Claude Skills vs Projects vs Prompts: What to Use When
What Claude skills, projects, prompts, connectors, and Cowork each hold, when Claude loads them, and the rule for putting every bid job in the right place.
Written by
Ops Lead, AutoRFP.ai
- Published
- Reading time
- 9 min
Claude skills vs projects comes down to what each one holds. A project holds the standing context for one body of work and loads into every chat inside it, while a skill holds a repeatable method that Claude loads only when a task calls for it.
On a live bid, Claude projects vs skills is rarely a choice between the two: the go/no-go and compliance matrix skills run inside the tender’s project.
Below are the rule, a side-by-side table, and what to put in each feature; the wider workflow for using Claude across the whole bid lives in the series hub. Plan details are as of November 2026.
Claude Skills vs Projects: The Rule for Bid Teams
Match each bid job to the feature that holds what it needs:
- A project for each live tender, holding the RFP, the buyer’s answers, and your bid facts.
- A skill for each method you run on every bid, such as go/no-go, the compliance matrix, and the contradiction check.
- A plain prompt for one-offs, such as rewriting one answer for an evaluator who is not technical.
- A connector for systems Claude should read live, such as the approved answer library and the CRM.
- Cowork for multi-step work that ends in files, such as a matrix built from a tender folder.
- A plugin to give the whole team the same skills and connectors.
Anthropic’s Skills explained (March 5, 2026) sums up the first two: “Projects say ‘here’s what you need to know.’ Skills say ‘here’s how to do things.’”
| Feature | What it holds | When Claude loads it | What persists | Where it breaks on a bid |
|---|---|---|---|---|
| Prompt | One request | When you send it | Nothing beyond the chat | Copies in people’s notes drift apart |
| Project | Files, instructions, and chats for one body of work | Always, in every chat inside it | Its files and instructions | Goes stale when the buyer’s answers are not added |
| Skill | A SKILL.md file and reference files | Its description always; the rest when a task matches | Installed per user, or provisioned by owners | Edited copies drift from the agreed method |
| Connector (MCP) | Access to another system | When Claude calls one of its tools | The connection, per user | Reaches all the user can reach |
| Cowork | A multi-step task on connected folders | When you hand it an outcome | Cowork projects keep files and memory | Paid plans only; outputs need a reviewer |
| Plugin | Skills, connectors, and sub-agents bundled | Once installed, in chat, Cowork, and Claude Code | Saved to the account | Local MCP servers run with the user’s full program permissions (Anthropic) |
What Are Claude Skills?
A Claude skill is a folder of instructions and reference files that Claude loads when a task calls for it, so a method your team runs on every bid is written once and applied the same way.
Anthropic’s help center calls skills “folders of instructions, scripts, and resources that Claude loads dynamically” (What are skills?, November 2026). AutoRFP.ai’s AI Go/No-Go Agent Skill shows how they load in three layers.
Only the description, which names requests such as “should we bid on this” and “go no-go”, stays in Claude’s context. The SKILL.md workflow loads when a request matches it, and the scorecard in references/go-no-go-framework.md (seven weighted categories, each question scored 1 to 5, and the decision thresholds) loads when a step needs it.
As of November 2026, Anthropic says skills work on every plan with code execution turned on, in Claude Code, and inside projects, and Team and Enterprise owners can provision them for all users.
Skills also travel between tools under the open Agent Skills standard, and AutoRFP.ai’s skills library, free and unlocked with an email address, ships each package for AutoRFP.ai, Claude, ChatGPT, Gemini, Grok, and Copilot.
When Should a Claude Prompt Become a Skill?
Turn a prompt into a skill once you find yourself pasting it into every tender. Anthropic’s March 5, 2026 guide gives the same signal: “If you find yourself typing the same prompt repeatedly across multiple conversations, it’s time to create a Skill.”
Bid work adds three signals of its own:
- The output needs the same format every time, such as a matrix your reviewers know.
- A scorecard must be applied the same way on every bid, so scores compare across bids.
- More than one person runs it, and each copy would drift.
Keep it a prompt while it is specific to one buyer, still changes on every run, or is needed once, like prompt 37 in our 72 Claude prompts for bid teams, which rewrites a technical answer for a lay evaluator.
Prompt 31, “Shred the RFP into a requirements register”, and the Compliance Matrix Builder skill do the same job. The prompt suits you while you are still changing it; once the team runs it on every bid, the skill keeps one agreed version.
Prompt, pasted per tender
- Pasted into the chat for each tender
- Each copy can be edited and drift apart
- Suits a method you are still changing
Skill, installed once
- Installed for you, or provisioned for the team
- Loads when a request matches its description
- One agreed version for every bid
When a prompt has earned it, build your first Claude skill from a method your team already trusts.
What Is a Claude Project?
A Claude project is a workspace with its own files, instructions, and chat history that loads into every chat you start there, which makes it the right home for one live tender.
Anthropic’s help center says projects “provide static background knowledge that’s always loaded when you start chats within them” (What are projects?, November 2026).
Free users get up to five projects, and Team and Enterprise members can share one (as of November 2026). A project is only as current as its files; see where Claude slips on a long tender.
What to Put in a Project for Any Live Tender
A tender’s project holds what is true for that one bid: the RFP and every appendix, every Q&A, clarification, or addendum, a facts file with the offer, the team, and the evidence you plan to cite, your win themes and clarification log, and project instructions that set the rules for every chat.
Save any spreadsheet embedded in the Word file as its own file, and add each Q&A or addendum the day it is published, because a mandatory requirement can appear only in the Q&A.
Goes in
One project per tender
Comes out
Leave out what belongs elsewhere: the scorecard and extraction method (skills), the whole answer library (a connector), and one-off requests (prompts).
For ready-made rules, use AutoRFP.ai’s Claude Cowork project instructions for RFP teams. They require a source on every answer and a GAP flag where evidence is missing, and name a skill for each stage, from ai-go-no-go to rfp-contradiction-checker.
The instructions keep past responses in the project folder. Where your approved answers live in AutoRFP.ai, its connector replaces that copy, and Claude reads each approved answer and its source from the live library.
On a Public Tender, the Misses Sat in the Embedded Workbook and the Q&A
We ran the Compliance Matrix Builder skill and prompt 31 on State of Maine RFP# 202506081, a state agency case management software RFP issued July 15, 2025, for a fictional bidder. Each ran once in a Claude Code session (Claude Opus 5.5, November 6 and 9, 2026), scored against 457 hand-listed requirements: 67 exist only in a workbook embedded in the Word file, and 44 only in the Q&A.
| Run (one run each) | What Claude was given | Requirements found |
|---|---|---|
| Skill, version 1.1 | The RFP and Q&A as text, without the embedded workbook | 391 of 457 (85.6%), and 390 of the 390 in its input; it named the workbook as not read and asked for it |
| Skill, version 1.1 | The same, plus the workbook as its own file | 457 of 457 (100%) |
| Prompt 31, no skill | The same full pack | 420 of 457 (91.9%); 7 of the 44 Q&A-only obligations became requirement rows |
The skill runs also returned duplicate rows (60 in the full-pack run) for a reviewer to merge.
Two lessons carry to any tender project. Save embedded spreadsheets and every Q&A as their own files, because Claude finds only what it is given. Once a method matters, put it in a skill: version 1.1 turns Q&A answers into requirement rows, while prompt 31 kept them in a separate table.
Claude Skills vs MCP Connectors
A connector gives Claude reach into another system and a skill gives it the method, so a bid workflow that needs approved answers uses both. In Anthropic’s words: “MCP connects Claude to data; Skills teach Claude what to do with that data.”
Each member signs in to a connector, so Claude reaches whatever that person can reach (Anthropic help center, as of November 2026). For what that covers, see how Claude connectors and MCP work for a bid team.
Two skills in AutoRFP.ai’s library show the pairing:
- Internal Research uses a CRM connector, or a CRM export when none exists.
- Assign Owners reads the AutoRFP.ai content library through MCP tools such as
search_content, then proposes owners and reviewers as a plan that changes nothing until the user approves it.
The video below shows Claude skills for bid writing and RFP responses that work from connected, approved company knowledge instead of leaving the model to fill the gaps.
AutoRFP.ai is AI proposal software that writes bid answers from your approved library, and through its MCP connector a Claude skill drafts from answers your team has already approved, each carrying its source, where an uploaded export would drift out of date.
The AutoRFP.ai connection offers read scopes for lookups and full scopes that write to projects and content, always as the signed-in user, with approvals kept in AutoRFP.ai’s workflow and anything the content cannot support routed to a person. Anthropic’s Connectors Directory lists it as “Anthropic verified” (added July 2026).
To give the whole team the same skills and connector, bundle them as a plugin, which owners can provision on Team and Enterprise plans (Use plugins in Claude, November 2026). If you are still deciding whether Claude alone is enough for RFP work, that comparison sets out the trade-offs.
When a Bid Job Needs Cowork
A bid job needs Cowork when the result is a finished file built from several documents. In Cowork you “describe an outcome, step away, and come back to finished work”, on files in folders you connect (Get started with Claude Cowork, November 2026).
| Use | When the job is | Bid example |
|---|---|---|
| Chat | A conversation or a single answer | Rewriting one answer for a new evaluator |
| Cowork | A multi-step job that ends in files | A matrix from a tender folder, or a completed security questionnaire |
Cowork is on paid plans only (as of November 2026).
Claude Code also runs multi-step work on files; see when Claude Code suits a bid manager better than Cowork.
Where to Start With Your Next Tender
Give your next tender one project, install the skills you run on every bid, connect the approved library, and keep one-offs as prompts.
Project. One project for the tender, with the RFP, appendices and every Q&A or addendum added as it lands.
Instructions. Project instructions that require a source for every answer and a GAP flag where evidence is missing.
Skills. The skills you run on every bid installed before you start, such as go/no-go and the compliance matrix.
Library. The approved answer library reached through a connector, with no uploaded copy to keep in sync.
Prompts. One-off requests kept as prompts, and any prompt you paste into every tender turned into a skill.
Cowork. Used when the job is a finished file built from several documents.
Begin from the free Claude skills library for bid teams, unlocked with an email address. Then watch for the prompt you keep pasting into every tender: that one is your next skill.
About the author
Ops Lead at AutoRFP.ai. Designs the pipeline, content, and handoff systems bid teams use to run RFPs end-to-end, and co-hosts live sessions on winning workflows.
LinkedIn
