Back
Claude loops for Monitoring and Evaluation

Claude loops for Monitoring and Evaluation

Beginner-friendly AI tutorial for M&E professionals

Build your first Claude loop—even if you have never created a project folder, Markdown file or automated AI workflow.

Claude loops for M&E can reduce repetitive review work when they are built with clear evidence rules and human oversight. This tutorial explains every step slowly: where to create the folder, which application to open, how to name and save each file, how to open the workspace in Claude, what to paste, what to approve and how to check that the result is safe and useful.

Check whether a task is ready for a Claude loop: Claude Loop Builder for M&E (Premium Tool)

Use practice data only. Do not use real beneficiary names, contact details, health information, complaint records, safeguarding information, passwords or confidential donor documents while learning this workflow.

Important: TASK.md, PROGRESS.md and LOOP_INSTRUCTIONS.md are ordinary text files that you create. They are not special built-in Claude commands.

What “.md” means: It is a Markdown file—a plain-text file that can contain headings, lists and simple formatting. You can open it with Notepad, TextEdit, Visual Studio Code or another text editor.

What You Will Build

You will build a small Weekly Monitoring Review Loop. It will read three fictional monitoring notes, prepare a structured weekly review, record what happened and flag anything that needs a person to check.

1. ReadClaude reads the instructions and sample monitoring notes.
2. DraftClaude writes a review inside the approved output folder.
3. RememberClaude updates a progress file for the next run.
4. VerifyClaude checks required sections and permission boundaries.
5. EscalateClaude stops when evidence is missing or judgement is required.

Estimated time: 60–90 minutes for a complete beginner. You can stop after the first successful manual run and return to the optional scheduling section later.

Beginner Terms Explained

The following words appear throughout the tutorial. You do not need technical experience; use these simple meanings:

FolderA container on your computer that holds files. It is also called a directory.
FileA saved document, such as TASK.md.
File extensionThe letters after the final dot. In TASK.md, the extension is .md.
WorkspaceThe main folder Claude is allowed to inspect for this exercise.
Source materialThe original notes Claude reads. In this tutorial, source files must remain unchanged.
OutputA new draft or report created by the workflow.
TerminalA text-based application used to type computer commands. Windows calls it Terminal, PowerShell or Command Prompt. macOS calls it Terminal.
Persistent stateInformation saved in a file so the next run can remember what happened previously.

Choose How You Will Use Claude

There are several ways to use Claude Code. For this tutorial, choose one of the first two options.

Easiest for beginners

Option A: Claude Desktop Code tab

You select the folder by clicking. Claude shows proposed file changes and lets you accept or reject them.

Choose this option when you do not want to use terminal commands.

Terminal option

Option B: Claude Code CLI

You open a terminal inside the folder, type claude and work inside a text interface.

Choose this option when Claude Code is already installed or you want to learn the terminal workflow.

Not recommended for the first exercise

Option C: Web or remote workspace

Remote sessions and cloud routines use a different environment. A normal chat where you upload files manually is not the same as a persistent local-folder workflow.

Return to this option later after the local version works.

Local folder or cloud-synced folder?

  • Local folder: stored directly on the computer, such as Documents or Desktop. This is the simplest choice.
  • Cloud-synced folder: OneDrive, Google Drive or Dropbox can work when the folder is also downloaded and visible in File Explorer or Finder.
  • Online-only file: a document that exists only in a browser cannot be selected by a local Claude session until it is downloaded or synced.

Recommendation: For the first exercise, create the folder in Documents. It is easy to find and less likely to become cluttered than the Desktop.

Choose a Suitable M&E Task

Do not begin by automating the biggest or most sensitive task in your organisation. A good first loop is small, repetitive, easy to check and safe when the output is wrong.

Use the following five questions. Each question includes a practical example.

1. Does the task repeat?

What this means: The same type of work happens again and again. You currently repeat similar instructions each week, month, meeting or reporting cycle.

Good example: Every Friday, you read three field updates and prepare a one-page monitoring summary.

Poor example: You need to design a completely new evaluation strategy once. That is a complex one-off task and may be better handled through a normal supervised conversation.

Simple test: Have you performed a very similar task at least three times before? If yes, it may be suitable.

2. Does the next run need to remember the previous run?

What this means: The next review is better when Claude knows what was already checked, what was missing and what remains unresolved.

Good example: Last week, one partner had not provided an attendance sheet. This week, the loop should check whether the missing document has arrived instead of treating the issue as new.

Poor example: You only need to convert one paragraph into a table. The result does not depend on an earlier run.

Simple test: Would a short note saying “what happened last time” improve the next output? If yes, persistent state is useful.

3. Can you check whether the output is acceptable?

What this means: You can write a checklist that distinguishes an acceptable draft from an incomplete one.

Good example: The weekly review must contain the reporting period, files reviewed, important changes, evidence gaps, risks and next actions. Every important observation must name its source file.

Poor example: “Improve the report until it feels excellent.” The word “excellent” has no clear test.

Simple test: Can another colleague use a short checklist and reach roughly the same pass/fail decision? If yes, the task is more suitable.

4. Is there a clear stopping point?

What this means: The workflow knows when to finish, when to try one correction and when to return the issue to a person.

Good example: Stop when the report contains all required sections and the state file is updated. Also stop when source files contradict each other or the evidence is insufficient.

Poor example: “Keep analysing until there is nothing else to say.” This can create endless output and unnecessary cost.

Simple test: Can you finish the sentence “The workflow is complete when…” with an observable condition? If yes, you have a stopping point.

5. Can the first version use limited permissions?

What this means: The loop can provide value without changing raw data, sending messages, publishing reports or updating an external system.

Good example: Read monitoring notes and write a draft to an outputs folder. A human opens the draft and decides what to do next.

Poor example: Automatically correct the source spreadsheet, send the report to the donor and mark the activity complete.

Simple test: Would the workflow still be useful if it could only read source files and write one draft? If yes, start there.

Decision rule: A task is a strong first-loop candidate when you can answer “yes” to all five questions.

If the output cannot be checked or the task requires broad permissions, redesign it as a smaller draft-and-review workflow.

Examples for M&E professionals

Original ideaSafer first-loop versionWhy it is safer
Correct all survey dataList possible duplicates, missing values and unusual recordsA person decides whether any correction is valid
Approve partner reportsCheck whether required sections and attachments are presentThe loop checks completeness, not approval
Decide whether the programme succeededSummarise available evidence, gaps and contradictory findingsEvaluation judgement remains human-led
Send the donor report automaticallyPrepare a draft and a pre-submission checklistAn authorised person reviews and submits

Task selected for this tutorial: Review fictional weekly monitoring notes and prepare a draft that highlights progress, evidence gaps, inconsistencies and questions requiring human review.

Step 1: Create the Folder and Workspace

A workspace is the main folder that contains everything for the exercise. Claude will only need access to this folder.

Recommended location

Create the folder inside Documents. The final location will look similar to:

Windows: C:\Users\YourName\Documents\me-monitoring-loop
macOS:   /Users/YourName/Documents/me-monitoring-loop

Windows: create the folders with File Explorer

  1. Open File Explorer. You can click the yellow folder icon on the taskbar.
  2. Click Documents in the left-hand menu.
  3. Right-click an empty area and choose New → Folder.
  4. Type me-monitoring-loop and press Enter.
  5. Open the new folder by double-clicking it.
  6. Create a folder named source-material.
  7. Create another folder named outputs.

Windows file-extension warning: Windows may hide extensions. In File Explorer, open View → Show → File name extensions. This helps you avoid accidentally creating TASK.md.txt.

macOS: create the folders with Finder

  1. Open Finder.
  2. Click Documents in the left-hand menu.
  3. Choose File → New Folder, or press Shift + Command + N.
  4. Name the folder me-monitoring-loop.
  5. Open the folder.
  6. Create a folder named source-material.
  7. Create another folder named outputs.

Recommended way to create the files

Use a plain-text editor. Visual Studio Code is convenient because it shows the folders and files in one window, but Notepad or TextEdit also works.

Using Visual Studio Code

  1. Open Visual Studio Code.
  2. Choose File → Open Folder.
  3. Select me-monitoring-loop.
  4. Use the New File icon to create each file.
  5. Press Ctrl + S on Windows or Command + S on macOS to save.

Using Notepad or TextEdit

  1. Open a new plain-text document.
  2. Paste the content from the tutorial.
  3. Choose Save As.
  4. Type the exact filename, including .md.
  5. Save it inside the correct folder.

When all files have been created, the structure should look exactly like this:

me-monitoring-loop/
├── TASK.md
├── PROGRESS.md
├── LOOP_INSTRUCTIONS.md
├── source-material/
│   ├── field-update-01.md
│   ├── field-update-02.md
│   └── partner-note.md
└── outputs/
    └── weekly-monitoring-review.md

Do not worry if the files are empty at this point. The next steps show exactly what to place inside each one.

Step 2: Create TASK.md

The task file describes the stable goal. Think of it as the short job description for the workflow.

Create and save the file

  1. Inside the main me-monitoring-loop folder, create a new file.
  2. Name it exactly TASK.md. Use capital letters for TASK.
  3. Open the file.
  4. Copy the complete block below and paste it into the file.
  5. Save the file.
# Weekly Monitoring Review Loop
## Goal
Review the files in `source-material/` and prepare a concise weekly monitoring review for human validation.
## Expected Output
Each run should create or update:
- `outputs/weekly-monitoring-review.md`
- `PROGRESS.md`
## Scope
Claude may read files inside this workspace.
Claude may write only to:
- `outputs/weekly-monitoring-review.md`
- `PROGRESS.md`
Claude must not:
- Modify source monitoring files
- Invent findings, quotations, dates or indicator values
- Delete, rename or move files
- Send, publish or approve the report
- Make final programme or evaluation decisions

What each section means

  • Goal: tells Claude the purpose of the workflow.
  • Expected Output: names the two files that should change after a successful run.
  • Scope: defines the boundaries. Claude can read the workspace but may write only to two approved files.
  • Claude must not: lists actions that remain forbidden even if Claude believes they would be helpful.

Beginner check: Close the file and open it again. If the text is still present and the filename is TASK.md, it has been saved correctly.

Step 3: Create PROGRESS.md

PROGRESS.md is the memory file. Claude reads it before each run and updates it before stopping. This is how one run can continue from the previous run.

Create and save the file

  1. Create a new file in the main workspace folder.
  2. Name it exactly PROGRESS.md.
  3. Paste the template below.
  4. Save the file. Leave blank fields blank; Claude will fill them during the first run.
# Loop Progress
## Current State
- Status: Manual testing
- Reporting period:
- Current focus: Validate the workflow before scheduling
- Last updated:
## Last Run
- Date:
- Trigger:
- Files reviewed:
- Summary:
- Output produced:
- Verification result:
## Open Issues
-
## Data-Quality Concerns
-
## Blockers
-
## Needs Human Review
-
## Next Run Should
- Read `TASK.md`, `PROGRESS.md` and `LOOP_INSTRUCTIONS.md`.
- Review files in `source-material/`.
- Update `outputs/weekly-monitoring-review.md`.
- Update this file before stopping.
## Decisions Made
- Source files remain read-only.
- The first version is manually triggered.
- No external tools or connectors are enabled.
## Do Not Repeat
- Do not infer missing results.
- Do not modify source files.
- Do not create additional output files unless instructed.

How the memory works in practice

Imagine that the first run identifies a missing attendance sheet. Claude records this under Open Issues. During the second run, it reads that issue and checks whether new source material resolves it.

Keep in PROGRESS.mdCurrent issues, blockers, decisions and what the next run should do.
Do not keep hereFull reports, complete chat transcripts, raw datasets or sensitive information.

Simple rule: If the information can change the next run, keep it in PROGRESS.md. If it is mainly a detailed record for a person to read, keep it in the outputs folder.

Step 4: Create LOOP_INSTRUCTIONS.md

This is the detailed procedure. TASK.md says what the workflow should achieve. LOOP_INSTRUCTIONS.md says how to perform the work safely.

Create and save the file

  1. Create a new file in the main workspace folder.
  2. Name it exactly LOOP_INSTRUCTIONS.md.
  3. Paste the complete instructions below.
  4. Save the file.
# Loop Instructions
You are running a weekly monitoring review workflow.
## Before You Start
1. Read `TASK.md`.
2. Read `PROGRESS.md`.
3. Review all files in `source-material/`.
4. Identify new information, unresolved issues, possible data-quality concerns and items requiring human review.
## Required Output
Write the review to:
`outputs/weekly-monitoring-review.md`
The report must contain:
1. Reporting period
2. Source files reviewed
3. Summary of important changes
4. Implementation concerns
5. Indicator or data-quality concerns
6. Missing or contradictory evidence
7. Emerging risks
8. Questions requiring human review
9. Recommended next actions
10. Verification result
## Evidence Rules
- Use only information found in the supplied files.
- Do not invent figures, findings, dates, quotations or explanations.
- Identify the source file supporting each important observation.
- Separate confirmed information from interpretation.
- Label incomplete evidence clearly.
- Do not treat missing information as proof of poor performance.
- When two files conflict, report the conflict and request human review.
## State Update
Before stopping, update `PROGRESS.md` with:
- Date and trigger
- Files reviewed
- Summary of the run
- Output produced
- Open issues and blockers
- Data-quality concerns
- Verification result
- What the next run should do
- Whether human review is required
## Permission Boundary
You may write only to:
- `outputs/weekly-monitoring-review.md`
- `PROGRESS.md`
You must not:
- Modify, delete, rename or move source files
- Create extra files without instruction
- Access files outside this workspace
- Send messages or publish content
- Make final programme, safeguarding, funding or evaluation decisions
## Verification Checklist
Before ending, check that:
- The output file exists
- Every required report section is present
- Important observations identify their source files
- Unsupported claims have been removed
- Missing and contradictory evidence is clearly labelled
- `PROGRESS.md` has been updated
- No file outside the permitted write list was changed
Record each check as PASS or FAIL.
## Failure and Stop Rules
- Correct a missing report section once.
- If the same verification item fails twice, stop.
- Stop immediately if a forbidden file was changed.
- Stop if sensitive personal data is discovered.
- Stop if the evidence does not support the requested conclusion.
- Mark the output as requiring human review whenever professional judgement is needed.

Why the instructions are detailed

  • Before You Start prevents Claude from acting without reading the task and previous state.
  • Required Output prevents an unstructured essay.
  • Evidence Rules reduce unsupported statements and hidden assumptions.
  • Permission Boundary limits which files may change.
  • Failure and Stop Rules prevent endless retries and return difficult decisions to a person.

Step 5: Add Sample Monitoring Notes

Now create three fictional input files inside the source-material folder.

File 1: field-update-01.md

  1. Open the source-material folder.
  2. Create field-update-01.md.
  3. Paste and save:
# Field Update 01
Reporting week: 6–10 July 2026
- Two planned community meetings were completed.
- Attendance records were received for one meeting.
- The second attendance record is still pending.
- Field staff reported delayed transport to one target area.
- No verified change to the beneficiary total was reported.

File 2: field-update-02.md

# Field Update 02
Reporting week: 6–10 July 2026
- The team distributed 180 information leaflets.
- The activity plan expected 200 leaflets.
- The note does not explain the difference.
- A follow-up visit is proposed for the following week.

File 3: partner-note.md

# Partner Note
- The partner reports that all planned community meetings were completed.
- Supporting documentation has not yet been uploaded.
- The partner requests confirmation of the reporting deadline.

Create the empty output file

  1. Open the outputs folder.
  2. Create a file named weekly-monitoring-review.md.
  3. Type # Weekly Monitoring Review.
  4. Save the file. Claude will replace or expand this content during the run.

What the test is designed to reveal: the notes report completed meetings, but one attendance record and other supporting documentation are missing. The loop should describe the activity as reported—not fully verified.

Step 6: Open the Folder in Claude

Choose the instructions for the method you selected earlier.

Method A: Claude Desktop Code tab

  1. Open the Claude Desktop application.
  2. Open the Code tab.
  3. Choose Local as the environment.
  4. Click Select folder.
  5. Navigate to Documents and select me-monitoring-loop.
  6. Confirm that the selected folder name appears in the session.
  7. Use Manual mode for the first run so proposed file changes require your review.

Windows note: Anthropic’s Desktop documentation states that Git is required for local Code sessions on Windows. Follow the official Desktop quickstart if the local folder option reports a Git problem.

Method B: Claude Code CLI

First check installation:

claude --version

If the terminal displays a version number, Claude Code is installed. If it says the command is not found, use the official Claude Code quickstart before continuing.

Windows: open Terminal in the correct folder

  1. Open the me-monitoring-loop folder in File Explorer.
  2. Right-click an empty area and choose Open in Terminal. On some Windows versions, use the File Explorer address bar, type powershell and press Enter.
  3. Type claude and press Enter.

macOS: open Terminal in the correct folder

  1. Open Terminal.
  2. Type cd with a space after it.
  3. Drag the me-monitoring-loop folder from Finder into the Terminal window. Its path appears automatically.
  4. Press Enter.
  5. Type claude and press Enter.

Claude should show the current working directory. Confirm that it ends with me-monitoring-loop. This matters because Claude reads files relative to the folder where the session starts.

Step 7: Run the Workflow Manually

Do not schedule anything yet. The first objective is to prove that the instructions work while you are watching.

Copy and send this prompt

Run the weekly monitoring review workflow for this workspace.
Before acting:
1. Read TASK.md.
2. Read PROGRESS.md.
3. Read LOOP_INSTRUCTIONS.md.
4. Review every file in source-material/.
Write the report to outputs/weekly-monitoring-review.md.
Update PROGRESS.md before stopping.
Complete the verification checklist.
Do not modify any other files.
Stop and request human review if evidence is missing, contradictory or sensitive.

What may happen next

Claude may ask for permission before writing to a file. Read the proposed path carefully.

  • Approve: changes to outputs/weekly-monitoring-review.md.
  • Approve: changes to PROGRESS.md.
  • Reject: changes to any file inside source-material.
  • Stop and investigate: any request to access an unrelated folder, secret file or external service.

Do not approve automatically because a request came from Claude. The path and action must match the permission boundary you wrote.

Step 8: Check the Result Manually

Claude’s verification checklist is useful, but a person should inspect the first runs. Open the files yourself.

Open the weekly review

  1. Open the outputs folder.
  2. Open weekly-monitoring-review.md.
  3. Check each item below.
  • The reporting period is included.
  • All three source filenames are listed.
  • The two completed meetings are described as reported.
  • The missing attendance record is clearly identified.
  • The difference between 180 and 200 leaflets is identified without an invented explanation.
  • The partner’s claim is not treated as fully verified when supporting documents are missing.
  • Questions requiring human review are listed.
  • The verification section contains clear PASS or FAIL results.

Open PROGRESS.md

Confirm that Claude recorded:

  • The run date or trigger.
  • The files reviewed.
  • Open evidence gaps.
  • Whether human review is needed.
  • What the next run should check.

Confirm the source files did not change

Open each source note and confirm the original text is unchanged. In Claude Desktop, also inspect the proposed or accepted file changes. Only the output file and PROGRESS.md should appear as modified.

A successful run is not a perfect report. It is a controlled run in which the output is reviewable, uncertainty is visible and forbidden files remain unchanged.

Step 9: Test a Second and Third Run

A loop must work repeatedly. One successful run only proves that it worked once.

Second run: resolve one open issue

  1. Create a new source file named attendance-update.md.
  2. Write: The missing attendance record for the second meeting was received and confirms 24 participants.
  3. Run the same manual prompt again.
  4. Check whether the new report recognises that the previous open issue has been resolved.

Third run: introduce conflicting information

  1. Create a file named conflicting-note.md.
  2. Write: The attendance list records 26 participants.
  3. Run the workflow again.
  4. Confirm that Claude reports the conflict between 24 and 26 instead of selecting one number.

Record the tests

TestExpected behaviourActual behaviourPass?Change needed
First runCreates structured report and state
Resolved issueRecognises the new attendance record
Conflicting evidenceReports 24 versus 26 and escalates

If the loop misses a problem, edit LOOP_INSTRUCTIONS.md to make the expected behaviour more explicit. Then repeat the test. This is normal workflow improvement—not a failure.

Optional: Schedule the Workflow

Skip this section during your first session. Add scheduling only after at least three manual runs behave predictably.

Session-based repetition with /loop

Anthropic documents /loop for repeating a prompt inside the current Claude Code session. It is useful for short polling or testing. The task is session-scoped, and recurring tasks expire after seven days.

/loop 24h Run the weekly monitoring review workflow.
Read TASK.md, PROGRESS.md and LOOP_INSTRUCTIONS.md first.
Review source-material/.
Write outputs/weekly-monitoring-review.md.
Update PROGRESS.md.
Run the verification checklist.
Stop and request human review when required.

For a short supervised test, you could use a shorter interval. Do not use frequent runs when the source folder is unlikely to change.

Completion-driven work with /goal

/goal is not a calendar schedule. It tells Claude to continue working until a verifiable completion condition holds.

/goal outputs/weekly-monitoring-review.md contains every required section, PROGRESS.md is updated, all verification checks pass, and no source file has been modified

Durable scheduling

  • Desktop scheduled task: runs locally and can access local files, but the application must be open and the computer awake.
  • Remote routine: runs on managed cloud infrastructure and is appropriate when the computer may be off.
  • External automation: GitHub Actions or another controlled system may be appropriate for repository-based workflows.

Scheduling rule: Never automate a task merely because the command is available. First define the input, output, verification standard, retry limit and human escalation point.

Permissions and Sensitive Data

Written instructions help, but technical permissions provide a stronger boundary.

Use Manual or Plan mode while learning

In Manual/default mode, Claude asks before file edits and commands. Plan mode uses read-only tools to prepare an approach without editing. Use higher oversight for sensitive or unfamiliar work.

StagePermission levelM&E example
1Read-onlyInspect files and list possible issues
2Draft outputWrite only to outputs/ and PROGRESS.md
3Controlled copy or sandboxEdit a duplicate dataset, never the original
4Human-approved actionSubmit or send only after explicit approval

Examples of information to exclude

  • Names, phone numbers, addresses and identification numbers.
  • Health, disability, protection or safeguarding information.
  • Confidential complaints and referral records.
  • Passwords, API keys, tokens, environment files and private keys.
  • Restricted evaluation findings or draft donor submissions.
  • Data used to make automated eligibility, service, funding or rights decisions.

For organisational use, follow your data-protection, information-security, informed-consent and responsible-AI requirements. The tutorial’s fictional notes are deliberately non-sensitive.

Troubleshooting Common Problems

My file is named TASK.md.txt

Windows added a hidden text extension. Turn on File name extensions in File Explorer, rename the file and remove the final .txt. The final name must be TASK.md.

Claude cannot see the files

Check that the session opened the main me-monitoring-loop folder—not Documents, the outputs folder or another location. In the CLI, confirm the working directory before starting Claude.

The output file remains empty

Check whether Claude requested permission to write and whether the request was rejected. Also confirm that LOOP_INSTRUCTIONS.md names the exact output path.

Claude tries to modify a source note

Reject the change immediately. Restore the original if needed, return to Manual mode and strengthen the permission boundary. Do not continue until source files remain read-only.

The report invents an explanation

Add a stricter rule: “When the source does not explain a difference, write ‘reason not provided’ and request human review.” Then repeat the test.

PROGRESS.md becomes very long

Keep only the current state and latest run. Move older detailed reports to an archive folder. State should help the next run, not become a full history.

The claude command is not found

Claude Code CLI is not installed or not available in the terminal path. Use Claude Desktop instead or follow the official Claude Code quickstart. After installation, restart the terminal and run claude --version.

The /loop or /goal command is unavailable

Check the installed version, update Claude Code and review the current official documentation. These features require recent versions and may depend on the interface or environment.

Practical Exercise

Complete two manual runs and submit:

  • TASK.md
  • PROGRESS.md after the second run
  • LOOP_INSTRUCTIONS.md
  • The source-material folder with fictional notes
  • outputs/weekly-monitoring-review.md
  • A completed verification checklist
  • A short reflection: What required human judgement, and what should never be automated?

Frequently Asked Questions

What are Claude loops for M&E?

Claude loops for M&E are repeatable AI-assisted workflows that read defined inputs, prepare a controlled output, record progress, verify requirements and stop or escalate when human judgement is needed.

Can I create the folder on my Desktop?

Yes. Documents is recommended because it is easier to keep organised, but Desktop works. The important point is that you can find the folder and select it as the Claude workspace.

Can I use Google Drive, OneDrive or Dropbox?

Yes, when the folder is synced and available locally in File Explorer or Finder. Avoid online-only files during the first exercise.

Can I do this in a normal Claude chat?

You can upload files and test the instructions manually, but a normal chat does not automatically become the same local, persistent, file-writing workflow described here.

Do I need to understand Markdown?

No. Copy the provided templates exactly. A line beginning with # is a heading, and a line beginning with - is a bullet point.

Why are TASK.md and LOOP_INSTRUCTIONS.md separate?

TASK.md contains the stable goal and scope. LOOP_INSTRUCTIONS.md contains the detailed operating procedure. Keeping them separate makes each file easier to understand and update.

Can the loop replace an evaluator or M&E specialist?

No. It can organise evidence, apply defined checks and prepare drafts. Interpretation, validation, accountability and professional decisions remain human responsibilities.

Should I schedule the first run?

No. Run it manually at least three times. Scheduling an unreliable workflow only repeats its mistakes more often.

Final Takeaway

A useful Claude loop begins with a small, understandable folder—not with complex automation.

Create the files, define a narrow task, keep source material read-only, require evidence traceability, inspect the first runs and stop whenever professional judgement is required. Scheduling and external connectors come later.

Official Claude Code Sources

Product interfaces, plan access, commands and limits may change. Check the official documentation before publishing or teaching this tutorial.

Academy

The courses and articles are developed by a team of experienced evaluators, collaborators, authors, and software developers, guided by Fation Luli. EvalCommunity Academy combines practical expertise in Monitoring & Evaluation and International Development with the latest advances in AI to create high-quality, accessible, and practical learning experiences for professionals worldwide.