CO-STAR Framework Prompts: Context to Response Format

The costar framework gives you six labeled slots for chat prompts: Context, Objective, Style, Tone, Audience, and Response. You separate background, goal, writing genre, emotional register, reader, and output shape instead of burying them in one vague paragraph. This page is a CO-STAR deep dive with field definitions, a blank template, paste-ready examples, review checks, and a soft path to https://promptmake.net/frameworks/costar. It is not the multi-framework hub and not the RISEN deep dive. You leave ready to write CO-STAR blocks for GPT-5.6 Sol, Claude Fable 5, and Gemini 3.5 Flash work where tone and audience matter as much as the task verb.

What the costar framework is and who should use it

CO-STAR is a prompt skeleton popularized in workshop and competition settings and widely taught for marketing, support, and operations copy. Each letter names one decision teams skip under deadline pressure. Context holds facts. Objective states the outcome. Style names the genre. Tone names warmth or steel. Audience names who reads the result. Response locks format and length.

Teams adopt CO-STAR when customer-facing voice drifts across writers. Shared labels make review a ten-second scan. New hires fill slots instead of copying a magic paragraph from a senior coworker. Solo builders use CO-STAR when a polite draft still misses the reader or the deliverable shape.

CO-STAR shines on emails, landing sections, training content, and internal memos where wrong tone costs trust. It is heavier than RTF for simple extraction. It differs from RISEN by front-loading Style, Tone, and Audience instead of Steps and Narrowing. Soft draft path: https://promptmake.net/frameworks/costar.

Skip CO-STAR for one-line lookups and for API calls that already use JSON schema. Lean in when external voice risk is high.

The six CO-STAR fields explained

Treat each field as a contract line. If you cannot fill Audience with a real reader, the model will write for a generic stranger. If Response lacks length and structure, you will get an essay when you wanted a table. The sections below define each letter, show failure modes, and give fill patterns that survive team review on ChatGPT, Claude, and Gemini as of September 2026.

Copy this blank block into your doc and fill it once per recurring job:

Context: [situation and facts the model must respect]

Objective: [one verb-led outcome sentence]

Style: [genre or register label]

Tone: [emotional register]

Audience: [reader role, knowledge, time]

Response: [format, sections, length, language]

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

Context and Objective

Context holds evidence, not vibes. Paste product facts, prior message snippets, policy lines, and constraints the model must not invent around. "SaaS billing product, duplicate charge ticket, refund in three business days" beats "customer is upset."

Objective is one verb-led sentence: draft, summarize, rewrite, compare, outline. Measurable where possible. "Draft a refund approval email under 150 words" beats "help with this ticket." If you need two unrelated outcomes, split into two prompts.

Weak Context dumps a novel. Strong Context is short and factual. Weak Objective stacks three verbs. Strong Objective picks one primary verb and parks secondary asks in Response as sections.

Style, Tone, Audience, and Response

Style names the genre: support email, QBR slide outline, lesson plan, LinkedIn post, executive memo. Tone names feeling: empathetic and policy-clear, direct with no hype, warm but brief. Teams often blur Style and Tone. Style is the form. Tone is the emotional dial inside that form.

Audience names who reads the result: frustrated SMB admin, VP who scans on mobile, new hire in week one. Audience calibrates vocabulary and assumed knowledge. Missing Audience is the most common CO-STAR failure in review.

Response defines structure and length: subject line plus body under 150 words, markdown with three headers, table with five rows, US English, bullets only. Response is where you stop the model from adding a closing essay when you asked for a checklist.

CO-STAR template you can paste today

Use this shell for any new job. Replace bracketed lines. Keep labels so teammates can scan.

Context: [facts, product, prior thread, policy]

Objective: [verb + object + success criterion]

Style: [genre label]

Tone: [emotional register]

Audience: [role, expertise, time pressure]

Response: [format + length + section names + language]

Source: [paste evidence below]

This shell works in ChatGPT, Claude, Gemini, and inside Custom GPT or Gem instructions when you want a user-message pattern the team repeats. Standing system prompts can hold brand voice; CO-STAR still helps each task message stay complete.

Paste-ready CO-STAR examples

Six teaching copies below cover support, launch, status, lesson, critique, and audience rewrite. Adapt nouns for your product. Keep all six fields so a reviewer can scan Context through Response in under a minute. These examples target costar framework practice for external and internal jobs where voice risk is real. They are not RISEN step lists and not RTF one-liners. Read one example aloud before you paste it into ChatGPT, Claude, or Gemini so you catch empty Style or Tone labels. After a live win, save the filled block in your team docs with the model name you used.

Examples 1-3: support, launch, and status

Example 1, support refund email:

Context: Duplicate charge on Pro plan invoice INV-4412. Policy: refund within three business days once approved. Customer already waited two days.

Objective: Draft a customer reply that confirms refund approval and sets timing expectations.

Style: Support email.

Tone: Empathetic, policy-clear, no blame theater.

Audience: Frustrated SMB admin who manages billing for a ten-person team.

Response: Subject line; body under 150 words; closing with Support Team; add three internal CRM bullets after a horizontal rule.

Example 2, launch email:

Context: Team project tool launches Tuesday. Current users: 200-person agencies on spreadsheets. Feature: shared timelines with comment threads.

Objective: Draft a launch email that drives free-trial signups from team leads.

Style: Short marketing email, scannable, no jargon.

Tone: Friendly and confident, not hype-heavy.

Audience: Operations and team leads, busy, skeptical of new tools.

Response: Subject line; body under 180 words; single CTA button text.

Example 3, weekly status for a VP:

Context: Payments squad notes from standups. Ship list and blockers pasted below.

Objective: Summarize progress, blockers, and next-week priorities.

Style: Executive mobile email.

Tone: Direct, calm, no hero language.

Audience: VP who scans email between meetings.

Response: Under 200 words; sections Shipped | Blockers | Next week; bullets only; no invented metrics.

Examples 4-6: lesson, critique, and rewrite

Example 4, onboarding lesson:

Context: New support hires need a ten-minute primer on refund policy. Source policy pasted below.

Objective: Turn the policy into a teaching script a lead can read aloud.

Style: Spoken lesson plan.

Tone: Patient, precise, no sarcasm.

Audience: Week-one support hire with no finance background.

Response: Three sections of under 80 words each; end with three quiz questions; US English.

Example 5, landing critique:

Context: Draft homepage hero for a fintech savings app. Current draft pasted below. Brand bans: get rich, guaranteed returns.

Objective: Critique the hero and propose one stronger alternative.

Style: Editorial memo.

Tone: Candor without mockery.

Audience: Founder who wrote the draft and wants blunt notes.

Response: Strengths (3 bullets), Risks (3 bullets), Alternate headline under 8 words, Alternate subhead under 20 words.

Example 6, rewrite for a different audience:

Context: Technical changelog paragraph about API rate limits. Engineers already understand 429 responses.

Objective: Rewrite for customer success managers who explain limits to clients.

Style: Plain-language briefing.

Tone: Helpful, non-alarmist.

Audience: CSMs with light technical literacy and five minutes before a call.

Response: Under 120 words; one analogy max; glossary of three terms as bullets.

Common CO-STAR mistakes

Mistake 1: Filling Style with creative and Tone with friendly. Those words do no work. Name a genre and a concrete dial.

Mistake 2: Skipping Audience. The model then writes for a generic reader and misses vocabulary and length needs.

Mistake 3: Stuffing Objective with three tasks. Split prompts or put secondary deliverables in Response as labeled sections.

Mistake 4: Context as vibes. "Make it premium" without product facts invites invention.

Mistake 5: Response without length. You get essays when you needed five bullets.

Mistake 6: Using CO-STAR for a pure extraction job that RTF would finish faster. Escalate to CO-STAR when tone and audience risk appear.

Mistake 7: Mixing CO-STAR and RISEN headers in one message. Pick one skeleton per job so reviewers know where to look.

Mistake 8: Treating the generator output as final. https://promptmake.net/frameworks/costar drafts labeled blocks. You still edit fences before production.

When to use PromptMake /frameworks/costar

Open https://promptmake.net/frameworks/costar when you want labeled CO-STAR text from a rough brief without memorizing headers. Paste a messy job line. Edit every field. Copy into ChatGPT, Claude, Gemini, or a standing GPT and Skill config. PromptMake returns text only. It does not run the model for you or store company vaults unless your org adds storage separately.

Guest and free quotas apply per tool path. Confirm current limits on promptmake.net. Use /frameworks/risen when Steps and Narrowing matter more than Style and Tone. Use /frameworks/craft when five fields feel enough. Use /text for freeform scaffolds outside named frameworks.

For team adoption, map each recurring external-copy job to CO-STAR in your wiki. Keep the multi-framework comparison hub for pick debates. Keep this page as the costar framework field manual.

FAQ

What is the costar framework?

The costar framework (CO-STAR) stands for Context, Objective, Style, Tone, Audience, and Response. It structures chat prompts so the model receives background, a clear goal, writing genre, emotional register, reader profile, and output format in labeled sections. Teams use it when voice and format consistency matter across writers.

How do Style and Tone differ in CO-STAR?

Style names the genre or form: support email, executive memo, lesson plan, social post. Tone names the emotional register inside that form: empathetic, direct, warm, formal. Mixing them into one vague line is how teams lose control of voice. Fill both with concrete labels a reviewer can check in five seconds.

When should I use CO-STAR instead of RISEN or RTF?

Use CO-STAR when tone and audience risk are high on customer-facing or executive copy. Use RISEN when ordered steps and narrowing constraints drive quality. Use RTF for fast daily extraction and rewrites. This page stays on CO-STAR; the hub article compares frameworks side by side.

Can I generate CO-STAR prompts on PromptMake?

Yes. Open https://promptmake.net/frameworks/costar, paste a rough brief, and edit the labeled output. PromptMake generates prompt text you copy into ChatGPT, Claude, Gemini, or standing configs. It does not execute the task. Guest and free account limits apply per path.

Who created CO-STAR?

CO-STAR is widely associated with practitioner teaching and competition settings, including work popularized by Sheila Teo in public prompt-engineering materials. Exact origin stories vary across blogs. For team use, the six field definitions matter more than lore. Stick to the labels your wiki documents.

Does CO-STAR work with reasoning models in 2026?

Yes. On GPT-5.6 Sol, Claude Opus 5, and similar reasoning-class models, keep Objective and Response sharp and skip cargo-cult "think step by step" lines. CO-STAR still helps by locking Audience and Tone so a strong model does not over-write for the wrong reader. For Flash-class daily chat, CO-STAR remains useful when voice risk is high.

How long should each CO-STAR field be?

Aim for one to four short lines per field. Context can grow when you paste evidence, but keep the label block scannable. If Context exceeds a few paragraphs of facts, move the bulk under a Source heading after Response so the contract stays visible first.