Advanced Prompting: Prompt Chains, Self-Checks and Reusable Templates That Work in Any Chatbot
~16 min
Before this: How to Talk to AI: A Simple Prompt Formula That Works in Any Chatbot
Next: Build Your Own AI Assistant: Gemini Skills, Claude Skills, ChatGPT Plugins and Grok Skills (Custom GPTs Are Retiring)
Five techniques that turn "pretty good" AI answers into great ones, every time: chain the steps, make the AI interview you, give it a scorecard, make it critique itself, and save the winner as a template. Each one is shown on a real job, from first draft to final.
Checked October 17, 2026. Researched and drafted with help from
Grok Bot.
Quick answer
- Big jobs fail as one big prompt. Break them into steps and check each step. That's a prompt chain.
- Let the AI ask first. "Interview me before you start" fixes most bad answers, because the AI stops guessing.
- Tell it what "good" looks like. A short rubric (scorecard) beats adjectives like "make it great."
- Make it check its own work. A self-critique pass catches errors, gaps and fluff.
- Save what works. Turn a winning chain into a template, Project or skill so you never start from scratch.
- Works in every chatbot, on free plans. For hard jobs, switch to a thinking mode (
ChatGPT Thinking,
Gemini Pro, Claude with Thinking,
Grok Expert).
This is our editorial method, built from testing across chatbots. It's not an official guide from any AI company.
Comparison table
| Technique | What it fixes | Best for | Extra time |
|---|---|---|---|
| 1. Prompt chain | Messy, shallow answers to big jobs | Business plans, long documents, trip plans | 5–10 min |
| 2. Interview me | Generic answers that ignore your situation | Anything personal: budgets, cover letters, plans | 2 min |
| 3. Rubric | Vague quality; "it's fine but not right" | Writing, comparisons, decisions | 1 min |
| 4. Self-critique | Errors, missing points, padding | Anything you'll send or act on | 1–2 min |
| 5. Template | Starting from scratch every time | Weekly or repeated jobs | 5 min once |
How the chatbots handle it: all of them follow chains and rubrics well. Long chains stay on track best in a thinking mode. For very long documents, Claude and
Gemini can hold the most text at once on paid plans (up to about 1 million tokens).
Grok's Expert mode works well for chains that need fresh web information.
Which one should I use? The verdict
- Start with #2 (Interview me). It's the single biggest upgrade for everyday jobs.
- Add #4 (Self-critique) to anything you'll send to a boss, client, landlord or school.
- Use #1 (Chain) for jobs bigger than one page.
- Use #3 (Rubric) when you know what good looks like but the AI keeps missing it.
- Use #5 (Template) for anything you do every week, then save it as a Project or skill.
Quick questions
Do I need a paid plan? No. Every technique works on free plans. Paid thinking modes help on hard, multi-step jobs.
Is "prompt engineering" still a thing? The tricks changed. Magic phrases matter less now. Clear steps, context and a definition of "good" matter more.
Should I use one long prompt or several short ones? Several short ones for big jobs. You can check and steer after each step.
Can I use two chatbots together? Yes. Asking a second chatbot to review the first one's work is one of the best self-checks there is (technique #4, cross-check version).
Where do I save templates? In a Project (ChatGPT,
Claude,
Grok, Perplexity), a
Gemini skill, a Claude Skill or a
ChatGPT plugin. See A3: Build Your Own AI Assistant.
Before you start
- Pick a real job. You'll learn faster on something you need: a cover letter, a budget, a landlord email, a lesson plan.
- Open a fresh chat. Old chats carry old context that can confuse the AI.
- Choose your mode. Normal mode for short jobs. A thinking mode (
ChatGPT Thinking,
Gemini Pro, Claude with Thinking,
Grok Expert) for long chains.
Technique 1: Chain the steps
Instead of "write my business plan," break the job into steps and do them one at a time. You check each step before moving on.
The chain starter:
I need to [big job]. Don't do it yet.
First, break it into 4–6 steps, each with a clear output.
Show me the steps. We'll do them one at a time, and I'll approve each one before the next.
Then, for each step:
Do step [N] only. Use what we agreed in earlier steps. End with: "Ready for step [N+1]?"
Why it works: each step is small enough to get right, and you catch mistakes early, before they spread through the whole document.
Technique 2: Make the AI interview you
Before you answer, interview me. Ask up to [5] short questions, one at a time, about anything you'd need to know to do this really well. Then do it.
Job: [your job]
Variations:
Ask all 5 questions at once(faster).If I say "skip", make a sensible assumption and tell me what you assumed.
Technique 3: Give it a rubric
A rubric is a short scorecard. It tells the AI exactly what "good" means.
Here's how I'll judge your answer:
1. [Clear: a 12-year-old could follow it]
2. [Specific: real numbers, names and dates, not generalities]
3. [Short: under 250 words]
4. [Actionable: ends with 3 next steps]
5. [Honest: says what it doesn't know]
Write your answer so it scores 5/5 on every point.
Tip: write the rubric from your last disappointing answer. What was wrong with it? Turn each problem into a rule.
Technique 4: Make it critique itself
The self-critique:
Now review your answer like a tough editor.
1. Score it 1–5 on each rubric point and explain each score.
2. List anything wrong, missing, vague or padded.
3. Check every fact, number and name. Mark anything you're not sure of.
4. Rewrite it to fix every issue. Show only the final version.
The cross-check (two chatbots):
Another AI wrote this for me: [paste]. The job was: [job].
Find mistakes, weak spots and anything missing. Be specific. Then give me an improved version.
Paste into a different chatbot than the one that wrote it. Different AIs have different blind spots.
The red team (for decisions):
Argue against this plan as strongly as you can. What would make it fail? What am I not seeing?
Technique 5: Save it as a template
When a chain works, ask the AI to turn it into a reusable template:
That worked well. Turn everything we did into one reusable prompt template I can paste next time, with [BRACKETS] for the parts that change. Include the interview questions, the steps and the rubric.
Save it in a note, a Project's instructions, or a skill (see A3).
Worked example: a small-business plan
Step 0: Interview
I want a one-page plan for a weekend dog-walking business in my neighborhood. Interview me first: up to 6 questions, one at a time.
Step 1: Chain
Now break the plan into 5 steps: (1) customers, (2) services and prices, (3) costs and break-even, (4) getting the first 10 clients, (5) risks. Show the steps only.
Step 2: Do each step with a rubric
Do step 2 only. Rubric: real local price ranges (say where they come from or mark them as estimates), 3 packages max, one line each. Under 120 words.
Step 3: Self-critique
Review the full plan as a skeptical small-business advisor. Score it 1–5 on: realistic numbers, clear first steps, honest risks. Fix everything below 5.
Step 4: Cross-check in a second chatbot with the cross-check prompt.
Step 5: Template
Turn this into a reusable "one-page business plan" template with [BRACKETS].
The others
Microsoft Copilot and Perplexity: chains and rubrics work the same way.
Perplexity is especially good for steps that need fresh sources ("Step 1: find current prices, with links").
Meta AI and
DeepSeek: follow the same techniques. Keep chains shorter; check facts with a second chatbot.
Try it: 3 practice jobs
1. Cover letter (Interview + Rubric + Critique)
Interview me (5 questions) for a cover letter for this job: [paste job post]. Rubric: under 250 words, 3 specific achievements with numbers, matches 3 keywords from the post, no clichés ("I'm passionate"). Then critique and rewrite.
2. Monthly budget (Chain)
Help me build a monthly budget. Don't do it yet. Break it into steps: income, fixed costs, variable costs, goals, final table. We'll go one at a time.
3. Landlord email (Rubric + Red team)
Write an email to my landlord asking them to fix [problem], which I reported on [date]. Rubric: polite, firm, under 150 words, a clear deadline, mentions the earlier report. Then tell me how the landlord might push back and adjust the email.
Common mistakes and fixes
| Mistake | Fix |
|---|---|
| The chain drifts off-topic | Restate the goal at each step: "Remember, the goal is…" |
| The AI forgets earlier steps | Paste a short summary of agreed steps before each new one |
| Self-critique says "it's great" | Ask for "at least 3 weaknesses" and a score for each rubric point |
| Too many questions | Cap it: "max 3 questions" |
| Rubric ignored | Put the rubric right before "Write your answer", not at the top |
| Long chat gets slow or confused | Start a new chat with a summary of where you are |
Prompt pack
Universal starter:
Job: [what you need]. Context: [who it's for, your situation]. Good looks like: [rubric]. First, ask me up to 3 questions. Then plan the steps. Then do step 1.
Fact-check pass:
List every fact, number and name in your answer. For each: how sure are you (high/medium/low), and how could I check it?
Simplify pass:
Rewrite this for someone who's never heard of the topic. Short sentences. Explain any term the first time you use it.
Tone pass:
Rewrite this in my voice. Here are 2 things I wrote: [paste]. Match sentence length, word choice and warmth.
Next level
Save your best chains as reusable assistants: A3: Build Your Own AI Assistant (
Gemini skills, Claude Skills,
ChatGPT plugins,
Grok Skills).
Get the free safety checklist
7 rules for letting an agent act for you, plus a note when prices or features change. We re-check every month. Read the 7 rules.