First task: find a flow

Find candidate flows for carbon dioxide emissions to air without creating data. Complete skill installation and sign-in first, with the AI client and terminal open in the same project directory.

1. Prepare the request file

Download the complete request JSON to the project root as flow-search.request.json, or save the following in an editor. It contains no credentials and is safe for the AI to read.

{
  "query": "Carbon dioxide emission to air",
  "filter": { "flowType": "Elementary flow", "asInput": false }
}

2. Give the AI the task and boundaries

Send this task to the client that has loaded the skill. Keep the search terms and JSON keys unchanged.

Use the installed flow-hybrid-search skill to find candidate flows using flow-search.request.json in this project.
This is read-only: do not save drafts, publish, delete data, or change configuration.
Check CLI sign-in first. If login is needed, pause and ask me to authorize in a trusted terminal/browser; do not request a password or token.
Run a dry run into flow-search.preview.json, then make the real query and save the raw response as flow-results.json.
Report whether execution succeeded and how many records were actually returned. Show at most 5 candidates with their IDs, versions, and names/types only where present in the response.
Explain the filters and limitations. Do not invent missing fields or treat relevance order as proof of scientific suitability.

3. Check the same operation in a terminal

These equivalent commands run from the project root. You can run them yourself first to separate client skill-loading problems from runtime problems. For a Codex installation:

Run commands one at a time. Immediately afterward, use echo $? on macOS/Linux or $LASTEXITCODE in PowerShell 7 to read the exit code; later commands can replace it. Except for explicitly expected sign-in/error demonstrations, stop on a nonzero code and resolve the error before continuing.

2> saves error output separately as text. Before sharing logs, check for account information, tokens, or private data. An empty result file is not evidence of success.

node .agents/skills/flow-hybrid-search/scripts/run-flow-hybrid-search.mjs --published-cli --input ./flow-search.request.json --dry-run --json > flow-search.preview.json 2> flow-search.preview-errors.txt

node .agents/skills/flow-hybrid-search/scripts/run-flow-hybrid-search.mjs --published-cli --input ./flow-search.request.json --json > flow-results.json 2> flow-search.errors.txt

For OpenClaw, replace the script prefix .agents/skills/ with skills/; all other arguments stay the same. --published-cli explicitly selects the pinned published CLI and avoids any preexisting local-development override.

4. Check the evidence, not just the AI’s answer

1. The preview should contain dryRun: true and a request.body matching your file. It makes no business query and is not completion evidence.
2. The real query must exit 0. Open flow-results.json: nonempty results use { "data": [...] }; no matches may use [] or { "data": [] }.
3. Compare the AI’s reported count with the raw array length. IDs, versions, and names must come from that response; missing fields must be identified as unavailable.
4. Before selecting a record, check flow type, classification, geography, and unit/flow property. Read-only search does not prove that the candidate fits your model.

For no matches, record a successful empty search, then refine terms or relax filters gradually. A failed command, 401/403, or server error does not mean “no data.” Keep request, response, and error code and follow troubleshooting.

You are done when you can locate the request, preview, and actual result, explain the real record count, and name at least one suitability check still requiring human judgment. A fluent AI answer without a real query result is not completion.

Next, try the other two standalone search skills or read Organize tasks safely.