CRAFT Framework Prompts: Context, Role, Action, Format, Tone

A craft framework prompt uses five labeled fields: Context, Role, Action, Format, and Tone. You front-load facts, name who speaks, state one verb-led job, lock output shape, and set channel voice. This deep dive matches PromptMake’s live CRAFT generator at https://promptmake.net/frameworks/craft. It is not the frameworks hub, not the CO-STAR deep dive, and not the RISEN deep dive. You leave with field definitions, a blank template, six paste-ready examples, a ten-minute build path, review checks, and an FAQ for ChatGPT, Claude, and Gemini work in September 2026. Keep Format countable so teammates can score outputs without taste debates on every run.

What CRAFT is and who should use it

CRAFT is a five-part chat prompt skeleton. Context holds situation and data. Role sets perspective. Action is the single primary task. Format defines structure and length. Tone names voice for the channel. PromptMake documents this expansion on the CRAFT product page: Context, Role, Action, Format, Tone.

Teams pick CRAFT when CO-STAR’s six fields feel heavy and RISEN’s Steps feel like overkill. Standups, meeting notes, internal wiki updates, short briefs, and ticket summaries fit. External brand campaigns with strict audience splits often still want CO-STAR. Multi-step audits often still want RISEN.

Solo builders use CRAFT as a daily checklist. Fill five lines faster than you debate acronyms. Soft drafts from rough notes: https://promptmake.net/frameworks/craft.

Skip CRAFT for one-line lookups and for API calls that already ship JSON schema. Lean in when background data and voice both matter on a short internal job.

The five CRAFT fields explained

Treat each field as a contract line. Weak CRAFT dumps vibes into Context and stacks three verbs in Action. Strong CRAFT pastes evidence, picks one Action verb, and makes Format measurable. Tone names the channel: Slack update, board memo, customer email, technical precise.

Order matters on PromptMake’s teaching page: Context before Role so the model reads facts before persona. That order cuts answers that sound confident while ignoring pasted data.

Copy this blank block into your doc:

Context: [background and data]

Role: [perspective]

Action: [primary task]

Format: [output structure]

Tone: [voice]

Put long source paste after the five fields so the contract appears first.

Context and Role

Context holds evidence: ticket tags, meeting bullets, product facts, deadlines, channel, constraints. "Q3 support themes: billing confusion, export failures" beats "customers are unhappy." Paste summaries or key excerpts; avoid dumping entire PDFs when a short excerpt works.

Role names who the model writes as or for: support lead for standup, analyst for VP skim, coach for a junior writer. Keep Role short. Inflated biographies waste tokens and rarely change outcomes on modern models.

Weak Context is empty. Weak Role is "helpful assistant." Strong pairs give facts plus a clear reader or speaker seat.

Action, Format, and Tone

Action is singular and verb-first: summarize, draft, compare, extract, rewrite. Split sub-tasks as bullets under Action only when they serve one deliverable. If you need research then write then edit as three products, run three CRAFT prompts or switch to RISEN for Steps.

Format kills vague "be concise." Prefer "5 bullets, max 15 words, start each with a verb" or "Markdown table with columns Feature | Risk | Owner." Tone names channel voice: neutral factual, casual Slack, formal report, empathetic support.

Add constraint bullets under Action or Format when needed: do not invent metrics, mark unknowns TBD, max 120 words.

Blank CRAFT template and filled skeleton

Use the empty template from PromptMake’s CRAFT page as your team default. Keep labels identical across writers so review is a scan, not a hunt.

Empty:

Context: [background and data]

Role: [perspective]

Action: [primary task]

Format: [output structure]

Tone: [voice]

Filled skeleton from the product example pattern:

Context: Q3 support ticket themes attached: top tags billing confusion, export failures, mobile login.

Role: Support team lead writing for internal product standup.

Action: Summarize patterns and propose two process fixes.

Format: 5 bullets covering pattern, frequency, suggested fix; max 15 words per bullet.

Tone: Neutral and factual, no blame language.

Soft path when you only have a messy paragraph: https://promptmake.net/frameworks/craft expands rough input into labeled sections plus a combined paste block.

Paste-ready CRAFT examples

Each example below is ready to copy. Replace bracketed sources. Keep Format ruthless.

Example 1: standup summary

Context: Yesterday’s engineering notes: API latency spike on checkout, fix deployed at 16:40 UTC, two flaky tests quarantined.

Role: Eng manager writing for a cross-functional standup.

Action: Summarize status and name today’s owners.

Format: 4 bullets covering item, owner, status, next check time.

Tone: Direct Slack voice, no jargon lectures.

Example 2: customer email draft

Context: Trial user hit export CSV failure; workaround is JSON export; permanent fix ships Friday.

Role: Support agent writing to a non-technical account owner.

Action: Draft an apology email with workaround and fix date.

Format: Subject line + body under 140 words + one CTA.

Tone: Empathetic and clear, no blame on the customer.

Example 3: meeting notes to decisions

Context: [paste transcript excerpt]

Role: Chief of staff writing for attendees who missed the call.

Action: Extract decisions and open questions only.

Format: Two headers (Decisions, Open questions), bullets only.

Tone: Neutral minutes voice, no commentary.

Example 4: product brief slice

Context: Feature: usage caps on free tier; goal reduce abuse; must not scare paying teams.

Role: Product marketer writing for landing-page draft review.

Action: Draft three benefit bullets and one FAQ answer.

Format: Bullets under 25 words; FAQ answer under 60 words.

Tone: Confident, plain English, no hype adjectives from the ban list in Context.

Example 5: data extraction

Context: [paste invoice table text]

Role: Ops assistant for finance intake.

Action: Extract vendor, amount, due date, and PO number.

Format: JSON object with keys vendor, amount, due_date, po_number; null if missing.

Tone: Precise, no prose outside JSON.

Example 6: rewrite for channel

Context: Long internal wiki paragraph about onboarding steps.

Role: Docs editor writing for new hires in week one.

Action: Rewrite as a checklist new hires can finish in 30 minutes.

Format: Numbered list, max 8 steps, one verb each.

Tone: Friendly coach, short sentences.

Example 7: competitive snippet for Slack

Context: Competitor launched usage-based pricing yesterday; our reply is "we already include overages in Pro."

Role: Product marketer writing for the #gtm channel.

Action: Draft a short internal brief on what changed and our talking point.

Format: 3 bullets covering change, our response, who owns follow-up.

Tone: Calm Slack voice, no dunking language.

Build a CRAFT prompt in ten minutes

Step 1: Name the recurring job in one line. Step 2: Paste facts into Context. Step 3: Set Role from reader or speaker. Step 4: Write one Action verb. Step 5: Make Format countable. Step 6: Name Tone by channel. Step 7: Run once on GPT-5.6 Sol, Claude Fable 5, or Gemini 3.5 Flash. Step 8: Score against Format. Step 9: Store the winner with date and model. Step 10: Link https://promptmake.net/frameworks/craft for future rough-to-labeled drafts.

Solo review pass

Read Format first. Does the output match structure and length? If not, tighten Format before you rewrite Role. Read Context. Did the model invent numbers? Move honesty rules under Action or Format.

Save failures as negative examples in your wiki. CRAFT templates rot when product names change.

Team handoff pass

One owner per template. New hires edit Context only in week one. Brand or legal owns Tone phrases for customer-facing CRAFT jobs. Monthly review deletes templates tied to retired workflows.

CRAFT versus CO-STAR, RISEN, and RTF

RTF is Role, Task, Format. CRAFT adds Context before Role and splits Task into Action plus Tone. Use RTF for instant asks. Use CRAFT when data and voice need labels.

CO-STAR adds Style, Audience, and Response as separate slots. CRAFT is faster for internal tasks. CO-STAR is richer for external comms where audience and response format need their own lines. Soft CO-STAR path: https://promptmake.net/frameworks/costar.

RISEN adds Steps, End goal, and Narrowing. Pick RISEN when sequence and hard bans dominate. Soft RISEN path: https://promptmake.net/frameworks/risen. Pick CRAFT when one Action and a clear Format are enough.

The hub prompt-frameworks-compared lines all four. This page implements CRAFT only.

Common CRAFT mistakes and fixes

Mistake: Action lists three unrelated jobs. Fix: split prompts.

Mistake: Format says concise. Fix: count bullets and words.

Mistake: Empty Context with a fancy Role. Fix: paste facts first.

Mistake: Tone contradicts Format (casual novel when you asked for JSON). Fix: align channel and structure.

Mistake: Treating generator output as verified customer copy. Fix: human fact pass.

Mistake: Using CRAFT for a six-step audit. Fix: move to RISEN.

Using PromptMake’s CRAFT generator

Open https://promptmake.net/frameworks/craft. Describe the task or paste a rough draft. The tool returns Context, Role, Action, Format, and Tone plus a combined prose block for hosts that prefer paragraphs.

PromptMake generates text only. Guests get about three text generations per day per path. Free accounts get about five. Confirm current limits on promptmake.net. Edit every fence before production send.

Daily habit: keep Format and Tone stable for recurring jobs; swap Context and Action per run. That stability makes evals comparable week to week.

Daily CRAFT workflow that compounds

Morning: paste yesterday’s notes into Context, keep Role and Format from your standup template, change Action only if the deliverable changed. Midday: run ticket replies with a fixed Tone line your support lead approved. Evening: rewrite one long wiki paragraph with the checklist Format from example six.

Weekly: pick one CRAFT job and A/B Format wording. "5 bullets" versus "5 bullets, max 12 words, verb-first" usually cuts edit time. Log the winner next to the template date.

Monthly: delete templates nobody ran. Link survivors to https://promptmake.net/frameworks/craft so new hires bootstrap from rough notes without inventing a sixth field.

When a CRAFT job grows Steps, graduate it to RISEN instead of stuffing procedure into Action. When a CRAFT job needs Audience as its own line for brand risk, graduate it to CO-STAR. Graduation is success, not failure.

Eval scorecard

Score each run 0 or 1 on four checks: used Context facts, obeyed Format, matched Tone, avoided invention. Four points means ship. Two or fewer means fix the template before blaming the model.

Naming note on Capacity versus Context

Some content plans label CRAFT with Capacity or Target. PromptMake’s live framework page uses Context, Role, Action, Format, and Tone. Match the product you paste into so team docs and generators stay aligned.

FAQ

What does CRAFT stand for in a craft framework prompt?

On PromptMake, CRAFT stands for Context, Role, Action, Format, and Tone. Context front-loads facts. Role sets perspective. Action states one primary task. Format locks structure. Tone sets channel voice. Other sites sometimes redefine the letters; match the generator you use.

When should I use a craft framework prompt instead of CO-STAR?

Use CRAFT for shorter internal jobs where five fields are enough. Use CO-STAR when Style, Audience, and Response need separate lines for external brand work. Soft CRAFT drafts: https://promptmake.net/frameworks/craft.

How is this different from prompt-frameworks-compared?

The hub compares CO-STAR, RISEN, CRAFT, and RTF for picker decisions. This craft framework prompt page teaches CRAFT field-by-field with paste-ready examples. Read the hub to choose among frameworks; read this page to implement CRAFT.

Can CRAFT prompts include constraints?

Yes. Add constraint bullets under Action or Format: do not invent metrics, mark unknowns as TBD, max word counts, banned phrases. Keep Action singular so constraints support one job instead of hiding a second task.

Does Context belong before Role?

Yes on PromptMake’s CRAFT pattern. Context first reduces persona-led answers that ignore pasted data. Put long sources after the five labeled fields so instructions stay visible.

Which models work with CRAFT in 2026?

CRAFT is model-agnostic labeled prose. GPT-5.6 Sol, Claude Fable 5, Gemini 3.5 Flash, and fast chat tiers all follow the five headers when Format is concrete. Skip chain-of-thought filler on reasoning-class models; keep goal, constraints, and format clear.

How do I start with CRAFT on PromptMake for free?

Open https://promptmake.net/frameworks/craft, describe your job in plain language, and edit the five fields. Guest and free account limits apply per text path. Register if you hit the guest cap mid-template build. Store approved templates in your wiki for reuse.