Module 2 — Prompting Fundamentals

Source: ai-research/anthropic-academy-claude-101-2026-05-04.md (snapshot of Claude 101 “Getting better results”) + ai-research/anthropic-academy-ai-fluency-framework-2026-05-04.md (snapshot of AI Fluency: Framework & Foundations) + ai-research/claude-tutorial-4ds-behavioral-indicators-2026-05-04.md (snapshot of 4 Ds of AI Fluency — Behavioral Indicators)

Time: Read 5 min | Watch 30 min | Practice 15 min

Watch First

Two Anthropic courses teach prompting — both free, both worth it. Start with the Claude 101 section; then do the AI Fluency framework if you want the deeper mental model.

1. Claude 101 — “Getting better results” section (~15 min)

Direct: anthropic.skilljar.com/claude-101

Practical tactics for getting Claude to produce what you actually want. This is the shortest path to “my prompts work better now.”

2. AI Fluency: Framework & Foundations (~60 min total; Modules 3–4 here — ~25 min)

Direct: anthropic.skilljar.com/ai-fluency-framework-foundations — free, certificate on completion.

This course introduces the 4D Framework — Delegation, Description, Discernment, Diligence — a mental model for effective AI collaboration co-developed with Prof. Joseph Feller (UCC) and Prof. Rick Dakan (Ringling). For WEO purposes, Module 4 “Description” (the prompting module) is the most immediately relevant.

Quick primer on behavioral indicators: 4 Ds of AI Fluency — Behavioral Indicators (~5 min read) — maps each of the 4 Ds to specific observable behaviors.

Why It Matters at WEO

Prompting is the single skill that separates “Claude saves me an hour a week” from “Claude saves me an hour a day.” The WEO bar is the second one. Most people who say “I tried Claude and it wasn’t that useful” wrote one-line prompts and gave up.

Two extra minutes structuring the prompt up-front usually saves ten minutes editing mediocre output downstream. Past that first-order win, prompting compounds:

• Muscle memory — after two weeks of RCTF, you stop writing one-line prompts at all. Your default-mode prompt is 80% of the way there before you’ve even thought about it.
• Personal prompt library — the handful of prompts that work well for your role become reusable templates. By month three, your library has 10–20 prompts and you’re mostly filling in placeholders rather than writing from scratch.
• Memory compounds it further (Module 4) — Claude learns your style, your common client shapes, your preferences. Your baseline output quality rises without you trying.
• Team quality floor — when everyone uses RCTF, the agency’s Claude output stops looking like it came from five different people. Review cycles shorten. Directors stop rewriting whole deliverables.

The team member who prompts well looks noticeably faster than the one who doesn’t, even when both have similar domain expertise. This is a skill with an unusually high ceiling and an unusually shallow learning curve — two weeks of intentional practice gets you competent; two months gets you fluent.

WEO-Specific Overlay — The RCTF Pattern

Anthropic’s material covers many prompting techniques. For WEO’s work (marketing copy, research, briefs, ad variants) we’ve standardized on one pattern that you’ll see in every prompt in Module 3:

RCTF = Role + Context + Task + Format.

Element	What it does	Example
Role	Who Claude should “be” — sets tone, expertise, vocabulary	”You are a senior dental marketing copywriter with 10 years of family-practice experience.”
Context	Background Claude needs to do the job — client, audience, constraints	”Practice is in Columbus Ohio, target patients 35-55, brand voice warm and plainspoken.”
Task	What you want Claude to do — clear and specific	”Write three variations of a homepage hero headline.”
Format	How the output should be structured	”Max 8 words each, numbered list, no emojis.”

RCTF in action — a worked example

Prompt without RCTF (what most people write first):

Write me a homepage headline for a dental practice.

Prompt with RCTF (what a WEO team member should write):

Role: You are a senior dental marketing copywriter specializing in family practices.

Context: Practice is Smile Springs Family Dental in Columbus, Ohio. Target patients are families with kids and adults 35–55. Brand voice is warm, plainspoken, trustworthy — not clinical. Key differentiator: Saturday appointments and no-wait booking.

Task: Write 5 homepage headline variations.

Format: Max 8 words each, numbered list, no emojis, no “world-class” or “state-of-the-art.”

The second prompt takes 90 seconds to write instead of 10. The output is 5× more useful. Module 3 has 25 more examples of this pattern across all 5 WEO roles.

Why RCTF actually works (not just what it is)

Each of the four labels collapses a different kind of ambiguity that Claude would otherwise fill with generic defaults:

• Role fixes the voice problem. Without a Role, Claude defaults to what it treats as a neutral “helpful assistant” tone — which reads corporate-bland in a dental-marketing context. With a Role, Claude narrows vocabulary, sentence cadence, and default assumptions toward the persona you named. “Senior dental marketing copywriter with family-practice experience” and “assistant” produce meaningfully different outputs from the same task.
• Context fixes the audience problem. Without Context, Claude writes for an average reader — which is no specific reader. With Context, Claude can make audience-specific moves: word choice for 35–55 parents differs from word choice for 65+ retirees in ways Claude will execute on if you name them.
• Task fixes the scope problem. “Write about implants” is an entire content calendar. “Write 5 homepage headline variations, max 8 words each” is a single bounded deliverable. Vague Tasks produce outputs where Claude guessed at the shape and got it wrong.
• Format fixes the shape problem. “No emojis, numbered list, max 8 words each” is a contract Claude holds itself to. Without Format, Claude picks the shape at random — prose when you wanted bullets, long when you wanted short, one example when you wanted five.

The absence of any label creates a vacuum Claude fills with defaults. RCTF is just the minimum viable structure to stop Claude from guessing in four places at once. That’s why it works on the first try, and why skipping any one label tends to produce the most common failure mode for that dimension (see When Output Goes Wrong below).

RCTF variations — add these when you need them

RCTF is the baseline. For harder work, stack one of these on top:

• RCTF + Constraints — when you need guardrails on what NOT to produce. “Never use the word ‘smile’ or ‘dazzle.’ No hyphenated adjectives. No rhetorical questions.” Constraints are often more valuable than instructions: telling Claude what to avoid prevents the failure modes you’ve already seen.
• RCTF + Examples (few-shot) — when the Role/Context/Format isn’t enough to nail the voice. Paste 2–3 real examples of the kind of output you want, then ask for more in the same style. Dental marketing often has subtle voice signatures (warm-but-credible, confident-but-not-salesy) that examples teach faster than instructions.
• RCTF + Grounding — when Claude tends to invent claims. Add “Never invent statistics or clinical claims. Mark any claim that needs verification by wrapping it in [VERIFY: …]. If you’re unsure, say ‘data not available’ rather than guessing.” This is especially important for anything client-facing.
• RCTF + Self-check — for high-stakes deliverables. End your prompt with “After writing, review your output against the Role and Format constraints. Flag anything that doesn’t meet them and propose a fix.” Claude will often catch its own issues on the re-read.

WEO-Specific Overlay — The Three Iteration Moves

When the first response isn’t quite right, you have three moves. Pick deliberately — piling on follow-ups without a plan is the most common time-waster with Claude.

1. Steer — same chat, same direction, narrower

Use when the output is close but needs refinement. Keep Claude’s working context intact.

You: “Make #3 punchier. Cut the word ‘quality.’ Try again.” Claude: (produces refined #3) You: “Yes — that’s the one. Regenerate #1 and #2 in that same voice.”

Steer is right when you’re adjusting 10–20% of the output. Beyond that, the chat accumulates conflicting instructions and starts producing mush. That’s the cue to move on.

2. Pivot — same chat, different direction

Use when the approach is wrong but you don’t want to re-explain everything (brand, audience, constraints).

You: “These feel too corporate. Try three more that feel like a family talking to another family — warmer, less ‘agency voice.’ Same audience, same constraints.”

Pivot is faster than starting over because Claude still has your Context loaded. If your pivot requires re-stating half the original Context, you should be using Edit-and-Resend instead.

3. Edit-and-Resend — replace the original prompt

Use when your original prompt was actually the problem — missing Context, wrong Role, vague Format.

Click the edit icon on your original message. Rewrite. Resend. The chat forks cleanly from that point.

Original: “Write homepage headlines for a dental practice.” Edited: “Role: Senior dental marketing copywriter for family practices. Context: Smile Springs in Columbus, OH — families with kids + adults 35–55, warm and plainspoken brand, Saturday appointments as the key differentiator. Task: 5 homepage headlines. Format: Max 8 words each, numbered, no emojis, no ‘world-class’.”

Edit-and-Resend is cleaner than piling on revisions because Claude sees the intended prompt, not the accumulated correction history. Chats where you pile on 5+ revisions often have Claude trying to satisfy contradictory instructions from turns 1, 3, and 5.

The 3-revision rule. If you’ve iterated three times and haven’t landed, stop and edit the original. You’re missing a Context element or a Format constraint. Another three revisions won’t fix what the original prompt structurally lacked.

Prompting for Different Output Types

RCTF is the baseline. Different kinds of work weight the four labels differently.

Creative output (headlines, hero copy, ad variants, taglines)

Weight: Role HEAVY, Context medium, Format LOOSE.

Creative work is a voice problem first, a fact problem second. The Role sets the voice. Format gives shape but shouldn’t over-constrain — variants are the whole point.

Role: You are [SPECIFIC PERSONA — e.g. a senior dental marketing
copywriter with 10 years of family-practice experience, direct-response
background, allergic to corporate jargon].

Context: [CLIENT, AUDIENCE, BRAND VOICE, DIFFERENTIATOR, 1-2 BANNED
WORDS OR PHRASES].

Task: Produce 5 [DELIVERABLE] variations spanning a voice range — two
that lean toward [VOICE A, e.g. emotional], two toward [VOICE B, e.g.
confident/factual], and one wild-card that could work as a loose cannon.

Format: Numbered list. Under each variant, one line naming the voice it
leans toward and one line flagging the single tradeoff.

The “spanning a voice range” move is the difference between 5 near-identical headlines and 5 meaningfully different starting points. Asking Claude to name the tradeoff after each forces self-awareness in the output.

Analytical output (teardowns, research, reports, audits)

Weight: Context HEAVY, Task precise, Format STRICT.

Analytical work is a precision problem. Claude needs to know exactly what to look at and exactly how to report. Fuzzy Context or loose Format produces analysis that sounds insightful but can’t be acted on.

Role: [ANALYST PERSONA — senior SEO auditor / competitive analyst /
content strategist specializing in dental vertical].

Context: [WHAT YOU'RE ANALYZING, the full asset. PASTE IT IN or attach
as a file. DON'T describe it in prose — Claude needs the actual thing.]

Task: Produce a teardown covering [3-5 SPECIFIC DIMENSIONS — e.g. 1)
keyword targeting, 2) content structure, 3) CTA strategy, 4) authority
signals, 5) missed opportunities]. For each dimension: one-paragraph
diagnosis, evidence from the asset (quote or cite), and a specific
recommendation.

Format: Markdown. H2 per dimension. Under each: "Finding," "Evidence,"
"Recommendation." No preamble, no conclusion — jump straight in.
Never invent claims or statistics. If uncertain, say "data not
available" rather than guessing.

The “no preamble, no conclusion” constraint saves editing time — Claude otherwise opens with “I’d be happy to help…” and closes with “In conclusion, this analysis reveals several actionable insights.” Nobody needs that.

Editing output (rewrite, tighten, translate, voice-shift)

Weight: Task HEAVY, Role calibrates voice, Context = the source material.

Editing is a constraint problem. You have fixed content; you need a fixed transformation. Claude needs to know what to preserve and what’s fair game to change.

Role: [EDITOR PERSONA — copy chief / brand voice custodian / tightener].

Context: [PASTE SOURCE MATERIAL — the existing content, verbatim.]

Task: [SPECIFIC TRANSFORMATION — e.g. "Tighten by 30% without losing the
three key selling points. Keep the testimonial quote verbatim. Rewrite
everything else."]

Format: [TARGET FORMAT — same structure as source, or specific new
structure.]

Preserve: [WHAT MUST NOT CHANGE — e.g. "Patient testimonial quote stays
word-for-word. Brand tagline stays. Service names stay. Everything
else is editable."]

Change: [WHAT YOU WANT DIFFERENT — e.g. "Cut corporate phrasing.
Shorter sentences. Active voice. Replace 'utilize' → 'use' throughout."]

The Preserve / Change split is the key move. Without it, Claude will either barely change anything (because it doesn’t know it can) or change things you wanted intact (the testimonial quote, brand tagline, etc.).

Structured output (tables, JSON, lists, comparison matrices)

Weight: Format HEAVY, everything else supporting.

When you need structured output for downstream use (pasting into a spreadsheet, feeding to another tool, importing to a CMS), get the Format exactly right the first time — fixing structured output by hand is tedious.

Role: [DATA PERSONA — research analyst / database-shaped thinker].

Context: [SOURCE MATERIAL or criteria for the list.]

Task: Produce [SPECIFIC QUANTITY] of [ITEM TYPE] with these attributes:
[ATTRIBUTE 1 — definition and example value], [ATTRIBUTE 2 — definition
and example value], [ATTRIBUTE 3 — definition and example value].

Format: Markdown table. Columns exactly in this order: [COL1 | COL2 |
COL3]. Row 1 is header. No leading or trailing rows. No prose before
or after the table.

Rules: If a value is unknown, use the literal string "UNKNOWN" —
never guess or leave empty. Limit each cell to [CHAR LIMIT] characters.

The “UNKNOWN” vs empty vs guess rule is the subtle one. Without it, Claude silently fills gaps with plausible-sounding invention, which is the most dangerous failure mode in structured output (because it looks real).

The Quality Bar — Pre-Send Checklist

Before hitting send on any prompt above a one-liner, ask yourself five questions. Any “no” is a signal to keep editing.

1. Would a new hire understand what I’m asking? If your prompt assumes context only you have (the client’s situation, the brand quirks, the deadline), Claude has even less of that context than a new hire would. Make implicit things explicit.
2. Could this prompt produce ten genuinely different acceptable answers? If yes, you’re under-constrained. Add Format or Constraints to narrow the acceptable output space. The goal isn’t “Claude produces something I can use” — it’s “Claude produces close to what I pictured.”
3. Can I picture what a GREAT answer looks like? If you can’t describe a great answer, Claude can’t produce one. Spend 30 seconds visualizing the shape of a good output before writing the prompt — you’ll write a better prompt as a result.
4. Can I picture what a BAD answer looks like — and am I guarding against it? Every deliverable has 2–3 common failure modes (“too corporate,” “too generic,” “includes banned phrasing”). Name them as constraints in the prompt. Prevention is cheaper than correction.
5. Am I asking for one output or several? If it’s several (e.g. “headlines plus meta descriptions plus ad copy”), split them into separate prompts — or separate numbered sections in the same prompt with clear Formats each. Mixing output types in one ask produces muddy results.

This checklist takes 20 seconds. It saves multiples of that in iteration.

When Output Goes Wrong — Five Symptoms and Moves

Most failed Claude outputs trace to a structural prompt gap, not a Claude failure. Learn to diagnose from the symptom:

Symptom 1 — Generic corporate output

“In today’s fast-paced world, dental health is more important than ever…”

Diagnosis: Missing or weak Role. Missing or weak audience in Context. Claude defaulted to “neutral-polite copywriting voice” which is generic by construction.

Move: Strengthen the Role with a specific persona. Name the audience concretely (not “people interested in dental implants” but “adults 45–65 comparing implants to bridges, who’ve already seen a consult”). Banned-phrase list helps: “Never use ‘in today’s fast-paced world,’ ‘state-of-the-art,’ ‘world-class,’ ‘revolutionize,’ ‘game-changer.‘”

Symptom 2 — Factually wrong or invented

Claude confidently states that a procedure cures something it doesn’t, or cites a statistic that turns out to be made up.

Diagnosis: No grounding instructions. Claude’s default mode is to produce fluent prose, and when it lacks a fact, fluent prose trumps accuracy.

Move: Add grounding: “Never invent statistics or clinical claims. Mark any claim requiring verification as [VERIFY: …]. If uncertain, say ‘data not available’ rather than guessing.” For client-facing work, require Claude to cite source material you’ve provided. For healthcare-adjacent content, treat every clinical-sounding statement as [VERIFY] until a clinician signs off (Module 7 covers this).

Symptom 3 — Wrong format

You asked for a table, got prose. You asked for 5 variants, got 3 long ones. You asked for under 50 words, got 90.

Diagnosis: Format under-specified, or Format appears buried in the prompt and gets deprioritized.

Move: Make Format the last thing in the prompt — recency bias applies to prompts too. Be specific: “Markdown table, columns exactly: Headline | Target Audience | Voice Lean | Tradeoff. 5 rows. No header row annotations. No prose before or after the table.” If Claude repeatedly ignores a constraint, ask: “You wrote 90 words; I asked for under 50. Rewrite under 50 — cut the example, keep the claim.”

Symptom 4 — Sycophantic or too agreeable

Claude rewrites your rough outline and tells you it’s already great. You ask if your headline is strong and Claude says yes when it’s actually weak.

Diagnosis: Claude defaults to supportive mode. Without explicit permission to push back, it hedges.

Move: Add: “Push back if you disagree. If my draft is weak, say so directly and propose a stronger alternative. Don’t hedge. I’d rather hear hard feedback now than produce weak work.” For critique requests specifically: “Role: senior editor critiquing an intern’s draft. Be direct. Name what’s weak and what’s actually good — don’t just list everything as ‘solid.‘” See Anthropic’s tutorial on sycophancy for the full picture.

Symptom 5 — Right content, wrong tone

Everything Claude says is correct, but it reads like a different brand. Word choice is off. Sentence rhythm is off. It feels like someone else wrote it.

Diagnosis: Role and brand voice under-specified. Generic Role (“dental marketing copywriter”) produces generic voice.

Move: Either use a Project with the brand guide loaded (Module 4) — the right answer for ongoing clients — or paste 2–3 examples of existing on-brand content directly into the prompt as few-shot references, then ask Claude to match the voice of those examples. Voice transfer via examples is substantially more reliable than voice transfer via instructions.

Common Mistakes (pattern-matched from WEO team feedback)

• No Role. → Claude defaults to “helpful assistant voice” and your copy reads generic.
• No audience in Context. → Claude writes for everyone, which means it writes for no one.
• Task is a vague noun. → “Write something about implants” vs “Write 3 variations of a 50-word service-page introduction for dental implants, audience patients 45–65.”
• No Format. → Claude picks a format at random (prose when you wanted bullets, long when you wanted short).
• Prompting in isolation when context lives in a file. → Upload the brand guide to a Project (Module 4) instead of re-pasting it every chat.
• Piling on revisions instead of editing the original. → Past three revision turns, edit the original prompt. Structural prompt gaps don’t fix themselves through more turns.
• No banned-phrase list. → You know your three pet-peeve corporate phrases. Tell Claude not to use them. The constraint is cheaper than the correction.
• Treating Claude like a search engine. → Chat is a conversation — you can set up context, iterate, and build. A single-shot prompt throws away that advantage.

Building Your Personal Prompt Library

Three months from now, the highest-leverage thing you’ll have built is a personal prompt library of 10–20 prompts you actually use. Start it in week one; don’t wait until you have a “finished” prompt. Libraries grow through iteration, not through planning.

Where to keep it

Wherever you already keep notes — Notion, Apple Notes, Obsidian, a Google Doc. The tool doesn’t matter. What matters is: one prompt per entry, searchable, and you actually open it when you start a task.

A working prompt library for an SEO strategist might look like:

📁 My Prompt Library
  📄 seo_topic-cluster-research.md
  📄 seo_meta-desc-batch-10.md
  📄 seo_keyword-gap-vs-competitor.md
  📄 seo_content-audit-single-url.md
  📄 content_blog-outline-from-keyword.md
  📄 content_rewrite-for-brand-voice.md
  📄 support_client-complaint-response.md
  📄 general_summarize-long-thread.md

Naming convention

{role}_{task}_{optional-variant} — lowercase, hyphenated. Role prefix makes library searchable by context; task names make it scannable.

Template placeholders

Save prompts with {{PLACEHOLDER}} syntax for the variable parts. Example entry:

Role: You are a senior dental marketing copywriter specializing in
family practices.

Context: Practice is {{PRACTICE_NAME}} in {{CITY}}. Audience is
{{AUDIENCE_DESCRIPTION}}. Brand voice is {{VOICE}}. Key
differentiator: {{DIFFERENTIATOR}}.

Task: Write {{NUMBER}} homepage headline variations.

Format: Max 8 words each, numbered list, no emojis, no
"world-class" or "state-of-the-art."

When you use it, fill in placeholders with the actual client details. Over time, you’ll build up 2–3 “flavored” versions of common prompts — the base template plus specialized variants for family practices, pediatric, cosmetic, etc.

When to add a prompt to the library

Not on first success — on second use. The signal that a prompt is library-worthy: you’d otherwise have to recreate it from memory. If the task comes up twice in a month, the prompt should be in your library.

When to retire or evolve

Every 90 days, skim the library. For each prompt, ask: “Is the current output still what I want?” Claude improves over time (new model releases), Memory makes some constraints redundant, and your own standards evolve. Prompts that used to need 5 constraints to work might now work with 2. Slim them down. If a prompt consistently produces output you then edit heavily, it’s under-specified — add the edits you always make as constraints in the prompt itself.

The Module 3 baseline

Module 3 has 25 prompts across 5 roles, each with before/after pairs showing the RCTF upgrade. Start by copying the 5 prompts for your role into your personal library. Use them for a week. Evolve them based on what you actually produce. That’s your first library version.

Prompt Hygiene — Chat Lifecycle

Knowing when to start a new chat is as important as knowing how to prompt.

Chats degrade over ~20 turns

Two mechanics cause this:

1. Accumulated constraints. Every turn adds instructions, clarifications, and corrections to the context. After 20+ turns, Claude is trying to satisfy 50+ implicit rules, some of which contradict. Output gets mushy.
2. Attention dilution. Claude’s context has limits. Past a certain length, early instructions (your original Role and Context) compete for attention with later conversation. You paid good attention up front; don’t let it erode.

Signal: outputs that used to be sharp on turn 3 are muddy on turn 25, even though you’re asking for similar things. Time for a new chat.

The Reset Move

Don’t abandon the accumulated work — consolidate it. At the end of a long productive chat, ask:

“Summarize the decisions we’ve made in this chat and the final prompt pattern that worked best. Format as a reusable template I can paste into a new chat.”

Claude produces a clean handoff artifact. You open a new chat, paste it as the opening message, and continue fresh with all the hard-won context preserved and none of the accumulated noise.

When to start a new chat

• Switching tasks. New deliverable = new chat. Don’t squeeze three different jobs into one thread “because the context is similar.”
• Switching clients. Always. Cross-client context bleed is a governance risk (Module 7) and a quality risk.
• Past the 20-turn mark. Reset move → new chat.
• When output quality suddenly drops. Usually means context pollution, not Claude having a bad day.
• When you’re about to share the conversation with someone else. New chat means they see a clean setup, not a 40-turn archaeology dig.

When to stay in the chat

• Same deliverable, same context, refining. That’s exactly what chat is for.
• Iterating on a draft Claude already produced. Steer and Pivot work. Don’t leave mid-iteration.
• Building on a successful previous turn. If turn 4 nailed the tone, turn 5 can request a variation and Claude remembers the tone.

Chats vs Projects

Chats are for tasks. Projects (Module 4) are for clients and workstreams. Rule of thumb: if the setup you’re typing at the start of a chat would apply to your next 10 chats too, that setup belongs in a Project’s custom instructions, not in the chat.

Key Takeaways

• Prompting is the highest-leverage skill in this whole course. Two weeks of intentional practice gets you competent; two months gets you fluent.
• RCTF = Role + Context + Task + Format. Use it every time. Each label fixes a different ambiguity.
• Variations when you need them: +Constraints (guardrails), +Examples (few-shot voice), +Grounding (anti-hallucination), +Self-check (high-stakes QA).
• Different work weights RCTF differently: creative weights Role, analytical weights Context + Format, editing weights Preserve/Change, structured weights Format.
• Three iteration moves: Steer, Pivot, Edit-and-Resend. Past three revisions, edit the original.
• The five-question pre-send checklist saves multiples of its cost in editing time.
• Five common failure symptoms trace back to specific prompt gaps — learn to diagnose from the output shape.
• Build a personal prompt library. Name by role_task. Use {{PLACEHOLDERS}}. Retire what’s stale.
• Chats degrade over ~20 turns. Use the Reset Move to preserve decisions and start fresh.
• The WEO Module 3 playbook is your starting-point library — 25 RCTF prompts across 5 roles with before/after.

Related

• Next: Module 3 — Role Playbooks by Discipline
• Exercise Set 2 — Fix These Prompts (20 min, hands-on RCTF practice)
• Anthropic Prompting Best Practices (wiki reference) (go deeper once RCTF feels natural)
• Anthropic Academy — AI Fluency: Framework & Foundations
• Anthropic: Getting good at Claude — a research-backed curriculum

Try It (25 min, hands-on)

Three exercises, in order. Do all three — they build on each other.

1. The Before/After diff (8 min)

Take one vague prompt you would’ve written yesterday — something you actually wanted help with last week. Rewrite it with RCTF.

Run both versions in Claude.ai in two separate chats (not same chat — context would pollute the comparison). Read the outputs side by side. Notice specifically:

• Which parts of the After output are obviously better?
• Which parts are worse or unchanged?
• What labels were you under-using? (Usually Role and Format.)
• What would RCTF+Constraints have caught that you had to edit by hand?

Save the winning prompt as the first entry in your personal prompt library.

2. Diagnose-and-fix (10 min)

Find a recent Claude output you weren’t happy with. Use the When Output Goes Wrong framework above to diagnose which symptom applies (1 of 5). Identify the specific prompt gap that caused it. Rewrite the prompt to fix that gap. Run it. See if the symptom goes away.

This is the single most valuable exercise in this module. Diagnosing your own bad prompts is how prompting becomes a skill rather than a recipe.

3. The Output-Type templates (7 min)

Pick your primary role. Look at the four Output-Type templates above (Creative / Analytical / Editing / Structured). Which one matches the task you do most often?

Copy that template into your prompt library. Fill in the placeholders with a real recent task from your work. Run it. Evolve the template based on what you learn. Name it {role}_{task}.md.

4. (Optional, recommended) Work through Exercise Set 2 — Fix These Prompts

Five more reps of diagnosing and fixing. Most valuable if you did Exercise 2 above and want more practice with the failure-mode framework.

Done? Move on to Module 3 — the WEO role playbook has 25 worked RCTF prompts ready for your library.