Prompting is a craft — and it has patterns.
Seven moves that compound. Climb the ladder only as high as the task requires. By the end of the session you will have a personal prompt library running on real tasks from your week.
- Climb the prompt ladder (one-shot → role + constraint) on demand
- Reach for the right pattern (role / examples / format / step / persona / critique / constraint) by the failure mode
- Diagnose a disappointing output with the failure-modes table
- Start a personal prompt library you commit to revisit Thursdays
The ladder of prompts.
Climb only as high as the task requires. Most professionals live one rung below where they should — and blame the model when the answer is generic.
Climb only as high as the task requires. The diagram, then the rungs.
live one rung
below this →
Click a rung. The right column shows when to reach for it.
One-shot
A single line. No context, no shape.
With context
Who you are, what you are working on.
With format
Plus the shape of the answer.
With examples
Plus a few-shot of what good looks like.
Role & constraint
The full structured prompt.
With iteration
Run, push back, refine. The dialog.
Climb the ladder on one of your real prompts.
- Rung 1 — one-shot: write it as a single line, no context, no shape. Note the result.
- Rung 2 — + context: add who you are and what you’re working on. Re-run.
- Rung 3 — + format: add the shape of the answer (bullets, table, max length). Re-run.
- Rung 4 — + examples: add a couple of your best past outputs. Re-run.
Same order as the diagram above — one-shot → context → format → examples → (role + constraint → iterate).
Stretch. Note the rung where the score flattened. That’s your team’s default level for this task type. Codify it as a template.
The Role pattern.
Tell the model who to be, and its outputs sharpen.
From the last unit: The ladder told you how high to climb. The next seven units are the rungs — the moves that compound. We start with the one that costs you nothing and sharpens everything.
Models are massively multi-modal in tone. They can imitate a tax attorney, a beat poet, or a thoughtful skeptic. Don’t make them guess.
Template
Worked example
Add a role to your worst-performing prompt.
- Identify the kind of expert who would do this task brilliantly (‘senior copywriter at a fintech’).
- Add it as the first line of the prompt: ‘You are a [exact role with seniority and domain].’
- Re-run on the same 3 inputs as before.
- Note quality change.
Stretch. Try 3 different roles. The best one becomes a template you reuse.
The Examples pattern.
Show, don’t tell. A single example is often worth a paragraph of instructions.
From the last unit: Role tells the model who to be. Examples tell it what good output looks like. Three concrete examples beat a paragraph of adjectives.
Especially for classification, formatting, or voice — three concrete examples beat a hundred adjectives.
Add 2 examples to a prompt.
- Pick 2 cases where the model previously produced good outputs.
- Add them as
<example>...</example>blocks in the prompt. - Pick examples that cover the *edges* (one easy, one tricky) — not the middle.
- Re-run on 5 fresh inputs. Compare.
Stretch. Examples grow stale. Refresh quarterly as your output style evolves.
The Format pattern.
Control the shape of the answer, and you control half the usefulness.
From the last unit: Role and examples shape what the model writes. Format controls the shape of what comes back. The cheapest 10x in your week.
Default model output is a paragraph because that’s what most internet text looks like. You usually want a list, a table, or a structured block.
Specify the output format explicitly.
- Add ‘Output as: ...’ with the exact structure (JSON keys, markdown sections, max length).
- Add a one-line example of the format if non-trivial.
- If the agent needs to be machine-parseable, specify JSON with a schema.
- Re-run on 5 inputs. Quality of structured fields should be more consistent.
Stretch. Validate parsing. If the agent fails JSON, the prompt isn’t specific enough — or you need tool_choice.
The Step-by-step pattern.
When the problem is hard, ask the model to think before answering.
From the last unit: You can shape the output. Now shape the thinking. When the problem is hard, make the model write down its reasoning before its answer.
Output quality goes up because the model writes the reasoning down — which means each next token is conditioned on the steps that came before.
This is sometimes called chain-of-thought prompting. It also lets you spot where the model went wrong, instead of just disagreeing with its conclusion.
Convert a reasoning prompt to step-by-step.
- Add: ‘Think step by step. List every consideration first. Then weigh them. Then conclude.’
- Re-run on a case where it previously made a mistake.
- Check: did the step-by-step trace catch the error?
- If the conclusion is still wrong, the missing step is the actual gap — specify it.
Stretch. For tricky reasoning, also enable thinking on Opus and compare.
The Persona pattern.
Teach the model your voice by showing samples of how you actually write.
From the last unit: Role gives the model a persona class. Persona makes it your specific voice — via your actual writing.
“Warm and direct” means twenty things to twenty people. Your actual writing means exactly one thing — yours.
Build a persona from 3 writing samples.
- Grab 3 samples of their writing (~1000 words total).
- Ask Claude: ‘Profile this writer’s style: sentence length, vocabulary, structure, what they include vs leave out.’
- Save the profile as a reusable persona block.
- Use it in a new prompt: ‘Write X in this voice [paste persona].’ Compare to the samples.
Stretch. The persona drifts. Re-extract every 6 months or after major writing changes.
The Critique pattern.
The most underused move: ask the model to disagree with you.
From the last unit: You can shape style. Now shape rigor. Critique is the move that turns Claude from cheerleader into sharpest reviewer.
The default is helpful, agreeable, encouraging. Helpful is sometimes the worst thing a draft can be.
Variants
- Steel-man — the strongest version of the opposing view.
- Red team — what a hostile reader would say.
- Pre-mortem — assume this fails; explain why.
Add self-critique to a finished prompt.
- After the main task, add: ‘Now critique your own output. List 3 things a thoughtful skeptic would push back on. Then revise once.’
- Re-run on 5 inputs. Compare initial draft to revised version.
- Note whether the critiques are substantive or performative.
- If performative, sharpen: ‘Critique like an editor of [domain]’.
Stretch. If revised matches initial, the critique step doesn’t fire — the model is too confident. Make the critique a separate call with a different role.
The Constraint pattern.
Telling the model what NOT to do is often more powerful than telling it what to do.
From the last unit: Six patterns so far told the model what to do. The seventh tells it what NOT to do — and that constraint is often where the real lift lives.
Constraints worth memorizing
- “No preamble. Start with the answer.”
- “No hedging. No ‘I think’ or ‘I believe.’”
- “Don’t use the words [X], [Y], [Z].”
- “Maximum N words / N lines / N bullets.”
- “If you’re not sure, say so.”
- “Don’t [send / delete / move]. Draft only.”
Compounding constraints
Add a constraint that eliminates a known failure.
- Name the failure precisely (‘adds marketing words’, ‘exceeds 200 words’).
- Write the constraint as a negative rule: ‘Never use the words leverage, robust, transformative.’ or ‘Stay strictly under 200 words.’
- Add to the prompt. Re-run on 5 cases that previously failed.
- If the constraint isn’t obeyed, the failure isn’t actually in the prompt — it’s in the model. Switch strategy (examples, format, or a different model).
Stretch. Add the constraint to your default prompt template. Don’t re-litigate it each time.
The prompt lab.
Compose a prompt with all six ingredients live. Edit any field; the assembled prompt updates beneath. Copy it and paste it into Claude.
From the last unit: You have the seven patterns. The lab is where you compose all of them at once and watch the prompt take shape live.
Paste the assembled prompt into the Claude chat app — or into the script from Day 01. The structure does the work. The model is just the engine.
Run the full lab on one task.
- Start with L1 (just ask). Score quality 1-5.
- Add each pattern in turn. Re-run on the same 5 inputs each time.
- Score quality after each addition.
- Note: which patterns added the most? Which added little? Which actively hurt?
Stretch. Patterns that added little can be dropped. Patterns that hurt should be removed. Document the final template.
Failure modes and their repair.
When a response disappoints, it usually fails in one of six predictable ways. Each has a one-line fix.
From the last unit: You ship a prompt. The output disappoints. Six predictable failures cover ~95% of those moments. Each one has a one-line fix.
| If the output is… | The cause | The repair |
|---|---|---|
| Bland or generic | Missing context or constraint | Add who you are + one constraint |
| Wrong shape | No format specified | Specify length, structure |
| Hedged | Default helpfulness | “No hedging. State it directly.” |
| Confidently wrong | Hallucination on a fact | Ground in source, ask for citations |
| Off-topic / drifting | Drift over a long conversation | Start fresh; restate the constraints |
| Doesn’t sound like you | No voice samples provided | Paste 2–3 samples of your writing |
Catalog the 5 most-common failure modes you’ve seen.
- Open your 30 most recent Claude conversations. Note where output was bad.
- Cluster failures into 5 categories (e.g. hallucinated facts, format breakage, off-topic, too long, too short).
- For each, write the repair: which pattern (Role/Examples/Format/Critique/Constraint) fixes it.
- Save as a playbook in your prompt library.
Stretch. Add categories quarterly. The taxonomy evolves as you push the model harder.
Your prompt library.
The compound asset that turns this from “I used AI once” into a practice.
From the last unit: The patterns are tools. The library is the compound asset. The compound is why this becomes a craft instead of a one-time win.
How to start one — in five minutes
- Open any text file. Name it
my-pradhya.md. - Add a section per pattern — Role, Examples, Format, Step-by-step, Persona, Critique, Constraint.
- Paste one working prompt under each. The first version of every prompt is allowed to be ugly.
- Every Thursday, spend 10 minutes opening the file. Run one prompt on something real. Improve one sentence.
# my-pradhya.md ## Role You are a senior product reviewer at a hard-news publication... ## Examples (email triage) "Hey, can you send me the deck by 3?" — ACTION "Quick note — we updated the policy page." — FYI ## Format (one-page brief) SUMMARY: ... THREE RISKS: ... RECOMMENDATION: ... ## Critique Critique this draft. Push back hard. List three things... # revisit every Thursday · 10 min · improve one sentence
The compound
After 30 days you have ten prompts you’d defend. After 90, a small library that does real work. After a year, a personal asset that no AI vendor can take away — because the value isn’t in the prompts, it’s in knowing which ones work on your stuff.
Save your top 5 prompts to a library.
- Create a folder:
~/prompts/or a Notion page or a Pradhya skills folder. - Each prompt gets: a name, a one-line description, the prompt itself, and one sample output.
- Cross-reference each with the patterns it uses (Role + Examples + ...).
- Test the library: a week from now, can you find the right prompt in < 30 seconds?
Stretch. Share one prompt with a teammate. The first time a colleague reuses your prompt, the library has paid for itself.