveniently add this to my personal calendar. To make adding this to my calendar more efficient, I decided to take screenshots, upload them, and prompt the assistant to extract the dates, format them in 24-hour time, and construct an ICS file. I download this file and import it into my calendar. Depending on the integration of your chosen assistant with calendar applications, this might be even more automated for you.
Additionally, I have document templates that take raw text and return completed, organized documents. This is particularly useful for audio transcription: I often transcribe videos, create organized notes, then watch the videos using those notes as a guide, keeping them as reference material afterward. This approach also works well for talks where I might use the template but can't organize information in real-time during the presentation.
This automated organization isn't always exactly what I want, since there's value in reviewing notes and organizing material yourself as part of the learning process. This organization is most useful when the task is more administrative in nature than something requiring comprehension.
I suggest giving it a try, finding something that works for you, and doing a little research. Take a look at Anthropic's AI Fluency course; it provides excellent foundational thinking about how to collaborate with these systems effectively.
There's actually much more I could discuss: my complete note-taking system, the standards for documents that are created, how some of the applications fit into taking those documents, organizing them, and archiving them. But hopefully this gives you helpful ideas for ways to integrate assistants into your workflows.
However, the most valuable concept is context management. Learning how to successfully manage the context of a specific task is what will give you the best results. More on this in my next post.
Enjoyed this article? I share thoughts on component architecture, and AI-enhanced workflows, regularly.