Claude Projects vs. skills vs. raw prompts — when to use which

Three Claude surfaces, three ways to do roughly-similar-looking things, and almost every solo consultant uses one of them for the wrong kind of work. The result is a workflow that feels like it should be smoother than it is. Pasting context every conversation. Reinventing structure on every prompt run. Living in a long-running Claude Project for tasks that should have been one-shot prompts.

The three surfaces:

• Raw prompts — a Markdown file you paste into a new conversation.
• Claude skills — auto-activated instructions loaded from ~/.claude/skills/.
• Claude Projects — long-running conversations with persistent context and uploaded files.

Each is genuinely useful. Each is wrong for some workflows. This article is the consultant-specific rule for picking between them.

What each one actually is

Before the picks, a clean definition of each — because the marketing language conflates them and most consultants form their mental model from the marketing.

Raw prompts are the simplest thing. You write a prompt once (the structure, the voice rules, the output format), save it as a .md file, and paste it into Claude when you need it. The prompt has no memory between runs. You feed it inputs each time. Output goes wherever you copy it. Nothing is persistent.

Claude skills are prompts that auto-load. The same prompt content you'd otherwise paste, but Claude reads it from ~/.claude/skills/ and activates it when trigger phrases in your conversation match. You skip the paste step. The skill's instructions stay the same across every activation; the only thing that varies is the input you give it.

Claude Projects are persistent conversations with shared context. You create a Project, upload files (PDFs, docs, transcripts, anything Claude needs to know about), set instructions for the Project itself, and then conversations inside the Project carry that context forward. Files persist between conversations. The Project remembers things from earlier sessions.

Three different things. The names sound parallel; the actual capabilities are not.

The picking rule

The cleanest way to decide which to use is to ask one question first: does the workflow have persistent state, or is each run independent?

If each run is independent — paste input, get output, move on — you want a skill or a raw prompt. The workflow has no memory and doesn't need one. Discovery debriefs are independent. Weekly client updates are independent. Each one stands alone; you give Claude the inputs from this week and the output is for this week.

If the workflow has persistent state — the same client, the same engagement, with context that builds up over weeks or months — you want a Project. Project context carries forward. You upload the kickoff brief on day one, the meeting transcripts as they happen, the deliverable drafts as you produce them. By week six, the Project knows the engagement; you can ask "what did the CEO say in the strategy session three weeks ago about the conversion target?" and Claude answers from the uploaded transcript.

That single question — independent or persistent — does most of the picking.

The follow-up: if independent, raw prompt or skill? Skill if the workflow runs more than three times. Raw prompt if it runs once or twice and you'd rather not clutter ~/.claude/skills/.

That's the rule, in 80 words. Most consultants get this wrong by reaching for a Project for everything, treating it as the default Claude surface, and ending up with twelve Projects (one per client, one per workflow) that each have unfocused contents. The Project becomes a junk drawer instead of a context window.

What each surface is good at (with consulting examples)

Raw prompts win when:

• The workflow runs once a quarter or less (the quarterly retrospective, the annual planning exercise).
• You're prototyping a new prompt and don't want to commit it to a skill yet.
• You're sharing the prompt with a peer who'll paste it into their Claude.

A raw prompt for the case-study harvest workflow, for example, runs once per engagement wrap — maybe 4-6 times a year for an active solo consultant. The friction of pasting it in is lower than the friction of maintaining it as a skill. Keep it in a prompts/ folder, paste when needed.

Skills win when:

• The workflow runs weekly or more often.
• The output structure needs to stay consistent across runs (every weekly update has the same five sections).
• You want Claude to auto-activate the prompt when you describe the situation.

The discovery debrief, the weekly client update, the meeting debrief — all skill territory. You run them constantly; auto-activation saves the paste step; structural consistency matters because the artifacts go to clients.

Projects win when:

• The work involves a single engagement with state that builds over time.
• You need to upload files Claude should reference repeatedly (the kickoff brief, transcripts, deliverable drafts).
• The conversation will return to the same context across days or weeks.

The right Project for a solo consultant is one per active engagement. Inside the Project: the kickoff brief, the weekly transcripts, the deliverable drafts, the running notes. Conversations inside the Project ask Claude about the engagement specifically. "Draft the case-study outline from what we've shipped so far." That's a prompt that needs the engagement context to produce useful output, and the Project is the right shape.

The wrong Project setup is one per workflow. "A Project for proposals." "A Project for discovery debriefs." Wrong shape — those are skill territory because each proposal is independent of the next.

If you'd rather load a finished version of this setup than assemble one, the Claude Project Kit ships the parts that don't change per engagement — a tuned instructions block plus the four practice files (ICP, offers and rate floor, voice, a per-client context template) that make any Project you create start grounded. (Since this piece ran, a fourth surface joined the three: pointing Claude at a folder — it reads your files and writes deliverables back as files. The Claude Workspace Kit is that setup assembled, and it shares its context files with the Project Kit on purpose.)

The combination most consultants miss

The interesting move — and the one most solo consultants haven't installed yet — is a skill that runs inside a Project.

The way this works: you have a Project for the Acme engagement (one Project, six weeks of context built up). You've installed the weekly-client-update skill globally in ~/.claude/skills/. In a new conversation inside the Acme Project, you type "Friday weekly update." The skill auto-activates AND the Project context loads. The skill provides the structure (five sections, output format, voice rules); the Project provides the substance (this week's actual work, the running history, the open items from last week).

Output: a draft of this client's update, in the consistent five-section format, based on the engagement-specific context. You edit one paragraph and ship.

That's the combination most consultants haven't connected. The skill is the structure; the Project is the substance. They're not competitors — they're a chain.

The Proposal-Closer Prompt Pack is built around this idea at the multi-prompt scale: one skill (proposal-closer) routes between eight prompts depending on which stage you're in. Drop the skill into ~/.claude/skills/, work inside a Project for the active deal, and the prompts route automatically while pulling on the deal's accumulated context. Skill provides the structure; Project provides the deal-specific substance.

The three failure modes

One Project for everything. A single mega-Project with all your clients' files in it. Claude can't tell which engagement you're referring to in a conversation; outputs are generic; uploads stack up until the Project becomes unusable. Fix: one Project per engagement. Delete the mega-Project.

Skills for one-off tasks. Building a skill for the annual planning exercise you run once. The cost of building, tuning, and maintaining the skill exceeds the cost of pasting a raw prompt twice a year. Fix: raw prompt for low-frequency work; skill for high-frequency.

Raw prompts for high-frequency work. Re-pasting the weekly update prompt every Friday for 18 months, never building it into a skill. You're paying the paste tax 78 times a year. Fix: when a prompt runs more than three times, promote it to a skill.

Most consultants have all three failure modes simultaneously when they first start using Claude seriously. The pattern resolves once you map workflow frequency × workflow persistence to the right surface. Independent + frequent = skill. Independent + rare = raw prompt. Persistent = Project. Combinations of those (skill inside a Project) when both apply.

The migration path (what to do this week)

Most consultants reading this are starting from one of two states: "everything is raw prompts" or "one mega-Project for everything." Both have a clean migration path that takes about 90 minutes.

If you're starting from raw prompts for everything:

1. List the prompts you've run more than three times in the last 90 days. There are probably four to six.
2. For each one, ask the picking question: independent or persistent? Most will be independent (discovery debrief, weekly update, meeting debrief, status email).
3. Convert the high-frequency independent ones to skills. Drop them into ~/.claude/skills/<name>/SKILL.md. Add trigger phrases.
4. Keep the low-frequency ones as raw prompts in a prompts/ folder. The annual planning prompt, the quarterly retrospective prompt — these stay as files you paste in.
5. Create one Project for each active engagement. Upload the kickoff brief and any existing transcripts.

Total time: 60-90 minutes. Payback: the next time you wrap a discovery call, the skill auto-activates and you skip the paste step. The next time you write a weekly update, the skill loads inside the engagement's Project and produces the draft with the engagement-specific context.

If you're starting from one mega-Project:

1. Create a new Project for each currently-active engagement. Three engagements = three Projects.
2. Move the relevant files into the relevant Project. Acme Widgets' kickoff brief goes in the Acme Project, not the mega-Project.
3. Archive (don't delete) the mega-Project. You may need to reference its history; don't lose it.
4. Repeat step 5 from the raw-prompts path above for any high-frequency workflows you didn't already skill-ify.

Total time: 30-60 minutes depending on how many files moved. Payback: when you next ask the Acme Project a question, Claude has only Acme's context to reason from, not the muddled mass of three clients' files. Outputs sharpen immediately.

The point of either migration: the surface should match the workflow, not the other way around. Surfaces are infrastructure. Workflows are what pays your rent — the AI-native habit of keeping the tool subordinate to the work.

Where to go deeper

Appendix B of The Solo Consultant's AI Playbook covers the skill format and the Claude/ChatGPT tooling differences in more depth. The Appendix walks through the Claude skill SKILL.md file structure, when ChatGPT Custom GPTs are the closer analog, and the tooling decisions to make once you've picked your surface for a given workflow.

If you've already built one Claude skill and are ready to build the next two, the prior article in this series — Claude skills for solo consultants — what they are, and the first three to build — names the three highest-frequency workflows and the skill structures for each.

The fastest move you can make after reading this: pick the workflow you're running most often this month, decide which surface it belongs in, and move it. The cost is 30 minutes. The compound return is every time you run that workflow for the rest of the year.