How do I share promptfoo results with my team?

Run `promptfoo eval -o results.html`, then publish that file to Comma — drag it in, or PATCH it to the API from CI. Everyone gets the same grid at one URL, with comments anchored to the rows that failed.

Why not just use `promptfoo view`?

`promptfoo view` starts a local web server on your machine. It is the right tool while you iterate, and useless for a teammate in another timezone. The HTML output is the artifact you can actually hand over.

Can I publish a failing eval from CI?

Yes, and you should. Run the eval step with `if: always()` (or the equivalent) and PATCH the same report id whether the thresholds passed or not — a red eval is the one people need to read.

Does the comparison grid stay interactive?

Yes. Report HTML renders with scripts enabled inside a sandboxed iframe, so sorting, filtering and the per-cell output panes behave as they do locally.

How do I track whether a prompt change actually helped?

PATCH one report id per eval suite. Each run appends a revision, so the URL shows today's grid and keeps every earlier one — which is the comparison you actually wanted when you asked whether the new system prompt is better.

Share a promptfoo eval report

Everyone on the team has an opinion about the system prompt. Exactly one person has the eval grid open, on localhost, behind promptfoo view.

That asymmetry is why prompt decisions get made from vibes and one cherry- picked screenshot of a good row.

Export the grid, publish the file

promptfoo eval -o results.html

Drag results.html into the app, or from CI:

promptfoo eval -o results.html --no-progress-bar

curl -fsS -X PATCH "https://commareports.com/api/v1/reports/$REPORT_ID" \
  -H "Authorization: Bearer $COMMA_API_TOKEN" \
  -H "Content-Type: application/json" \
  -d "$(jq -n --rawfile html results.html \
        --arg title "Prompt eval — $GITHUB_SHA" '{title: $title, html: $html}')"

Run the publish step with if: always(). An eval that dropped below threshold is the exact artifact the team needs in front of them, and it is the one a continue-on-error: false job throws away.

Why the review layer matters here more than usual

Eval output is not a pass/fail number, it is a set of judgment calls about what a model should have said. Those judgments belong next to the row:

Anchored threads — "this assertion is wrong, the model's answer is better than the reference" sits on that cell. See commenting on HTML.
Revisions — one report id per suite, one revision per run. The prompt change either moved the grid or it didn't, and both are visible.
Access per report — evals often carry customer-shaped inputs. Keep them private, team-only, or domain-gated. See the sharing model.

Close the loop with the agent

If a coding agent is the one editing the prompts, it can read the review comments back through the MCP server and act on them — the same loop described in let an agent respond to comments. And a routine can re-run the suite nightly so the URL never shows last week's model.

Limits

Entry HTML: 5 MB. A large grid with full outputs per cell can approach it; narrow the eval or split by suite. Assets: 25 MB per file, 250 MB and 500 files total.
60 requests/minute per token.

Try it

Comma is free — unlimited reports, unlimited commenters, unlimited revision history.

Publish an eval report →