Use maniva
Ask maniva real questions from your work through the AI assistant you already use. maniva is in alpha, so tell us where it helps and where it fails.
Before you start
- Do not upload unpublished or sensitive data. Your assistant's own privacy terms apply to anything you share with it.
- Treat answers as a starting point. Check anything you plan to act on against the cited source.
- Questions about access or setup go to k.desousa@cgiar.org.
-
Open an AI assistant that can read the web
maniva works best with assistants that can open several files from GitHub in one conversation, such as Claude with web access. Copilot and Gemini can read the main page but often cannot follow links into the layers, so answers may stay general. Start a new conversation so earlier chats do not mix in.
-
Share the maniva link
Paste the link, then ask your first question in the same message.
-
Ask questions from your own work
The best tests are questions you actually have. If you need a start, try one from each group.
- Evidence
- How many farms does a tricot trial need?
- What do farmers' rankings add to yield data?
- Protocols
- Which survey questions and traits should my bean trial use?
- What are the steps to run a tricot trial from start to finish?
- Data
- Which trials and datasets exist for sorghum?
- Load the common bean trials and summarise them by country.
- Analysis
- How should I analyse trials of my crop?
- How do I link farmer rankings with rainfall data?
-
Check what you get
A good maniva answer does three things. Watch for answers that do not.
- It cites its source. Findings name the paper, dataset or protocol they come from, with a link you can open.
- It says what it does not know. When maniva has no answer, the assistant should say so instead of inventing one.
- It keeps the layers apart. Evidence, protocols, data and methods are named as such, so you can tell a published finding from a suggestion.
-
Tell us how it went
The feedback form takes about 5 minutes.
Your AI assistant can also write a short report of how you used maniva, with personal details removed. Paste the text below into the same conversation where you used maniva, not a new one. Read the report it writes, edit anything you do not want to share, then paste it in the feedback form. This step is optional but it helps us most.
Report promptPrepare a feedback report on how I used maniva in this conversation. The report goes to the maniva team. PRIVACY. Remove all identifying information, including names, emails, phone numbers, institutions, locations finer than country, GPS coordinates, paths of my own files, private URLs, credentials and unpublished data values. Replace them with [PERSON], [ORG], [LOCATION] or [DATA]. Never include raw data rows or the contents of files I uploaded. Paths of public maniva files are fine. FORMAT. One markdown block, under 600 words. 1. Session - AI assistant and model - How you reached maniva (link, uploaded skill or connector) - maniva version and commit. Fetch https://raw.githubusercontent.com/AgrDataSci/maniva/main/version.json and report "version" and the first 7 characters of "commit". If you cannot fetch it, write "unknown". Never guess a commit. - Number of exchanges - My overall goal, in one sentence 2. Tasks. For each request I made, give - type (evidence, data, analysis, protocol, setup, other) - my prompt, anonymised, as close to verbatim as possible - the maniva files you actually opened, by repository and path - outcome (solved, partly solved, failed) - any wrong or invented content, and how it was corrected - number of follow-up prompts needed 3. Gaps. What I asked that maniva could not answer or do. 4. Friction. Setup problems and code errors, with anonymised error messages. 5. My suggestions. Only what I said or asked for. 6. Your observations. Anything you inferred yourself, kept separate from my suggestions. Report only what happened in this conversation. Show me the report so I can check it before sharing.