Want to access all premium content?
Become a paying subscriber to get access to all courses, guides like this, prompts and skills library, and $500+ worth of AI tools.
Ask Claude for a welcome email in one sentence and you might get a generic draft with placeholders and an offer to tailor it later. You still have to supply the details, spot what’s missing, and decide when it’s ready.
This setup gives Claude a way to check its work. Before drafting, it writes three to five tests for what a finished result should include. After drafting, it reports which tests it met. A fresh chat then checks the result against the same list.
You’ll need a Claude account and one task from this week with notes, data, or a draft. The Free plan covers the first four steps.
Step 1. Catch the lazy tells. Run a task the usual way and see what you’d still have to fix.
Step 2. Set the finish line. Define what the result must do before Claude starts.
Step 3. Save the trigger. Start a message with full job when you want the full workflow.
Step 4. Add the inspector. Use a fresh chat to check the draft against the original tests.
Step 5. Make it a skill. Save the full workflow so you can reuse it without pasting the instructions.
Step 1. Catch the lazy tells
Choose a task you’d normally ask for in one sentence. For this walkthrough, open a new chat and ask for a newsletter welcome email.
Write a welcome email for new subscribers to my newsletter.
Read the reply and look for these four signs.
What to look for | What it looks like | What you still have to do |
First draft only | A draft with no checks against your requirements | Decide whether it’s ready |
An offer to do more | “If you tell me more, I can…” | Start another round |
Sample details | Placeholders, invented numbers, or generic perks | Replace them with real information |
Self-grading | Claude presents the work as complete without showing what it checked | Verify the result yourself |
A one-line request leaves Claude to guess what a send-ready email should include. Keep this chat open so you can compare it with the next draft.
Step 2. Set the finish line
Give Claude your task and the material it should use. Ask it to write the tests before it drafts anything.
Before you start, set the finish line for this task. 1. Write 3 to 5 pass/fail tests for "done". Each test names something I can check outside this chat: a number, a count, a rule from my material, or a quick test with a person. "It reads well" doesn't count. 2. Show the tests in a code block, then do the work in full. Use my material. If something is missing, ask once at the start instead of filling it with sample details. 3. End with a report: each test marked Met, Missed or Not checked, with the number or the reason. Then list what I still need to check myself. Task: [your task] Material: [paste your notes, data or draft]
Read the tests before the draft. If one says “friendly tone,” ask Claude to turn it into something you can actually check.
Rewrite test 2 so I can check it with a number or a person.Then read the report. A Missed item tells you which requirement needs work. A Not checked item tells you where you still need evidence or a manual check.
Step 3. Save the trigger
Add a shorter version to your personal instructions so you can call it up with full job.
Click your initials in the lower left, then open Settings → Your Claude → Instructions. In the older layout, look under Account for Instructions for Claude.
Paste this into the field.
When my message starts with "full job", set a finish line before you work: write 3 to 5 pass/fail tests for "done" that I can check outside this chat, show them in a code block, do the work in full with my material, and end with each test marked Met, Missed or Not checked. Ask once for missing material instead of filling gaps with sample details. For every other message, answer as usual.
Personal instructions apply across your chats. Open a fresh chat and test the trigger with your notes.
full job: Write a welcome email for new subscribers to my newsletter. Material: [paste your notes]
Then ask a quick question such as “What’s 15% of 240?” It should answer directly. If it starts writing tests for quick questions, check that the saved instruction still begins with When my message starts with "full job".
Step 4. Add the inspector
Use a fresh chat to review the draft without showing it the first chat’s verdict. Copy the tests and the work into this prompt.
You're checking someone else's work. You didn't write it, so don't defend it. Here are the tests it has to pass, then the work. For each test, mark Pass or Fail and quote the line or number that proves it. Then list up to 3 gaps, most important first, each with the exact fix. Tests: [paste the tests block] Work: [paste the draft]
Take the inspector’s findings back to the original chat for one revision.
An outside check found these gaps. Fix them, then rerun your report against the same tests.
[paste the gaps]
After that round, review any remaining failure yourself. Check whether the draft still needs a specific edit, the test is unclear, or Claude needs more material from you.
Step 5. Make it a skill
A skill can hold the full workflow, including a ready-to-copy inspector prompt.
Open Settings → Capabilities and enable Code execution and file creation. In the updated layout, look for Create and edit files.
Start a new chat and paste this prompt.
Make me a skill called finish-line and package it as a ZIP I can upload. The folder name must be finish-line. Description: Use when my message starts with "full job". Sets pass/fail tests before the work, does the work in full, then reports each test as Met, Missed or Not checked. Instructions: 1. Before any work, write 3 to 5 pass/fail tests for "done". Each test names something I can check outside the chat: a number, a count, a rule from my material, or a quick test with a person. 2. Show the tests in a code block, then do the work in full with my material. Ask once for anything missing instead of filling gaps with sample details. 3. End with a report: each test marked Met, Missed or Not checked, with the number or the reason, then what I still need to check myself. 4. Last, give me a ready-to-copy inspector prompt with the tests and the work filled in, for a fresh chat to grade.
If Claude shows a Save skill card, open Content, review the instructions, and click Save.
If there’s no card, download the ZIP and open Settings → Customize → Skills. Click + → Create skill → Upload a skill and select the file. Use one installation route to avoid duplicate copies.
Once the skill is enabled, remove the Step 3 instruction so the two versions don’t overlap. Start your next task with full job. If Claude doesn’t pick up the skill, type /finish-line or ask it to use your finish-line skill.
If the upload fails, check that the ZIP contains a folder named finish-line with a SKILL.md file inside.
In Claude Code v2.1.273 or later, sign in with the same account and run /skills to check for it under claude.ai sync.
Use cases
Use case #1. Write a landing page
Check whether someone can identify the offer from the top of the page.
full job: Write the copy for a one-page landing page for my course. One test must be a 5-second test: a fresh chat sees only the headline and the first two lines, and has to say what the course is, who it's for and what it costs. Material: [paste your course notes, price and who it's for]
Run the check in a new chat with just the headline and first two lines. See whether it can name the course, audience, and price from that copy alone.
Use case #2. Clean a spreadsheet
Attach your CSV and define the checks the cleaned file should pass.
full job: Clean this sales sheet. Every row total must equal price times quantity, there must be no duplicate orders, and each monthly total must match the sum of its rows. List the row numbers you changed.
Before using the file, recalculate two changed rows yourself.
Use case #3. Research a decision
Enable web search so Claude can find and link the sources for its numbers.
full job: Write a one-page brief on whether I should do this: [your plan in one or two lines]. Every number needs a source link from the last 12 months, and one section must make the strongest case against my plan.
Open three of the sources and compare their figures with the brief.
You can use the same approach for a pitch deck outline, a sponsor reply, or a weekly report. Choose tests that would catch a mistake you’d otherwise have to fix later.
Limitations
The tests depend on the material you supply. With a one-line request and no supporting details, Claude may ask for more information or mark a test Not checked.
A fresh chat can compare the draft with the requirements you give it. Checks that depend on outside results, such as open rates or feedback from a real reader, are yours to run.
Personal instructions apply across chats, and Claude has to recognize the full job trigger. Project instructions apply within their own project.
Skills require code execution. Anthropic’s help pages list Skills on all plans, including Free. Claude Code sync requires v2.1.273 or later.
A full run adds a test list, report, and sometimes a second review, so it uses more of your plan allowance than a quick reply. The settings and menus in this walkthrough were checked on November 11, 2026.
Closing
Try the Step 2 prompt on one real task this week. Read the tests first and rewrite any you couldn’t check yourself.
When the prompt works for you, save it as a trigger or a skill. Use the inspector before sending an important draft, and keep the tests that catch problems you actually need to fix.
See you in the next one :)