What you will be able to do
- Turn a recurring task done by ad-hoc prompting into a repeatable workflow, and choose between a Project, a Skill, a connector and a code-executed step for each part of it
- Explain why calculations done in Claude's written reply vary between runs, and fix the problem by moving the arithmetic into code
- Put the build of a shared workflow in the right order: prove it on your own work, test it against known answers, then share it with a named owner
- Hand off a workflow so a colleague can own it, and pick the sharing mode that gives them control or keeps it with you
Key concept
Repeatable workflow — A recurring task whose instructions, reference material and procedure are stored once in a reusable container, such as a Project or a Skill, instead of being retyped into each new chat. The container is what makes the output match across runs, and across the different people who run it.
1.Why retyped prompts drift, and where the workflow should live instead
If a recurring task depends on a prompt retyped each time, the output varies from run to run. Take a coordinator who opens a fresh chat every Friday and writes the status-update prompt from memory. The wording changes slightly each week, so the sections, tone and length of the update change with it, and clients notice. Claude is not being unreliable here. It is being given a slightly different task every week.
The fix is to write the procedure down once, in a place Claude reads on every run. Claude offers several such places, and each holds a different part of a workflow. Projects hold standing context. Skills hold procedures. Connectors give access to data in other services. Code execution handles the steps that must come out exactly the same every time. Custom instructions are a fifth candidate, but they apply to every conversation, so they suit general preferences better than one recurring task.
| Mechanism | What it holds | When it is in play |
|---|---|---|
| Project | Static background knowledge and instructions | Always loaded when you start chats within the Project |
| Skill | Specialized procedures: how to complete a specific task or workflow | Activates dynamically when relevant, and works everywhere across Claude |
| MCP connector | Access to external services and data sources | When Claude needs a tool; a skill can teach Claude how to use it effectively |
| Custom instructions | Broad preferences | Apply to all your conversations |
The deciding question is whether the workflow is tied to one body of material or to one kind of task. A client-report workspace that always uses the same layout guide and approved examples fits a Project. A procedure such as 'format any meeting notes our way' that should apply wherever the task comes up fits a Skill. The two can be combined: a Project can supply a client's background while a Skill supplies the house format.
Checkpoint 1 of 10· Match them up
A team is turning its recurring work into repeatable workflows. Match each part of the work to the mechanism where it belongs.
Tap a term, then the definition that fits it.
Projects hold standing context tied to one body of material and load in every chat inside them. Skills hold procedures for one kind of task and activate wherever that task comes up. Connectors give access to outside data, code handles steps that must be exact, and custom instructions apply to every conversation, so they suit general preferences rather than one recurring task.
“Projects provide static background knowledge that's always loaded when you start chats within them. Skills provide specialized procedures that activate dynamically when needed”Source: support.claude.com
Sources1
2.Projects: fixed instructions and an approved example for every run
A Project is a self-contained workspace with its own chats and a knowledge base. It has two places to store a workflow. Project instructions hold the rules: the fixed outline, the tone, the length, the perspective to write from. Project knowledge holds the material: the uploaded documents, templates and files Claude uses as background in every chat within the Project.
For a recurring deliverable, the strongest setup pairs the two. The instructions state the required outline. The knowledge holds one previously approved deliverable, so Claude has a concrete example of what 'right' looks like. Every Friday chat then starts from the same rules and the same example, because they belong to the workspace rather than to whatever the coordinator typed that morning. Pasting last week's update into a new chat looks similar, but the example changes each week and any drift carries forward. Saving a prompt in a notes file still leaves a person rebuilding the context and correcting the output by hand.
Projects are also not the same as having Claude remember a preference from one earlier conversation. The layout lives in a configuration that anyone with access can open and read. How memory differs from configured context is covered under context and memory management. Writing durable instructions belongs to persistent-instruction design.
Checkpoint 2 of 10· Check yourself
Setting up a Project for the Friday status update, the coordinator has two things to store: the rules (fixed outline, tone, length) and one update the client already approved. Which setup gives every Friday chat the same starting point?
Project instructions hold the rules, and Project knowledge holds a fixed approved example. Both load into every chat started in the Project. Custom instructions apply to all of your conversations, not just this task. Swapping in each week's newest update lets drift carry forward, and one long pinned chat is not a fresh chat that starts from the Project's stored context.
“you can define project instructions for each project to further tailor Claude's responses. For example, instructing Claude to use a more formal tone”Source: support.claude.com
Checkpoint 3 of 10· Exam question
Every Friday, a project coordinator writes a client status update by typing a fresh prompt into a new Claude chat. The prompt wording changes each week, so the update's sections, tone and length keep changing, and clients have noticed. The coordinator wants the same structure every week with the least manual effort. What is the best change?
Correct answer: A — Set up a Project for client reports, store the layout, voice guidelines and one approved example as its instructions and files, then start each Friday chat in it
- A. Correct. A Project's instructions and files are applied to every chat started inside it, so the layout, voice rules and approved example are applied the same way each week without anyone rewriting a prompt.
- B. Incorrect. Last week's update is a single example, and the request wrapped around it is still written fresh each Friday, so the wording keeps varying. Content from the old update can also carry into the new draft, so the format is still not standardized.
- C. Incorrect. Memory holds general preferences that Claude may or may not apply. It is not a fixed, reviewable template that is applied the same way each time, so the coordinator cannot rely on it to keep the update's format identical.
- D. Incorrect. A saved prompt fixes the wording, but it must still be copied in each week and carries no approved example or persistent setup. Any drift in the output is then repaired by hand, which keeps the manual effort the coordinator wants to cut.
Sources2
3.Skills: packaging a procedure that runs wherever the task appears
A Skill is a folder of instructions, and optionally scripts and resources, that Claude loads when a task calls for it. Its typical uses are recurring procedures: applying brand guidelines, generating emails from company templates, structuring meeting notes in a house format, or creating tickets in tools such as Jira following team conventions. Skills require code execution to be enabled.
The best Skills are narrow. They solve one specific, repeatable task. Their instructions are clear, they include examples where those help, they state when they should be used, and they cover one workflow rather than everything. A Skill can be a few lines of Markdown, or a multi-file package with executable code attached.
Every Skill contains a skill.md file. It opens with a required name and description, and the description does the most work. Claude reads it first to decide whether the Skill is relevant, and loads the full body only if it is. This staged loading, called progressive disclosure, keeps unused Skills from crowding the context window. A vague description therefore produces a Skill that never fires, however good its body is.
## Metadata
name: Brand Guidelines
description: Apply Acme Corp brand guidelines to all presentations and documentsMaterial that only applies in particular cases can go in extra files inside the Skill folder, such as a REFERENCE.md. When skill.md refers to that file, Claude can judge whether it needs to open it on a given run.
Checkpoint 4 of 10· Check yourself
An analyst's meeting-notes Skill has a detailed, well-written body, but its description just reads "Notes helper". When they ask Claude to tidy up meeting notes, the Skill is not used. What should they change first?
Claude reads the description first to decide whether a Skill is relevant, and loads the body only if it is. A vague description stops the Skill from firing however good the body is, so the fix is the description, not more body content.
“This is critical—Claude uses this to determine when to invoke your skill”Source: support.claude.com
4.Connectors: bringing source data into the workflow
Many recurring tasks start with copying data out of another tool. A connector removes that manual step. Connectors let Claude reach your apps and services, retrieve data from them and take actions inside them. Examples include searching Google Drive, sending a Slack message or creating a Linear issue. Once an app is connected, Claude can bring it into a conversation on its own when the request calls for it.
Connectors and Skills fill different roles in the same workflow. The connector supplies access to a tool. The Skill supplies the procedure for using that tool well. A Skill that says 'create tickets with our naming convention and these required fields', running alongside a Jira connector, produces tickets that follow team conventions without anyone pasting data. Partner-built Skills in the Skills Directory are designed to work with their matching connectors for exactly this reason.
Connectors do weaken consistency across users in one respect. Claude inherits each person's permissions from the connected service. Two colleagues running the identical workflow can therefore get different results if one of them cannot open a folder or channel the other can. Before rolling a connector-based workflow out to a team, confirm that everyone can reach the same sources. Setting up and governing connectors is covered under connector and knowledge management.
Checkpoint 5 of 10· Check yourself
A support lead wants Claude to open Jira tickets that follow the team's naming convention and required fields, without anyone copying data out of other tools first. Which setup fits this workflow?
The connector gives Claude access to Jira so no one has to paste data in. The Skill stores the procedure for using it the team's way, so the convention isn't retyped each run. The other setups either retype the rules or keep a manual copy-and-paste step.
“MCP connections give Claude access to tools, while skills teach Claude how to use those tools effectively.”Source: support.claude.com
5.Code-executed steps: making calculations come out the same every run
Some steps in a workflow must produce exactly the same answer every time: totals, percentage changes, counts, reformatting of a data file. When Claude does these sums in its written reply, the model has to track and add the numbers itself while it writes. The reply is generated as text, one piece after another, and no program computes the figures, so on a 400-row file each run can come out with slightly different totals. Code behaves differently: it adds the same numbers the same way and returns the same result on every run. Asking Claude to show more working, or to run the task three times and take a majority, keeps the arithmetic in the same place the errors come from.
The fix is to route that step through code. With code execution and file creation enabled, Claude writes and runs code on uploaded files to perform advanced analyses. A quarterly expense workflow can then work like this: the totals and per-department changes are computed by code over the spreadsheet, and Claude's reply explains the figures the code produced. The calculation becomes a repeatable step. Claude's own work is the part that needs judgement: interpreting the numbers and writing them up.
Code can also be built into the workflow container itself. Custom Skills can include executable scripts. Anthropic's own document skills use Python libraries such as pandas and numpy, so a Skill can carry the same calculation logic to every run and every user.
Checkpoint 6 of 10· Check yourself
Five analysts on a finance team each run the same quarterly variance calculation on their own. Their figures should match from run to run and from analyst to analyst. Which design carries the calculation itself to every run and every analyst?
Custom Skills can include executable scripts, so the calculation runs as code and the same logic goes to every run and every user. The other designs leave the arithmetic in Claude's written reply, which is where the variation between runs comes from.
“no coding required for simple skills, though you can attach executable scripts to custom skills for more advanced functionality.”Source: support.claude.com
Checkpoint 7 of 10· Exam question
A finance assistant uses Claude each quarter to total expenses across 400 rows of an uploaded spreadsheet and compute the percentage change per department. Different runs of the same data have produced slightly different totals when Claude calculated in its written reply. The team wants the totals identical on every run. Which workflow design fixes this?
Correct answer: C — Have Claude run code on the uploaded file to compute the totals and percentages, then explain the results in its reply
- A. Incorrect. Writing out running sums still has the model producing the arithmetic as generated text, so a slip can occur in any step and differ between runs. Showing the work makes errors easier to spot but does not make the totals identical on each run.
- B. Incorrect. The spreadsheet formula adds the subtotals deterministically, but each subtotal is still calculated by Claude in prose. Those four figures can still change between runs, so the final total can too.
- C. Correct. Code executed on the uploaded file computes the totals and percentage changes the same way from the same data on each run. Claude's role is then limited to explaining the output, so the numbers no longer depend on text-based arithmetic.
- D. Incorrect. Each of the three runs is still arithmetic done in generated text. Two runs agreeing does not make them correct, and a rerun next quarter can settle on a different majority, so the totals are still not guaranteed to be identical.
6.Building order: prove the workflow on your own work before others rely on it
Sharing a workflow before it works spreads its errors to every user. The safe order has four stages, and each depends on the one before it.
Start by running the task yourself several times and recording which wording and which corrections produced output the audience accepted. That record becomes the raw material for the workflow. Next, turn it into a container. That means Project instructions holding the fixed outline, with an approved past deliverable in knowledge, or a Skill holding the procedure. Then test the container against cases where you already know the right answer, such as past months whose correct figures are on file, and fix any gaps it shows. Only then share it, naming the person who owns it and telling users how to report problems.
Testing a Skill follows a concrete loop. Enable it, try several different prompts that should trigger it, and check that Claude actually loads it. If Claude does not use the Skill when you expected it to, change the description, because that is the field Claude uses to decide. Also confirm that every file the Skill refers to exists where skill.md says it is.
Checkpoint 8 of 10· Put it in order
An analyst has finished building a meeting-notes Skill and wants to test it before sharing it with the team. Put the steps of the testing loop in order.
- 1.If Claude isn't using it when expected, revise the description and test again
- 2.Check that Claude actually loads the Skill for those prompts
- 3.Try several different prompts that should trigger it
- 4.Enable the Skill
The Skill has to be enabled before any prompt can trigger it. You can only tell whether it loads after trying prompts. If it doesn't load, the fix is the description, because that is the field Claude reads to decide whether to use the Skill.
“Iterate on the description if Claude isn't using it when expected.”Source: support.claude.com
Checkpoint 9 of 10· Exam question
A regional manager writes a monthly branch performance memo by hand-prompting Claude, and results vary. They are turning this into a repeatable workflow that four branch leads will also run, and the memo goes to executives. Which sequence best builds and rolls out the workflow?
Correct answer: C — Run the task and note what worked → Write instructions, outline and approved example → Test on past months and fix gaps → Share, name an owner, set reporting route
- A. Writing the instructions before running the task means they rest on guesses rather than on the wording and corrections that executives actually accepted. The later test would then be checking guesswork instead of a proven approach.
- B. Writing the instructions and testing them before running the task and recording what worked means the instructions and the test are built without evidence of the accepted result. The manager would be testing guesses and only afterwards learn what wording and corrections the executives accepted.
- C. Running the task first records what actually produced an accepted memo, that evidence is turned into instructions, an outline and an example, the Project is tested on past months with known figures, and only then is it shared with a named owner and a reporting route.
- D. Sharing with the branch leads before testing on past months with known figures means the leads rely on an unproven Project, and defects surface in front of them and the executives instead of being fixed in private. The owner is named too early, before the workflow has been shown to work.
Sources3
7.Handing a workflow to the colleague who will own it
A workflow that others rely on needs one owner who can change it. The deciding detail is whether the sharing mode actually gives that person edit rights. Projects and Skills handle this differently.
On Team and Enterprise plans, a Project can be shared at two permission levels. Can view members can read the contents, knowledge and instructions and chat in the Project, but cannot change it. Can edit members can change instructions and knowledge and manage members. To hand a Project to a new owner, give them Can edit. Without it they cannot fix the workflow when something changes. A common case is a required template that has been revised while the old template file still sits in the Project's knowledge. Because knowledge loads on every run, the stale file keeps shaping output until someone with edit rights replaces it.
| Aspect | Share | Publish to org |
|---|---|---|
| Who receives it | Specific people, or groups on Enterprise | The organization's library, for anyone to install |
| Who controls it | You remain the owner; recipients get it view-only | The organization manages it after publishing |
| How updates reach users | Recipients get your updated version at next use | Republish; users stay on the approved version until the update is approved |
Sharing a Skill does not hand it over. Recipients can use a shared Skill but cannot edit it, and it still lists you as owner. Use sharing when you will keep maintaining the Skill. Publishing moves the Skill into the organization's control, and in organizations that require review, each new version is checked before anyone receives it.
These same rules keep results consistent across users. A view-only Skill cannot be changed by the people using it. Updates reach recipients automatically, so nobody keeps an old copy. Organization Owners can also provision Skills so that they appear for every member, enabled by default. Who can configure provisioning and review settings is covered under project configuration. Keeping a workflow current over time is covered under configuration lifecycle.
Checkpoint 10 of 10· Match them up
Before handing a weekly-report workflow to the colleague who will own it, you compare the sharing options. Match each option to what it gives the people who receive it.
Tap a term, then the definition that fits it.
Sharing a Skill keeps control with you, because recipients get it view-only. Publishing hands the Skill to the organization. For a Project, only Can edit lets the new owner replace outdated knowledge or change instructions, so that is the permission a handoff needs.
“Sharing gives a skill to specific people or groups, and you keep control of it. Publishing hands it to your organization.”Source: support.claude.com
Sources
(8)
Every claim above is drawn from one of these pages, quoted as it was written on the date shown.
- 1.
“Custom instructions apply broadly to all your conversations.”
↩︎ Why retyped prompts drift, and where the workflow should live instead“Skills provide specialized procedures that activate dynamically when needed and work everywhere across Claude.”
↩︎ Why retyped prompts drift, and where the workflow should live instead“This feature requires code execution to be enabled.”
↩︎ Skills: packaging a procedure that runs wherever the task appears“MCP connections give Claude access to tools, while skills teach Claude how to use those tools effectively.”
↩︎ Connectors: bringing source data into the workflow“you can attach executable scripts to custom skills for more advanced functionality.”
↩︎ Code-executed steps: making calculations come out the same every run“Distribute approved workflows consistently across all employees”
↩︎ Handing a workflow to the colleague who will own it“Skills teach Claude how to complete specific tasks in a repeatable way”
↩︎ Key concept“Projects provide static background knowledge that's always loaded when you start chats within them.”
↩︎ Exam trap 1“Projects provide static background knowledge that's always loaded when you start chats within them. Skills provide specialized procedures that activate dynamically when needed”
↩︎ Checkpoint“Projects provide static background knowledge that's always loaded when you start chats within them.”
↩︎ Prediction“no coding required for simple skills, though you can attach executable scripts to custom skills for more advanced functionality.”
↩︎ Checkpoint - 2.
“you can define project instructions for each project to further tailor Claude's responses.”
↩︎ Projects: fixed instructions and an approved example for every run“You can upload relevant documents, text, code, or other files to a project's knowledge base”
↩︎ Projects: fixed instructions and an approved example for every run“Members can see project contents, knowledge, and instructions, and chat within the project, but cannot edit it.”
↩︎ Handing a workflow to the colleague who will own it“you can define project instructions for each project to further tailor Claude's responses. For example, instructing Claude to use a more formal tone”
↩︎ Checkpoint - 3.
“Are focused on one workflow rather than trying to do everything”
↩︎ Skills: packaging a procedure that runs wherever the task appears“Test with example prompts to ensure Claude invokes it appropriately.”
↩︎ Building order: prove the workflow on your own work before others rely on it“Verify all referenced files exist in the correct locations.”
↩︎ Building order: prove the workflow on your own work before others rely on it“Iterate on the description if Claude isn't using it when expected.”
↩︎ Exam trap 4“This is critical—Claude uses this to determine when to invoke your skill”
↩︎ Checkpoint“Iterate on the description if Claude isn't using it when expected.”
↩︎ Checkpoint - 4.https://support.claude.com/en/articles/11176164-use-connectors-to-extend-claude-s-capabilitiesOfficial docs
“Connectors let Claude access your apps and services, retrieve your data, and take actions within connected services.”
↩︎ Connectors: bringing source data into the workflow“Claude inherits each person's permissions from the connected service.”
↩︎ Connectors: bringing source data into the workflow - 5.
“perform advanced analyses on uploaded data”
↩︎ Code-executed steps: making calculations come out the same every run - 6.
“Rather than requiring the model to mentally track and sum dozens of expenses with complex metadata”
↩︎ Code-executed steps: making calculations come out the same every run“Claude delegated the arithmetic and aggregation logic to Python code. This reduced cognitive load on the model, ensured precise calculations”
↩︎ Exam trap 2“Claude delegated the arithmetic and aggregation logic to Python code. This reduced cognitive load on the model, ensured precise calculations”
↩︎ Prediction - 7.https://claude.com/blog/analysis-toolSecondary source
“mathematically precise and reproducible”
↩︎ Code-executed steps: making calculations come out the same every run - 8.
“Shared skills are view-only. Recipients can enable and use the skill, but they can't edit the contents.”
↩︎ Handing a workflow to the colleague who will own it“everyone who uses the skill stays on the approved version until the update is approved.”
↩︎ Handing a workflow to the colleague who will own it“Sharing gives a skill to specific people or groups, and you keep control of it. Publishing hands it to your organization.”
↩︎ Exam trap 3“Sharing gives a skill to specific people or groups, and you keep control of it. Publishing hands it to your organization.”
↩︎ Checkpoint