Claude can feel inconsistent when every job begins with a blank chat. You paste the background again, explain the audience again, restate your standards again, then spend time correcting familiar mistakes. The issue is rarely that you need a longer one-off prompt. You need a place for each kind of context.
This is how to set up Claude for a repeatable work workflow: separate what is always true about how you work, what is true for one body of work, and what must happen every time you create a specific deliverable.
That distinction matters more now because Claude can support longer-running work with files, connected tools, and recurring tasks alongside ordinary conversations. The useful response is not to automate everything. It is to build a reliable manual workflow first, with clear human approval points.
If you need help choosing the first deliverable to systematize, use the AI Project Scope Generator to turn an idea into a practical implementation brief. It helps you define the outcome, scope, inputs, constraints, and review points before you build context around a vague task.
The three layers: preferences, context, and procedure
Claude gives you three distinct places to put information. Treating them as interchangeable is the reason setups become cluttered or unreliable.
- Instructions hold account-wide preferences. Use them for enduring choices about how Claude should communicate and collaborate with you.
- Projects hold focused context for a line of work. Use them for the facts, files, examples, and constraints that apply to a client, publication, team, or initiative.
- Skills hold a repeatable procedure. Use them when a deliverable has stable steps, required inputs, a defined format, and a quality bar.
A simple rule helps: Instructions are how you generally work. A Project is what this work is about. A Skill is how a recurring deliverable gets made.
Do not put a client’s positioning into account Instructions. Do not turn a folder of background documents into a procedural Skill. And do not expect context from one Project conversation to automatically become available in every other conversation. If information must be reused, add it to the Project knowledge rather than leaving it buried in a previous chat.
Choose one workflow worth systematizing
Start with one deliverable, not your entire job. A good candidate meets three conditions:
- You create it at least twice a month.
- It needs roughly the same inputs, standards, or sections each time.
- Someone else will read it, use it to make a decision, or act on it.
Examples include a client-meeting brief, a proposal outline, a source-based newsletter package, a weekly operating report, or a content repurposing pack. Avoid highly novel work for your first build. You are looking for repetition with judgment, not a task that can be left unsupervised.
For a wider process covering task selection, prompting, and review, read How to Use AI. The workflow below is the practical setup layer for a task you have already chosen.
Step 1: Write minimal account Instructions
Your account Instructions should be short because they follow you everywhere. They should improve most conversations without forcing a client-specific rule into unrelated work.
A professional creator might use Instructions like these:
- Use UK English unless I ask otherwise.
- Write directly and avoid inflated claims.
- When facts, figures, or dates are missing, flag the gap instead of inventing an answer.
- For recommendations, state assumptions and give the next practical action.
- When reviewing a draft, identify the highest-impact changes first.
Notice what is absent: brand voice documents, customer details, campaign goals, confidential files, and one-off formatting requests. Those belong elsewhere.
Review your Instructions after a few weeks. If a rule only helps one project, remove it from the global layer. Global preferences should stay boring, stable, and broadly useful.
How to set up Claude for a repeatable work workflow with a Project
Create one Project for the ongoing area of work, then make it the source of truth. A Project is where Claude should find the information it needs to understand the work without being reminded in every chat.
Add only material that is current, approved, and genuinely useful for repeated jobs. For a client-briefing Project, that could include:
- Client positioning and audience notes.
- Current priorities, agreed terminology, and known constraints.
- Approved past briefs that demonstrate the desired standard.
- Meeting templates and reporting formats.
- Reference documents that the brief must reflect.
Then add Project instructions. These differ from account Instructions because they define the local operating environment. For example:
Write for the client’s senior commercial team. Distinguish supplied facts from analysis and recommendations. Do not make claims that are not supported by Project materials or the current task inputs. Use the agreed briefing structure. Flag conflicts or missing information before drafting.
Keep the Project clean. Remove outdated files, replace superseded templates, and do not add documents merely because they might be useful one day. A smaller, maintained knowledge base is easier to trust than an archive of contradictory material.
Project and Project-instruction availability can vary by plan, account, and rollout. Check your Claude settings before designing a workflow that depends on a particular feature.
Step 3: Build a Skill for the recurring deliverable
Use a custom Skill when you can describe the job as a dependable method. Skills can package instructions, resources, and supporting files for specialised work, and a simple custom Skill can begin as a Markdown document. You do not need to code to define a useful first version.
A strong Skill answers five questions:
- Purpose: What is this deliverable for, and who will use it?
- Required inputs: What must be supplied before work starts?
- Method: What steps should Claude follow, in order?
- Output format: What sections, length, and formatting are required?
- Quality control: What must be checked before the result is presented?
A starter Skill template
Purpose: Produce a decision-ready client research brief for a meeting.
Required inputs: Meeting objective, attendee list, recent account updates, relevant source material, and decisions needed.
Method: First list missing inputs. Next extract confirmed facts from supplied material. Then identify implications, risks, and open questions. Draft only after the input check is complete.
Output: Use these sections: meeting objective, account snapshot, confirmed developments, opportunities and risks, recommended talking points, questions to ask, and next actions.
Quality check: Separate facts from recommendations, flag unsupported claims, verify names and numbers against supplied material, and ensure every recommended action has an owner or decision point.
This is deliberately specific without pretending the output is final. The Skill tells Claude how to perform a repeatable first pass; it does not transfer accountability for the advice.
Worked example: the client-ready research brief
Imagine your recurring request used to be: “Make me a client-ready research brief for tomorrow.” That request contains a hidden system. Pull it apart rather than rewriting it more eloquently.
- Account Instructions: direct writing, your spelling preference, a requirement to flag uncertainty, and your preferred feedback style.
- Project knowledge: client strategy, audience, prior meeting notes that need to remain available, approved examples, and account documents.
- Project instructions: who the brief serves, what evidence is acceptable, which commercial constraints apply, and how uncertainty should be handled.
- Skill: the exact briefing sequence, mandatory sections, missing-input check, and final audit.
- Current chat input: tomorrow’s objective, new developments, attendees, source materials, and the decision needed from the meeting.
The current chat still matters. No system can know what changed since the last run unless you provide the update. Your first instruction for each live job should be: “List missing inputs and assumptions before drafting.” That one step prevents polished output built on an incomplete brief.
Run, review, and improve the system
Run the workflow manually several times before considering scheduled work. Review each result against the Skill’s checklist, then capture only repeated corrections.
Use this routing rule:
- If the correction changes how you work across nearly everything, update account Instructions.
- If it corrects a fact, example, standard, or constraint for this area of work, update the Project.
- If it changes the steps or quality checks for this deliverable, update the Skill.
Keep human approval for claims, numbers, source quality, confidential material, external-facing recommendations, and final judgment. Claude can accelerate preparation and structure, but it cannot know whether a source is authoritative, a number is current, or a recommendation is appropriate without your review.
Once several manual runs produce consistently useful drafts, you can consider a scheduled task for a weekly brief or report. Schedule the stable workflow, not the experimental one. Define the inputs it can access, the output owner, the delivery destination, and what should happen if required information is missing.
If you want to formalise the underlying procedure beyond Claude, the Template and SOP guide can help you create a maintainable source of truth for repeatable work.
Build the system, not the mega-prompt
Better results do not come from cramming every instruction, document, and exception into one message. They come from putting stable information in the right layer and improving that layer when real work exposes a flaw.
Start with one recurring deliverable this week. Write five account-level preferences, create one focused Project, and draft one Skill around the output someone needs to use. After three real runs, revise it from the mistakes you actually see. That is how Claude becomes a durable working system rather than another blank chat.