Core Concepts
Using It Well in Practice — Instructions, Verification, and Workflows
Core Concepts, part 3. The craft of specific instructions, making the AI ask first, verification magic words, security rules, how to turn repeated work into skills, and two real-world workflow examples.
This is the final part of the Core Concepts series. Part 1 covered principles, part 2 the tool; this one covers the user's craft. It's why the same tool produces wildly different results in different hands. The principle is simple:
Input determines output. The quality of AI's work is proportional to the quality of your instructions.
The craft of specific instructions
The most common mistake is giving the AI instructions so vague they wouldn't work on a person either. Think of it as delegating to a smart new hire. A new hire needs scope, criteria, and format to do well.
| Vague | Specific |
|---|---|
| "Write a report" | "March sales report: month-over-month changes in a table, three key insights, one page max" |
| "Sort my email" | "From my inbox, pick only the messages that need a reply; table of sender, gist, and deadline" |
| "Research this" | "Market size and top 5 players for [topic] domestically. Include source links; sources from the last 2 years only" |
Four things to cover when you make an instruction specific.
State the purpose
Say who reads it and why, and the tone and depth adjust accordingly. "A summary for investors" and "an internal team memo" require completely different writing even with the same content.
Cut the scope
"Everything" is not an instruction. Cut by period, target, and count. Qualifiers like "the last 3 months," "top 5," "domestic only" sharpen the result.
Specify the format
Tables, bullets, length, even the file type. "In a table," "one page," "save as a markdown file" — specifying format is the cheapest way to raise quality.
Provide a good example
If you have a past deliverable that came out well, attach it: "same format as this file." One example beats a hundred words of description.
Make the AI ask first
You don't need to write perfect instructions. Just append this one line to your prompt:
If anything is unclear or information is missing, ask me before you start.
WHY — Why does this work?
By default, AI tries to "somehow manage with what it's given." When information is missing, it fills the gap with plausible guesses — and when a guess is off, the entire deliverable drifts. Permitting questions replaces guessing with confirming, which eliminates long runs in the wrong direction.
Verification magic words
AI output can be wrong even when it looks right. Numbers and facts especially deserve a second pass. Use these sentences as-is.
| Situation | Magic words |
|---|---|
| After number work | "Verify the calculation you just did, from scratch" |
| After research | "Flag every item from that research where the source is uncertain" |
| After writing a document | "Point out what a first-time reader of this document wouldn't understand" |
| Before an important decision | "Build the strongest possible case against this conclusion" |
TIP — Verify in a separate session
Having the same session verify its own work is letting the student grade their own exam. For important deliverables, open a new session, hand over only the output, and have it verified there — the assessment is far more dispassionate.
Security rules: what's fine to enter and what isn't
Anything you enter into Claude is sent to Anthropic's servers. Ordinary work is fine, but as a rule, the items below should never be entered directly into any AI service.
| Fine to enter | Handle with care |
|---|---|
| Email and report drafts | Customer personal data (names, contact details, ID numbers) |
| General work materials | Undisclosed financial or contract information |
| Public market data | Passwords, API keys |
| Internal meeting summaries | Documents related to legal disputes |
CAUTION — When you must handle sensitive data
If a task involves personal data — say, organizing a customer list — replace names with "Customer A, Customer B" and contact details with "***" before the work, then restore them afterward (pseudonymization). Claude can even do the substitution itself: run the replacement on the local file and let it work only on the substituted copy.
How to think about turning repeated work into skills
Part 2 called skills "recipes." So which tasks deserve a recipe? Three criteria:
- Do you do it at least twice a month? Work that doesn't repeat isn't worth a skill
- Is the procedure reasonably fixed? Work requiring entirely fresh judgment each time fits conversation better than a skill
- Can you articulate the quality bar? If you can explain "it's done well when it looks like this," it can become a recipe
Found a task that fits all three? The build order is simple.
Do it well once, conversationally
Don't design the skill first. Just delegate the task in conversation and refine with feedback until you like the result.
Freeze the good moment
The instant the output satisfies you, say: "Turn exactly what you just did into a skill I can reuse."
Refine in the next real use
Next time the task comes up, call the skill; if anything falls short, say "fix this part of the skill." Two or three rounds and the recipe fits your hand.
Two real-world workflow examples
Here's how the principles above combine in practice, using two common tasks.
Example 1: Meeting notes
Turning a recording transcript or raw notes into organized meeting minutes.
Put the source in your folder
Drop the transcript or handwritten notes into the work folder. If attendee information is sensitive, pseudonymize at this step.
Instruct with a specified format
"Turn this transcript into meeting minutes. Don't list chronologically — reorganize by topic, and for each topic: discussion → decisions → action items (owner, deadline). Mark anything not actually decided as 'proposal'."
Have it verified
"Cross-check whether anything in the minutes was added that isn't in the transcript. Also check whether anything recorded as a decision was actually left unresolved."
Freeze it as a skill
Once you like the format, say "turn this minutes format into a skill" — and from then on, one /minutes command repeats the same quality.
Example 2: Competitor monitoring
Semi-automating a weekly research routine.
Define the watch list in a document
Write the competitor list, the items to check (new products, price changes, job postings, press releases), and the output format into one file. This document becomes the standing brief for every weekly run.
Run the first pass conversationally
"Research according to this document's criteria. Summarize per company as 'changed / no change' first, then detail only the ones that changed. Attach a source link to every item."
Verify the sources
"Check each source link one by one — does it actually contain the claim attributed to it?" Skip this step in research work and plausible misinformation ends up in your report.
Make it a skill and call it weekly
Once the output format settles, freeze it as a skill and call it the same day every week. Hand it last week's result file and say "highlight only what changed since last week" to get a comparison report.
Wrapping up the series
CHECK — Part 3 in five points
- Make instructions specific along four axes — purpose, scope, format, example — as if delegating to a smart new hire
- One line — "ask me first if anything is unclear" — prevents off-target guessing
- Numbers and facts get a second pass with verification magic words — important ones in a fresh session
- Personal and confidential data: don't enter it, or pseudonymize first
- Repeats twice a month + fixed procedure + articulable quality bar = a task worth turning into a skill
That covers principles (part 1), the tool (part 2), and the craft (part 3). What remains is applying it to real work, one task at a time. Start small — any task you repeated even once today is the best starting point.