Skip to content

launch.md

Latest commit

 

History

History
169 lines (123 loc) · 5.71 KB

launch.md

File metadata and controls

169 lines (123 loc) · 5.71 KB
Raw

Public Alpha Launch Playbook

Use this after docs/release.md passes. The goal is a small, credible launch: one demo GIF, one install command, one clear product promise, and a tight feedback loop.

Launch Positioning

One sentence:

redline turns prompt-response logs you already have into a local eval suite that catches prompt regressions before you ship.

What to show:

  • A prompt change makes answers shorter.
  • redline catches lost JSON keys, URLs, numbers, refusals, tables, lists, or other high-signal behavior changes.
  • The user can mark intentional changes and accept reviewed outputs into the baseline.
  • No cloud account or judge is required for the default path.

What not to claim:

  • Do not claim full semantic correctness.
  • Do not claim hallucination detection without a configured judge or explicit requirements.
  • Do not describe neutral as safe to ship without review.

Asset Checklist

Generate launch artifacts from a clean release checkout:

bash scripts/release_check.sh
bash scripts/demo_gif.sh .redline/launch .redline/launch/redline-demo.gif
bash scripts/build_release.sh /tmp/redline-dist-v0.3.0

Required artifacts:

  • .redline/launch/redline-demo.gif or .redline/launch/redline-demo-transcript.txt
  • .redline/dashboard.html from the demo path
  • /tmp/redline-dist-v0.3.0/redline_ai-0.3.0-py3-none-any.whl
  • /tmp/redline-dist-v0.3.0/redline_ai-0.3.0.tar.gz

Website Checklist

The GitHub Pages homepage should make the product clear without reading the README:

  1. The first viewport says redline turns existing prompt-response logs into eval suites.
  2. The primary command is the proof path: redline demo --public --compact.
  3. The page shows the trust boundary before the adapter list: structural checks are deterministic; factual correctness, tone, hallucinations, and subtle reasoning need requirements or an optional judge.
  4. The page shows the closed loop: generate suite, evaluate prompt change, mark intentional changes, accept reviewed baselines.
  5. The page links back to GitHub and the quickstart, not to a marketing waitlist.

Demo GIF Storyboard

Keep the GIF short and concrete:

  1. Run redline demo --public --compact.
  2. Pause on the summary line: ten cases, ten regressions, high confidence.
  3. Let the viewer see four representative reasons: invalid JSON, lost table, new refusal, lost bullet list.
  4. Record history so the loop is not a one-shot diff.
  5. Generate the dashboard so the viewer sees the review artifact.

Do not show install logs, dependency downloads, or a long scrollback. The demo should sell the loop, not the package manager.

Publish Sequence

  1. Confirm main is clean, pushed, and certified.

  2. Tag the exact commit:

    git tag v0.3.0
git push origin v0.3.0

  3. Upload the release distribution:

    python -m twine upload \
  /tmp/redline-dist-v0.3.0/redline_ai-*.whl \
  /tmp/redline-dist-v0.3.0/redline_ai-*.tar.gz

  4. Create the GitHub release from v0.3.0.

  5. Attach the demo GIF or link the public README asset: https://raw.githubusercontent.com/gowtham0992/redline/main/site/assets/redline-product-demo.gif.

Launch Post

Short post:

I built redline, a local-first tool that turns prompt-response logs into eval
suites automatically.

You point it at existing JSONL logs, groups behavior by deterministic signatures, picks representative
cases, and catches regressions when a prompt change drops structure, numbers,
URLs, entities, refusals, tables, code blocks, or required fields.

No cloud account required. Optional judges are supported only for ambiguous
changed cases.

Install:
python -m pip install redline-ai

First run:
redline demo

Repo: https://github.com/gowtham0992/redline

Longer post:

Most eval tools start with "write test cases." redline starts from the logs you
already have.

It watches or imports prompt-response JSONL, groups observed behavior by deterministic signatures, builds
a representative suite, and compares new prompt runs against the accepted
baseline. The default checks are deterministic and local: JSON validity, missing
keys, empty outputs, refusals, URLs, numbers, entities, code blocks, tables,
lists, output shape, and obvious allow/deny polarity flips.

The point is not to replace semantic review. The point is to catch the boring,
expensive regressions that slip through prompt iteration and CI.

Try:
python -m pip install redline-ai
redline demo
redline init --runner stdio --copy-runner

First 10 Feedback Loops

Ask early users to file structured feedback with the GitHub dogfood issue template. Reports should include sanitized commands, what redline caught, where they hesitated, and any safe snippets from doctor, summary, history, or reports.

Track every early user report as one of:

  • blocked: they could not install, run redline demo, or load their logs.
  • confused: they understood the value but did not know the next command.
  • false-positive: redline flagged a case that felt noisy or unactionable.
  • false-negative: redline missed a regression they expected it to catch.
  • adapter-gap: their app did not fit an existing runner/log adapter.
  • docs-gap: the command existed but the documentation did not make it obvious.

For the first ten users, fix in this order:

  1. Anything blocking redline demo.
  2. Any private-log import issue.
  3. Any runner setup issue affecting OpenAI, Anthropic, LiteLLM, HTTP, LangChain, or app logs.
  4. Any trust-calibration issue that makes users over-trust neutral results.
  5. Any false positive that appears in more than one user workflow.

Do not add a desktop app during this phase. Improve the CLI, generated reports, dashboard, runner adapters, and docs until the first-use loop is obviously valuable.