Practices/ Capable Series/ Day 02 · Cadence
02 / 04
Pradhya Day 02 · Cadence 11 units · ~45 min reading · 3 labs to do in session, 8 to carry into your week

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.

By the end of Day 02 · Cadence
  • 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
§ 02.01 · Unit 08 · Do now · in session

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.

one-shot + context + format + examples role + constraint iterate value
most pros
live one rung
below this →
The ladder of prompts · climb only as high as the task requires

Click a rung. The right column shows when to reach for it.

01

One-shot

A single line. No context, no shape.

“Summarize this email.” · For trivia and quick edits where you don’t care about the form.
02

With context

Who you are, what you are working on.

“I lead a product team and need to cut this email by half.” · Most everyday work.
03

With format

Plus the shape of the answer.

“Three bullets, under 12 words each.” · Anything shared, saved, or sent.
04

With examples

Plus a few-shot of what good looks like.

Show the model your three best past outputs. For repeatable patterns.
05

Role & constraint

The full structured prompt.

All six ingredients (role, context, goal, examples, format, constraint). For high-stakes work where the cost of a generic answer is real.
06

With iteration

Run, push back, refine. The dialog.

For hard problems where the first response is a starting position, not the answer.
The fastest 10x The fastest 10x in your week is to climb one rung. The ten-second prompt that gets you a generic email becomes a 90-second prompt that gets you the real one.

Climb the ladder on one of your real prompts.

You’ll do
Take a prompt you wrote today. Walk it up the ladder rung-by-rung.
Steps
  1. Rung 1. one-shot: write it as a single line, no context, no shape. Note the result.
  2. Rung 2. + context: add who you are and what you’re working on. Re-run.
  3. Rung 3. + format: add the shape of the answer (bullets, table, max length). Re-run.
  4. Rung 4. + examples: add a couple of your best past outputs. Re-run.
  5. Rung 5. + role and constraint: name the perspective to use and the failure to avoid. Re-run.
  6. Rung 6. + iteration: critique the strongest output against the same criteria, revise once, and score the revision.

Same order as the diagram above. One-shot → context → format → examples → (role + constraint → iterate).

Verify
Keep all six outputs and six scores, using the same three criteria at every rung. Scores may rise, flatten, or fall; the simplest rung that meets your quality bar is the right default for this task.

Stretch. Note the rung where the score flattened. That’s your team’s default level for this task type. Codify it as a template.

§ 02.02 · Unit 09 · Lab: this week

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.

? no role cast a role tax atty poet skeptic
Same model · different persona by instruction

Template

You are a [role] with [N years / specific background]. You care about [priorities]. Review this [thing] for [specific goal]. Be specific.

Worked example

You are a senior product reviewer at a hard-news publication. You care about clarity, specificity, and the absence of marketing language. Review the draft below for places where I'm relying on jargon or assertion instead of evidence. Be specific. Quote the line, then suggest a fix.
Where it doesn’t work Role does not give the model knowledge it didn’t have. “You are a doctor” doesn’t add medical facts. It shifts style and priorities, not capability.

Add a role to your worst-performing prompt.

You’ll do
Take a prompt that keeps disappointing you. Add a specific role.
Steps
  1. Identify the kind of expert who would do this task brilliantly (‘senior copywriter at a fintech’).
  2. Add it as the first line of the prompt: ‘You are a [exact role with seniority and domain].’
  3. Re-run on the same 3 inputs as before.
  4. Note quality change.
Verify
Side by side, the role version is better on at least 2 of your 3 inputs. Keep the tally (no-role vs role, win/lose per input). If it’s 0–1 wins, the role isn’t telling the model anything it lacked. Spell out what the role implies (the audience, the quality bar, what this expert would check) and re-run. That context is usually where the lift comes from.

Stretch. Try 3 different roles. The best one becomes a template you reuse.

§ 02.03 · Unit 10 · Lab: this week

The Examples pattern.

Show, don’t tell. A few varied examples are 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.

input 1 label A input 2 label B input 3 label A new input ? model infers the pattern
Few-shot · show the pattern, let the model finish it

Especially for classification, formatting, or voice. Three concrete examples beat a hundred adjectives.

Classify each email as REPLY, FYI, ACTION, or SKIP. Here are examples: "Hey, can you send me the deck by 3?". ACTION "Quick note. We updated the policy page.". FYI "Are you free Thursday at 11?". REPLY "Newsletter from a blog I subscribed to". SKIP Now classify these: [list]
Pro move Include one edge case in your examples. One well-chosen edge example is worth five typical ones.

Add 2 examples to a prompt.

You’ll do
Pick a prompt where quality is uneven. Add 2 examples and re-test.
Steps
  1. Pick 2 cases where the model previously produced good outputs.
  2. Add them as <example>...</example> blocks in the prompt.
  3. Pick examples that cover the *edges* (one easy, one tricky). Not the middle. Make them differ in length and wording, and say they’re illustrative: the model copies an example’s length, tone, and structure closely, so one lone sample becomes a template.
  4. Re-run on 5 fresh inputs. Compare.
Verify
On your 5 fresh inputs, tally before vs after. The edge-case inputs improve on at least 2 of 5; the routine inputs score the same or better (never worse). If a routine case got worse, your examples are pulling it off-pattern. Swap one.

Stretch. Examples grow stale. Refresh quarterly as your output style evolves.

§ 02.04 · Unit 11 · Lab: this week

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.

Without format a wall of prose With format SUMMARY: ___ RISKS: • ___ • ___ RECOMMEND: ___ a shape you can read
Format = half the usefulness · for free

Left to choose, the model picks a shape for you, and different models pick differently. Say the shape you need. You usually want a list, a table, or a structured block.

Answer in this exact format: SUMMARY: [one sentence] THREE RISKS: [list] RECOMMENDATION: [one sentence, specific] No preamble. No qualifying phrases.

Specify the output format explicitly.

You’ll do
Take a prompt where the format is implicit. Make it explicit.
Steps
  1. Add ‘Output as: ...’ with the exact structure (JSON keys, markdown sections, max length).
  2. Add a one-line example of the format if non-trivial.
  3. If code will parse the output, don’t ask for JSON in prose. Use structured outputs (Day 03 · Unit 22): hand the API a JSON schema and the reply is guaranteed to match it.
  4. Re-run on 5 inputs. Quality of structured fields should be more consistent.
Verify
Format is consistent across runs. Downstream code never has to handle “sometimes markdown, sometimes plain”.

Stretch. Validate parsing. If a prose-requested JSON reply ever fails to parse, stop tuning the wording and switch to structured outputs (output_config.format). Forcing a tool call with tool_choice was the older trick; current models reject it (as of Oct 2026).

§ 02.05 · Unit 12 · Lab: this week

The Think & check pattern.

Current models already think before they answer. Your job is to set how hard, and to ask for the plan or the check you actually need.

From the last unit: You can shape the output. Now decide how much thinking the task gets. On current Claude models the reasoning happens in a built-in thinking step; you steer its depth, not its wording.

problem list considerations weigh them answer each token conditioned on the prior steps
Reasoning before the answer · now built into the model

The mechanism hasn’t changed: when reasoning comes before the answer, each next token is conditioned on the steps that came before, and quality goes up. What changed is where it happens. Current Claude models (Opus 5.5, Sonnet 5.5, Haiku 5.5, Fable 5.1) think adaptively in a separate thinking step before the visible answer, so “think step by step” in your prompt is mostly redundant. You steer depth with the effort setting in the API (and /effort in Claude Code), not with prose (as of Oct 2026).

What still earns its place in the prompt: the check you want run, and the short explanation you want back.

Recommend one option for [goal]. Before you commit, check it against these criteria: [criteria]. If it fails one, say which and pick again. In your answer, give the recommendation, the two or three considerations that decided it, and what would change your mind.

The older version of this pattern was chain-of-thought prompting: “think step by step, show your reasoning.” Asking for a short explanation of the deciding factors gives you the same benefit (you can spot where it went wrong, instead of just disagreeing with its conclusion) without asking for a reasoning dump. On Opus 5.5, Sonnet 5.5, and Fable 5.1, an instruction to reproduce the model’s internal reasoning in the answer can be declined outright (as of Oct 2026).

Swap the incantation for a check.

You’ll do
Pick a prompt whose answer disappointed you. Test whether “think step by step” still helps, then test what usually does: a stated check and the missing context.
Steps
  1. Re-run the disappointing prompt as-is and save the answer (A).
  2. Add ‘Think step by step.’ and re-run (B). Note whether anything changed.
  3. Replace that line with a check: ‘Before you answer, verify it against [the specific thing it got wrong]. Then give the answer and the two or three factors that decided it.’ Re-run (C).
  4. If C is still wrong, the gap is usually context: a fact, constraint, or definition the model couldn’t know. Add it and re-run.
Verify
You have A, B, and C side by side and can name which change fixed it: the incantation, the stated check, or the added context. On current models it’s rarely the incantation. If nothing fixed it, name the missing fact explicitly. That’s your real gap.

Stretch. In the API, run the original prompt at effort low and high and compare quality against output tokens. Thinking is adaptive by default on current models; effort is the dial.

§ 02.06 · Unit 13 · Lab: this week

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.

sample 1 sample 2 sample 3 match the voice new piece in your voice
“Warm and direct” means many things · your writing means one

“Warm and direct” means twenty things to twenty people. Your actual writing means exactly one thing. Yours.

Here are three things I've written recently. Study the voice. Sentence length, vocabulary, what I include and what I leave out: [Sample 1] [Sample 2] [Sample 3] Now write [the new thing] in that voice. Match the rhythm. Don't be precious.

Build a persona from 3 writing samples.

You’ll do
Pick someone (yourself, a colleague, an author). Extract their voice into a persona.
Steps
  1. Grab 3 samples of their writing (~1000 words total).
  2. Ask Claude: ‘Profile this writer’s style: sentence length, vocabulary, structure, what they include vs leave out.’
  3. Save the profile as a reusable persona block.
  4. Use it in a new prompt: ‘Write X in this voice [paste persona].’ Compare to the samples.
Verify
Blind-match it: drop the model’s output into a set with 2 real samples from that writer and 2 from someone else, then ask a colleague (or a fresh Claude chat) “which of these five share an author?” The generated piece is grouped with the real samples. If it lands in the wrong pile, the profile is too generic. Add 2 more samples and re-extract.

Stretch. The persona drifts. Re-extract every 6 months or after major writing changes.

§ 02.07 · Unit 14 · Lab: this week

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.

your draft three paragraphs push back hard 1. Unsourced claim on line 4 2. Headline buries the lede 3. Weak ending revise v2
The model is more useful as your sharpest critic than as your cheerleader

The default is helpful, agreeable, encouraging. Helpful is sometimes the worst thing a draft can be.

Critique this [draft / argument / plan]. Push back hard. List the three things a thoughtful skeptic would point out. Be specific. Quote the line you're objecting to.

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.
Why this works You don’t always need to be right. You need to know where you’re wrong before someone else finds it.

Add self-critique to a finished prompt.

You’ll do
Take a prompt that produces decent output. Add a critique pass.
Steps
  1. After the main task, add: ‘Now critique your own output. List 3 things a thoughtful skeptic would push back on. Then revise once.’
  2. Re-run on 5 inputs. Compare initial draft to revised version.
  3. Note whether the critiques are substantive or performative.
  4. If performative, sharpen: ‘Critique like an editor of [domain]’.
Verify
Side by side, the revised draft beats the initial draft on at least 3 of your 5 inputs. Keep the tally. If the critique step changes nothing, it isn’t firing (see the Stretch).

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.

§ 02.08 · Unit 15 · Lab: this week

The Constraint pattern.

Constraints carve the answer. State the real ones plainly, give the reason, and say what you want instead of listing what you don’t.

From the last unit: Six patterns so far told the model what to do. The seventh sets the limits: the hard lines (draft only, don’t guess) and your house style. The lift comes from stating them precisely, with a reason.

space of possible responses "leverage" hedging preamble >200 words close off the bad answers · what remains is the good one
Constraints carve the answer · state them plainly, with the reason

Current models follow instructions closely and literally. That cuts both ways: a precise constraint with its reason gets applied well, while a list of bare “don’ts” gets over-applied, and banning a failure the model wasn’t going to make can steer it toward that failure. So keep the hard lines as hard lines, and phrase style limits as what you want.

Constraints worth memorizing

  • “Lead with the answer.”
  • “State conclusions directly; flag real uncertainty once, where it applies.”
  • “Use plain words; our readers distrust terms like [X] and [Y].” (the reason makes the rule generalize to words you didn’t list)
  • “Size it for where it lands: one phone screen / a two-line Slack update.” Use a hard number only when the slot is hard (a 160-character SMS).
  • “If you’re not sure, say so.”
  • “Don’t [send / delete / move]. Draft only.” A real hard line. Keep it as a prohibition.

Compounding constraints

Rewrite this for executives who will skim it on a phone: about half the length, plain words, conclusions stated directly. They tune out marketing terms, so swap words like "leverage," "robust," and "transformative" for what you actually mean. Start with the verb. If anything in the original is unclear, mark it [UNCLEAR] rather than guessing.
Common mistake Don’t stack ten contradictory constraints. Pick three that pull in the same direction, and give each its reason. A rule with a “because” generalizes; a bare prohibition gets applied too literally.

Add a constraint that eliminates a known failure.

You’ll do
Take a prompt that produces a recurring failure (long, hype-y, off-topic). Constrain it out.
Steps
  1. Name the failure precisely (‘adds marketing words’, ‘exceeds 200 words’).
  2. Write the constraint as a plain rule with its reason, phrased as what you want: ‘Use plain words; our readers distrust marketing terms like leverage, robust, transformative.’ or ‘Keep it short enough to read on one phone screen.’
  3. Add it to the prompt. Re-run the same 5 cases that previously failed and keep the before/after result for each.
  4. If the constraint still isn’t obeyed, look for what pulls against it (an example that breaks the rule, a conflicting instruction elsewhere in the prompt) before blaming the model. Then switch strategy (examples, format, or a different model).
Verify
Your before/after table shows fewer failures, with a target of 0/5 after the change. Do not report a percentage from this tiny sample; keep every remaining counterexample for the next revision.

Stretch. Add the constraint to your default prompt template. Don’t re-litigate it each time.

§ 02.09 · Hands-on · 20 min · Do now · in session

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.

role context goal examples format constraint composed prompt six fields → one prompt →
The prompt lab assembles the parts 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.

You’ll do
Pick one real task. Apply the builder’s six fields: Role + Context + Goal + Examples + Format + Constraint. Measure each addition.
Steps
  1. Start with L1 (just ask). Define three scoring criteria and score quality 1–5.
  2. Add Role, Context, Goal, Examples, Format, and Constraint in turn. Re-run on the same 5 inputs each time.
  3. Score quality after each addition.
  4. Note: which patterns added the most? Which added little? Which actively hurt?
Verify
You have a layered prompt + a score table. The marginal value of each pattern is visible.

Stretch. Patterns that added little can be dropped. Patterns that hurt should be removed. Document the final template.

§ 02.10 · Unit 17 · Lab: this week

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. These six failures recur often enough to make a useful first diagnostic, and each has a one-line experiment to try.

bland shape hedged wrong drift not you each has a one-line fix · see the table below
Six failures · six repairs · no mystery
If the output is…The causeThe 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 youNo voice samples provided Paste 2–3 samples of your writing

Catalog the 5 most-common failure modes you’ve seen.

You’ll do
Across all your prompts this month, what fails most? Build the repair toolkit.
Steps
  1. Open your 30 most recent Claude conversations. Note where output was bad.
  2. Cluster failures into 5 categories (e.g. hallucinated facts, format breakage, off-topic, too long, too short).
  3. For each, write the repair: which pattern (Role/Examples/Format/Critique/Constraint) fixes it.
  4. Save as a playbook in your prompt library.
Verify
Your playbook has 5 rows, each pairing a named failure with one repair pattern (Role / Examples / Format / Critique / Constraint). Test it on the next disappointing output: classify it into one of the 5 rows and apply that row’s repair. The fixed version lands in 1 follow-up, not the 3–4 it took before you had the table. If your failure doesn’t fit any row, add a 6th row with its repair.

Stretch. Add categories quarterly. The taxonomy evolves as you push the model harder.

§ 02.11 · Hands-on · 10 min · Do now · in session

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.

day 0 day 365 leverage 30 days 90 days 1 year personal asset
The library compounds · one sentence improved per week

How to start one. In five minutes

  1. Open any text file. Name it my-pradhya.md.
  2. Add a section per pattern. Role, Examples, Format, Think & check, Persona, Critique, Constraint.
  3. Paste one working prompt under each. The first version of every prompt is allowed to be ugly.
  4. 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.

Commitment card Before Day 03, run three prompts from your library on real tasks. One should disappoint you. Bring it to the next session. The disappointments are where the patterns get sharper.

Save your top 5 prompts to a library.

You’ll do
Pick the 5 prompts you’ve refined this week. File them where future-you will find them.
Steps
  1. Create a folder: ~/prompts/ or a Notion page or a Pradhya skills folder.
  2. Each prompt gets: a name, a one-line description, the prompt itself, and one sample output.
  3. Cross-reference each with the patterns it uses (Role + Examples + ...).
  4. Test the library: a week from now, can you find the right prompt in < 30 seconds?
Verify
The file (or page) exists with 5 entries, each having all four parts: name, one-line description, the prompt, one sample output. A week from now, open it and find the right prompt for a new task in under 30 seconds.

Stretch. Share one prompt with a teammate. The first time a colleague reuses your prompt, the library has paid for itself.