Getting Started

Choose a task, copy one complete stage to your assistant, check the output, and resume from retained artifacts. Claude packages are optional.

Status: Start with the shared task library and its portable text instructions for the assistant you choose. The Claude package examples in this guide retain earlier, bounded evidence; they do not establish execution of the current task route or current source version. Earlier Claude Code checks and one paid-account browser skill check were recorded separately. Free-account execution and practitioner usability were not checked. Confirm actual account capabilities and inspect every output before use.

Start Here

  1. Choose an Outcome in the shared task library. Read its minimum inputs and suitability checks.
  2. Copy One Complete Stage into the Claude or ChatGPT conversation you choose, then supply only the readable, de-identified material that stage needs. No plugin installation is required for this portable route.
  3. Inspect the Output and Named Handoff. Check sources, decisions, calculations and unresolved issues before continuing to the next stage. Text and Markdown tables are the baseline; downloadable files depend on actual host tools.
  4. Revise and Resume from Actual Artifacts. Retain the source manifest, v1, review issues and user dispositions. A summary or prior chat memory does not replace them.
  5. Use Existing Claude Packages Only as Optional Shortcuts. The command tables below preserve those package routes and their earlier evidence. Current portable methods can differ from retained package bytes.

Words used here: a skill is a set of method instructions for one job. A plugin is a package of skills. A handover is a short note that lets a new session continue the work. Proposed means the assistant suggests it and nobody has approved it.

1. Choose the Job

There are 24 jobs: 12 document reviews and 12 job skills that draft, plan or check. Each name links to its page on this site, with the start instructions, a fictional example and downloads. The full method (SKILL.md) governs if it ever differs from the short start instructions.

Eight Core Job Skills

What you needCode invocationMinimum starting materialMethod and start instructions
Frame evaluation decisions/me-workspace:evaluation-questionsEvaluation purpose; intended decision/use if known; program and scopeEvaluation questions
Match questions to evidence and methods/me-workspace:question-method-alignmentEvaluation questions or purpose; population/program context; known evidence sourcesQuestion-method alignment
Draft an evaluation TOR/me-workspace:evaluation-tor-draftProgram description and evaluation purpose; known users, scope, timing, or constraintsEvaluation TOR draft
Document indicators/me-workspace:indicator-reference-sheetsIndicator name or result statement; any available framework/source/definitionIndicator reference sheets
Assess measurement burden/me-workspace:measurement-feasibilityIndicator list or measurement plan; at least one known constraint to assessMeasurement feasibility
Draft an M&E plan/me-workspace:me-plan-draftProgram summary/results framework; indicator list or monitoring purposeM&E plan draft
Check progress evidence/me-workspace:progress-evidenceProgress report; a framework/definitions help check coverage; prior comparable period for trendsProgress evidence
Turn findings into proposed actions/me-workspace:findings-to-actionsFindings or recommendations and their source; program context; any existing responseFindings to actions

Twelve Document Reviews

These review skills assess a supplied document or instrument. For cross-document checks, attach the related framework or source documents too. The linked method explains which inputs matter for that review.

What you need reviewedCode invocationMinimum starting materialInstructions
Logframe or results framework/me-review:logframe-reviewLogframe; proposal or ToC if checking alignmentLogframe review
Theory of Change/me-review:theory-of-change-reviewToC document or description of causal logicToC review
Indicators/me-review:indicator-qualityIndicator list and the results they are intended to measureIndicator quality
M&E plan/me-review:me-plan-reviewM&E plan; framework and requirements help check coherenceM&E plan review
Proposal M&E section/me-review:proposal-me-reviewProposal M&E section; relevant call or framework if checking fitProposal M&E review
Evaluation TOR/me-review:tor-reviewTOR or commissioning brief; program framework if checking alignmentTOR review
Survey instrument/me-review:survey-reviewQuestionnaire or survey instrument; indicator definitions if availableSurvey review
Sampling plan/me-review:sampling-plan-reviewSampling plan or description of methodology; purpose/population/frame helpSampling plan review
Baseline report/me-review:baseline-report-reviewBaseline report, study, or data summary; logframe for coverage checksBaseline review
Progress report/me-review:progress-report-reviewProgress report; framework for coverage and prior period for comparisonProgress report review
Data-quality assessment/me-review:data-quality-reviewDQA report, methodology, or findings and the data stream assessedData-quality review
Evaluation report or findings/me-review:evaluation-report-reviewEvaluation report or findings brief; TOR/questions help check coverageEvaluation report review

Four More Job Skills

Earlier versions of these four job skills had bounded Claude Code checks on ordinary, incomplete and unsuitable inputs, through their own commands and seven older command names. Later reruns repaired initial failures while retaining minor limitations. That evidence does not establish the current portable task route. Practitioner feedback was not gathered.

What you needCode invocationMinimum starting materialInstructions
Compare results documents/me-workspace:results-chain-alignmentAt least two distinct design artifacts, including one results structureResults-chain alignment
Draft a sampling plan/me-workspace:sampling-plan-draftPurpose/question, intended population/unit, and known sampling frameSampling-plan draft
Plan a data-quality assessment/me-workspace:data-quality-planningIndicator or data stream, intended use, and available data-flow/risk informationData-quality planning
Plan baseline evidence/me-workspace:baseline-planningIndicator set, population, and exposure timing if knownBaseline planning

2. Pick an Input Route

Optional Claude Code Package Route: Earlier Evidence

Earlier bounded checks used the local workspace plugin and review plugin together, with an existing Claude subscription and readable Markdown files. The local launch form is:

claude --plugin-dir <me-workspace-folder> --plugin-dir <me-review-folder>

Replace each placeholder with the folder containing that plugin on your computer. If you already choose the optional marketplace route, the existing commands are /plugin marketplace add mestudiohq/me-review, /plugin install me-review@me-review and /plugin install me-workspace@me-review. Check the package version and permissions before installing. The portable task route does not need these commands. After launch, use one of the exact /me-workspace:<skill> or /me-review:<skill> invocations above. Name the specific files or one scoped folder. If you name a folder, ask the assistant to list candidate files and use only the files you identify.

Those earlier method and workflow checks apply only to the recorded source versions, Code setup and editable Markdown. They do not show that every route works with every account, provider setting, model, file type, or machine.

Optional Browser Upload: Earlier Limited Evidence

The earlier package work recorded an evaluation-questions skill chat check on one paid Claude browser account. Four reserve ZIPs uploaded without errors; upload alone was not a chat-execution check. No Free-account execution was established. These earlier checks do not establish current source, portable-copy, ChatGPT production-task or native-output acceptance.

Current task pages supply complete stage methods for copying. Confirm that your chosen assistant can read the actual material and perform the requested operation. A browser cannot infer a local folder path. If upload or execution is unavailable, use readable text and the task's specification/table fallback.

Official references: Claude Skills and Claude Code subscription access. Check current account settings and provider documentation; this guide does not guarantee a plan, installation or file capability.

No Site Execution or Required Installation

The task library supplies instructions; it does not process your documents or call a model API. You supply material to the assistant you choose under its actual account controls and policies. No M&E Studio signup or plugin installation is required for portable use. Provider access, limits and charges are separate.

3. Start One Job

  1. Choose a task from the tables above and open its linked start instructions.
  2. Identify the source files you intend to provide. Prefer only the documents needed for the job. Use sources you are authorized to share with the selected service.
  3. State the purpose and desired artifact. Include known scope, decision, timeframe, and constraints. If a fact is unknown, say so.
  4. Use the starter prompt from that skill's START.md, replacing bracketed text. Do not paste a machine path into a browser request; upload or attach the readable file instead.
  5. For a complete workflow, use the relevant workflow pack's own START.md: Evaluation planning, Feasible measurement, or Reporting to action. Workflow packs name their component jobs and expected Markdown files.
  6. Ask for a bounded draft when information is missing. A useful output can include explicit gaps and decisions instead of guessing.

4. Inspect the Actual Markdown Artifact

The earlier package workflows targeted editable Markdown. The current task library uses text and tables as its baseline. Native DOCX/XLSX is conditional on an actual attachment being created and reopened; otherwise use the clearly labeled text/table fallback. A chat summary is not proof that a file exists. After a run:

  • Confirm which .md files were actually created and open each one. If no file was created, say so; use the inline Markdown only if you can save it yourself.
  • Compare key facts, figures, dates, definitions, and commitments to the supplied sources. Follow evidence IDs back to the source and verify the cited location.
  • Separate facts in the source from assumptions made for drafting and proposals suggested by the assistant. Proposed owners, deadlines, actions, crosswalks, or revisions are not approvals.
  • Use three evidence states: confirmed, contradicted, and unconfirmed. A missing record is unconfirmed, not proof of a negative. A source that directly conflicts with a claim is contradicted. Do not convert “not found in supplied files” into “does not exist.” Skills may also say not supplied, not assessable or decision required: these are kinds of unconfirmed, never contradicted. Use contradicted only when a supplied source directly conflicts with the claim.
  • Preserve source versions and unresolved disagreements. Do not silently choose which version governs, or rewrite an approved indicator, target, or commitment. If one version is signed and a later one is not, the skills use the signed one as a proposed working version and flag the other; you decide.

Before a client, donor or manager sees an output:

  • A person who knows the program reads the whole file, not only the summary.
  • That person checks every figure, date and commitment against the source files.
  • Every proposed owner, date and action is either approved by the right person or still marked proposed.
  • The file says it is an AI-assisted draft and who reviewed it. The skill's own checklist is not a review.
  • Recheck arithmetic and comparison basis. Keep period and cumulative values distinct, use matching definitions and populations, and do not describe a single period as a trend. Pre/post change alone does not establish causation.
  • If your interface cannot create a file, request a Markdown response and save it manually. Do not call inline text a downloaded artifact.

5. Pause and Resume with a Real Handover

Before pausing, save the original inputs, every actual intermediate or final artifact, and a short handover. Use the optional getting-started handover template. In a new session, provide those same source files, artifacts, and handover. Resume only from material you supply in that session; do not rely on chat memory or assume the prior workspace is available.

The handover is a navigation aid, not authority for facts or decisions. Recheck inherited claims against the original sources. Record source filenames and versions/dates, completed artifacts, unresolved choices, and the exact next step. If sources changed, flag the differences. Never upgrade a draft proposal, assumption, pending approval, or unknown status into an approved decision.

Troubleshooting

What happenedWhat to do
The assistant cannot identify a document or several files have similar namesProvide the specific file, version, or an explicit file list. Do not let it guess which version controls.
A file could not be readProvide a readable text or Markdown export and state the original source and version. Do not claim the unread file was reviewed.
Information is missingAsk for a bounded draft with “Not supplied,” “Unconfirmed,” or “Decision required” labels. Do not fill targets, budgets, sample sizes, roles, dates, or results from guesswork.
Two sources disagreeShow both with their source references and identify the decision needed. Do not silently reconcile them.
Only a chat summary appears, with no fileAsk for the expected Markdown content or save the inline response yourself. Verify the saved file's contents.
The browser route does not load or produce a usable resultTreat that route as unverified. Check actual account capabilities and use readable text with the task's supported fallback. Preserve completed work; do not claim a missing file was read or generated.
A result seems causalCheck whether the supplied design supports attribution. If not, describe observed change, association, or the stated hypothesis without claiming causation.
Work is interruptedSupply the original inputs, actual artifacts, and a source-checked handover when resuming.

Earlier Evidence and Current Limits

The earlier PD3 acceptance selected 52 accepted Code/Markdown checks from 85 completed invocations at frozen source 38fac1f. Coverage included eight pilot jobs, four adapted reviews, seven adapters, three workflows and interruption/resumption; it was not a blanket retest of the catalogue. Later reserve, routing and authority reruns repaired initial failures and retained minor baseline-fallback, DQA-owner and local-path limitations. Earlier fresh-session worked-project checks used these guides.

Earlier browser evidence covers one evaluation-questions skill on a paid account. Four reserve ZIPs uploaded without errors, which does not establish execution. No Free-account execution or practitioner usability acceptance was recorded. The other eight review methods have separate earlier-release evidence.

Current method files are projected from their recorded source revision and exact digests. Retained ZIPs preserve earlier bytes and provenance; use the current portable text for source repairs such as the DQA threshold correction. Source projection, earlier host execution, installation, package release and website deployment are separate facts. These historical checks do not pass the new task route or every format, language, source version or assistant. Consult each current task/resource evidence record for the exact scope; unperformed checks remain unperformed.

Source: M&E Studio, monitoringevaluationstudio.com. Guide and fictional attachment use is governed by the existing free-content terms; public plugin files retain their separate MIT license.

See usage and credit for how to credit M&E Studio.