Agent Skills Generator Guide: Claude, GPT & Gemini in One Flow

Agent skills are reusable instruction packs for repeating jobs: Claude SKILL.md, Custom GPT builder fields, Gemini Gem instructions. This guide maps one workflow across all three so you stop rewriting the same playbook three times. You draft once, export per platform, paste into each vendor builder, and test with five samples. PromptMake at https://promptmake.net/skills generates English config text only. It does not install skills, publish GPTs, or save Gems. High keyword difficulty on agent skills rewards clear workflow and honest scope, not one-click deploy claims. You leave with trigger rules, mapping tables, step-by-step loops, copy-paste examples, mistake fixes, and an FAQ for teams shipping their first skill this week.

What agent skills are (and what they are not)

Agent skills are durable rules for a named job: rewrite support replies, grade rubrics, draft release notes, or run a weekly research brief. Claude exposes them as Agent Skills with a SKILL.md file and a short description that tells the runtime when to load the body. OpenAI exposes a parallel idea as Custom GPT instructions plus optional knowledge files inside ChatGPT. Google exposes it as Gemini Gem instructions plus knowledge uploads on gemini.google.com. Same product intent, three paste targets.

Agent skills are not a marketplace download you can trust blindly. They are not a single system prompt that replaces your tools API. They are not magic memory. A skill tells the model how to behave when the job matches. Your runtime still supplies files, retrieval, and permissions. PromptMake does not host a skill store and does not run agents. It helps you draft the text you paste into each builder.

Skip a formal skill when a one-shot chat prompt already works, when the job changes every message, or when you have no stable output shape to test. Ship a skill when the same brief returns every week, when teammates need one named entry point, and when you can write a trigger line that separates this job from general chat.

As of mid-2026, confirm access on vendor help pages before you promise teammates a new Custom GPT or a Claude Code skill path. Workspace rules shift. The workflow below stays valid even when UI labels move: draft, map, preview, save, test with five real samples.

Why one workflow beats three separate essays

Teams stall because they open Claude docs, then ChatGPT builder, then Gemini Gems, and rewrite the same role paragraph three times with slightly different adjectives. Each rewrite drifts. Triggers get vague. Fences drop. A single source draft with explicit sections (ROLE, TASK, FORMAT, FENCES, EXAMPLES) exports cleaner than three improvised blobs.

A cross-platform agent skills workflow also makes evals portable. You run the same five sample inputs against Claude Code with SKILL.md loaded, against a Custom GPT preview, and against a saved Gem. If one platform fails on missing-info behavior, you fix the shared FENCES block once, then re-export. That beats debugging three unrelated instruction essays.

People search agent skills when they mean any standing assistant config, not only Anthropic's file format. This article uses agent skills as the umbrella term and names Claude SKILL.md, Custom GPT, and Gemini Gems where the paste target matters. Winning on a high difficulty keyword means showing a clear workflow and honest scope, not claiming one-click deploy everywhere.

Claude Agent Skills: SKILL.md shape

Claude Agent Skills center on a SKILL.md file in a folder the product expects. The description field at the top is the trigger surface: short, specific, rich in nouns and verbs the runtime can match to user intent. The body holds imperative steps, output sections, and refusal rules. Anthropic's public pattern favors progressive disclosure: keep the description scannable, put long examples and edge cases in the body or linked files, not in the trigger line.

Claude Code users add skills beside project or user skill directories depending on team convention. Claude web and API paths differ by account and product surface. Hedge exact folder names with Anthropic docs at draft time. The writing job stays the same: a trigger line that says when to load the skill, and a body that says what to do once loaded.

PromptMake /skills can emit a SKILL.md-shaped draft from one brief. You still create the folder, add any supporting files, and register the skill in the environment you use. The generator does not push to your repo or Claude account. For deep Claude Code layout notes, pair this page with our claude-code-skills-guide post.

Description triggers that actually fire

Write the description as a job label plus trigger phrases. Bad: "Helpful writing assistant." Good: "Rewrite B2B support replies to match the Acme tone sheet. Use when the user pastes a draft support email or ticket reply and asks for a customer-ready version."

Pack concrete nouns: support email, rubric, lesson plan, release note, SQL review. Pack verbs: rewrite, grade, summarize, extract, refuse. Skip metaphor. Skip "world-class." The runtime matches text, not vibes. Keep the description under the length your surface recommends; put long policy in the body.

Add a negative trigger when collisions are common: "Do not use for legal advice or pricing quotes." That line cuts wrong-skill loads when you run multiple skills in one project. Test with five user phrases that should load the skill and five that should not.

Body steps, format blocks, and fences

Open the body with numbered steps the model can follow in order. Name the output sections every time: Subject lines, Body, Risk flags, or Rubric scores, Feedback, Next actions. Fences belong in the body, not buried in chat: "Never invent refund policy. Write Unknown when the source omits the rule."

Progressive disclosure means the description sells the job; the body carries examples, sample inputs, and edge cases. Link or attach long PDFs instead of pasting them into the description. For Claude Code, keep commands imperative: "Read the pasted draft, apply the tone rules in section 2, return the structured blocks in section 4."

Test the body with a missing-info sample before you share the skill. If the model invents policy on the second sample, strengthen FENCES and add an explicit Unknown path. Store reference files in a references/ folder when the body would otherwise exceed maintainable length.

Custom GPT: the OpenAI agent skills path

A Custom GPT is ChatGPT's standing assistant: name, description, instructions, conversation starters, optional knowledge, and enabled capabilities. For agent skills work, map your shared draft into Instructions (standing rules), Knowledge (stable PDFs), and conversation starters (how humans begin). The GPT description field is partly human-facing in the store and partly model-facing; write it so teammates know which GPT to open.

As of mid-August 2026, OpenAI help docs emphasize that new Custom GPT creation and publishing focus on Business, Enterprise, and Edu workspaces when admins allow it. Personal accounts may still use existing GPTs without new create rights. Confirm before you plan a team rollout. PromptMake does not create GPTs inside ChatGPT. It drafts fields you paste into the builder on chatgpt.com when your account can edit.

Map ROLE and TASK from your shared draft into Instructions. Map FORMAT into a labeled section inside Instructions. Map FENCES into a short Unknown and refusal block. Map EXAMPLES into Knowledge or a collapsed section at the bottom of Instructions, not into every conversation starter.

Builder fields that matter

Name: job-based ("Support reply rewriter"), not vibe-based ("Magic helper"). Description: one sentence for humans plus trigger nouns. Instructions: role, steps, output headers, fences. Conversation starters: three real opening patterns (paste draft, paste ticket, paste rubric). Knowledge: tone PDF, FAQ, rubric file. Capabilities: enable browsing or actions only when the job needs them; extra capabilities invite drift.

Keep ticket detail out of Instructions. Order IDs, client names, and this week's campaign belong in the chat message. Standing rules belong in Instructions. That split mirrors Claude SKILL.md and Gemini Gems, which keeps your cross-platform draft stable.

When the job also needs Midjourney or FLUX dialect, do not stretch one Custom GPT into image syntax. Export chat prompts through PromptMake /text or photo prompts through /image instead. Custom GPTs excel at ChatGPT-native work, not at Discord parameter factories.

Preview and five-sample testing

Preview with a happy path and a missing-info path before Save. After Instructions change, open a fresh chat with the GPT. Old threads may keep older behavior. Run five samples: two happy paths, one wrong input, one empty paste, one out-of-scope ask. Log which fence failed. Freeze prompt text as version N when the set passes.

Snapshot Instructions in git even though ChatGPT hosts the live GPT. When OpenAI shifts UI copy, you regenerate from the canonical brief instead of archaeology in old chats. For builder field drafts, /custom-gpt-generator is an alias on the skills hub when you want GPT-shaped export only.

Gemini Gems: the Google agent skills path

Gemini Gems are saved profiles on gemini.google.com: name, description, Instructions, optional knowledge files, and default tools on many accounts. Agent skills language maps cleanly: Instructions hold ROLE, TASK, FORMAT, and FENCES; knowledge holds long references; chat holds the ticket. Google's public framing uses Persona, Task, Context, and Format. You do not need those labels in the box if the content is there.

PromptMake /skills can target a Gem-shaped export from the same brief you used for SKILL.md and Custom GPT. You still create, preview, and save the Gem in Google's UI. The generator does not upload knowledge files to Google Drive on your behalf. See our gemini-gems-guide for full Gem product context beyond this cross-platform map.

Workspace availability and file caps depend on admin settings. Hedge limits with Google Help at draft time. For education teams, Gems often replace retyping the same rubric instructions every Monday. For support teams, Gems replace a shared doc of macros if Instructions stay tight.

Persona, task, and format mapping

Persona: who the Gem plays (editor, coach, analyst). Task: the repeating outcome (rewrite, outline, grade). Context: what stable facts the Gem can assume (audience, product line, grade band). Format: section headers every reply must include. Paste those four blocks from your shared draft without rewriting tone three ways.

Keep Instructions under the length you can maintain monthly. Long essays rot. Short blocks with clear headers survive handoffs. If the Gem serves multiple grades or locales, split Gems instead of if/else novels in one Instructions field.

Default tools (Canvas, Deep Research, image modes) change the starting mode. Pick a default only when almost every chat with that Gem needs that mode. Leave no default when the job switches often.

Knowledge files and preview discipline

Upload only files that change answers: style guide, SKU sheet, district rubric. Skip unrelated PDF piles. If citations matter, leave knowledge citation behavior on unless policy says otherwise. Preview with a real classroom or ticket sample, not "hello."

After Instructions edits, start a new Gem chat. Share settings only when policy allows. Redact student names and customer PII from test pastes you send through public generators.

One cross-platform workflow on PromptMake /skills

Use one brief for all three exports. Open https://promptmake.net/skills. State the job in one sentence, the audience, the output sections, three fences, and two trigger phrases. Pick the export family you need first (Claude SKILL.md, Custom GPT fields, or Gem instructions). Generate once. Edit invented constraints. Re-run export for the next platform from the same brief notes, not from memory.

Budget thirty minutes for your first skill. Ten minutes on the brief, ten on edits, ten on five-sample tests per platform you ship. Soft sell once: the skills hub generates text; you own paste, preview, save, and version control in git or an internal wiki.

Store the winning brief beside the exported files with date and model tags. When Anthropic or OpenAI shifts UI copy, you regenerate exports from the same brief instead of archaeology in old chats. Guests get about three runs per day on skills paths; free accounts get about five. Quotas stay separate from /text and /image.

Steps 1-3: Draft the shared skill brief

Step 1: Write JOB in one sentence with a verb and deliverable. Step 2: List OUTPUT sections as headers the model must return every time. Step 3: Write FENCES: Unknown rule, refusal rule, and one scope limit (what this skill does not do).

Add two TRIGGER lines: when to use, when not to use. Add one FORMAT example with placeholder text, not customer data. Stop if you cannot name a success check ("returns three subject lines and a body" beats "be helpful").

Paste the brief into PromptMake /skills. Select Claude, GPT, or Gem export. Generate. Delete any invented policy the model added. Add your real product names and links offline if needed.

Steps 4-6: Export, paste, and test per platform

Step 4: Export for platform A. Paste into SKILL.md or GPT Instructions or Gem Instructions. Add knowledge files on GPT/Gem paths if the brief references them. Step 5: Run five samples on platform A. Fix the shared brief if fences fail. Step 6: Re-export for platforms B and C without rewriting role paragraphs from scratch.

For Claude Code, commit SKILL.md to the repo path your team uses. For Custom GPT, snapshot Instructions in git even though ChatGPT hosts the live GPT. For Gems, copy Instructions into an internal doc so edits are diffable.

Promote a skill to teammates only after two people pass the same five samples. Shared skills fail when only the author tested happy paths.

Common mistakes in agent skills projects

Mistake 1: Vague triggers. "Writing help" loads on everything. Fix with nouns and verbs tied to one deliverable.

Mistake 2: Ticket detail inside standing instructions. Client names become stale rules. Move ticket facts to the chat message.

Mistake 3: One mega-skill for every department. Split by output shape. Support rewrites and research briefs fight inside one block.

Mistake 4: Treating the generator output as published. PromptMake text is a draft. You still preview, save, and test in each vendor UI.

Mistake 5: Expecting skills to replace missing tools. A skill cannot call a CRM that does not exist. Skills reduce invention; they do not invent integrations.

Mistake 6: Skipping the missing-info sample. Happy-path-only tests ship skills that invent policy under pressure.

Mistake 7: Chasing marketplace packs instead of owning your brief. Public repos are inspiration. Your fences and tone need local edits.

Model notes for agent skills in 2026

Claude Fable 5, Opus 5, and Sonnet 5 follow labeled contracts and long instruction blocks well. XML-style tags in the body can help if your team already uses them. Keep triggers short.

GPT-5.6 Sol and ChatGPT GPT-5.5 Instant paths reward clear FORMAT sections and refusal fences. Put frozen instructions in the GPT builder; keep volatile facts in the user message.

Gemini 3.5 Flash suits high-volume Gems with short menus. Gemini 3.1 Pro suits longer knowledge-backed Gems. Tighten "only these sections" lines on Flash; measure format drift before you widen scope.

Cross-model habit: same ROLE, TASK, FORMAT, FENCES source brief; platform-specific paste; five-sample eval per surface. Swap models after the brief is frozen so you compare apples to apples.

FAQ

What are agent skills in plain terms?

Agent skills are reusable instruction packs that tell a model how to handle a specific repeating job. Claude uses SKILL.md files with trigger descriptions. ChatGPT uses Custom GPT instructions. Gemini uses Gem instructions. All three load standing rules before the user types the ticket of the day. None of them replace your responsibility to paste, test, and save in the vendor UI.

Does PromptMake install agent skills for me?

No. PromptMake at https://promptmake.net/skills generates English configuration text: SKILL.md-shaped drafts, GPT field blocks, or Gem instruction blocks. You create folders, open ChatGPT or Gemini builders, paste, preview, and save. The product does not push files to Anthropic, OpenAI, or Google accounts.

How is Claude SKILL.md different from a Custom GPT?

SKILL.md is a file-oriented skill format aimed at Claude Code and Anthropic's skills model, with a trigger description and a markdown body. A Custom GPT is a ChatGPT profile with Instructions, knowledge uploads, and conversation starters. The writing job overlaps. The paste target and runtime do not. One shared brief can feed both exports.

Can one agent skills brief work for Gemini Gems too?

Yes, if you keep a shared ROLE, TASK, FORMAT, and FENCES block and map it into Gem Instructions. Knowledge files replace long PDF paste. Preview in gemini.google.com after paste. Expect to shorten examples for Gems that serve fast daily chats.

What makes a good skill description trigger?

A good trigger names the deliverable, the input shape, and when to load the skill. Use concrete nouns (support draft, rubric, lesson plan) and verbs (rewrite, grade, summarize). Add a when-not-to-use line if the skill collides with general chat. Skip vague labels like "helpful assistant."

Is a skill the same as a system prompt?

Close cousin, different packaging. A system prompt is one text block for a session. A skill is a reusable file or builder config meant to load when a job matches. Skills add trigger discipline and team sharing patterns. You still test both like product code.

How do I start with agent skills for free?

Draft a one-sentence job and three output headers offline. Open https://promptmake.net/skills, paste the brief, generate one export, edit fences, then test five samples in the platform you actually use. Guest tiers cover light tries; registered accounts raise daily caps. Confirm current limits on the site before a big team workshop.

When should I skip agent skills and use a normal chat prompt?

Skip skills when the task changes every message, when you have no stable output shape, or when a single RTF prompt already passes evals. Use skills when teammates repeat the same job weekly and need one named entry point with tested fences.