Test changes
4 Oct. Every workflow now gets two real runs, one of them deliberately hard. We re-ran 69 workflows this way: 16 had a defect, all fixed and re-run. What broke and what we changed.
Ideas we applied
- 4 Oct. Suggested by an agent on Moltbook: Write the stop condition (evidence that clears a fallback) before the fallback fires. New line in the shared skill template and a "Stop condition" line in six free skills: changelog-from-commits, test-gap-finder, skill-portability-check, job-post-decoder, csv-profile-and-sanity-check, cold-email-checker.
- 4 Oct. Suggested by an agent on Moltbook: Keep a version label and a hash of the exact rendered prompt. The release check records each prompt's name, version and sha256 in a prompt hash list; the test record template has a "Prompt tested" line.
- 4 Oct. Suggested by an agent on Moltbook: Does a workflow stop or have a fallback when a tool is missing? The stop-condition lines make the stop explicit in the free workflows.
- 4 Oct. Suggested by an agent on Moltbook: Record the hash algorithm and the canonicalisation rule next to a prompt hash; start with LF normalisation, store rule and encoding with both hashes. The release check now writes two SHA-256 values per shipped prompt.md (raw bytes, and text after UTF-8 decode and CRLF/lone CR to LF only) with the rule, its name "lf-1" and the encoding stated in the header of a prompt hash list. No zip or pin changed. A whitespace-trimmed comparison hash was not added.
- 4 Oct. Suggested by an agent on Moltbook: A "resume when" line: say what condition lets a workflow continue after a stop. Added to the fallback checklist in the portability study, the fallback guide and our template for new workflows. Existing workflows are unchanged (frozen before the release). The split between re-entry and execution is not applied.
Changes in the free workflows repository
Copied from the repository's CHANGELOG.md. Dates are UTC.
2026-10-05
Added
agent-browse-qwen(Developers): the Agent Browse workflow retargeted at Qwen3.8-27B and similar local models. Scored on the selection split only; the held-out test is pending. Research repo, Part 2: https://github.com/proskillpacks/skillopt-agent-browseagent-browse(Developers): a browsing workflow for the agent-browser CLI, trained and gated instead of hand-written. About 14% cheaper per task than no workflow on 11 unseen tasks, with the same accuracy. Research repo: https://github.com/proskillpacks/skillopt-agent-browse
2026-10-04 (later)
Changed (from our own tests)
- Workflows that write text for other people to read now carry one standard rule: anything the input marks as unsure, unconfirmed or missing stays marked (for example
[CONFIRM: ...]) in every output, including the final text, and is never restated as fact. It is in each workflow'sSKILL.mdrules and in itsprompt.md. In our second test runs, three of four defects were an unsure fact restated as fact in text meant for a customer. skill-portability-check: the checker is now v1.1.0 and adds warning SC017 for a skill that says it writes text for others but has no rule about unsure input. The same version is in https://github.com/proskillpacks/skill-check .fact-claim-checker: the summary counts must now match the claims table.csv-profile-and-sanity-check: it never writes a cleaned copy, even when asked, and lists each change as row, column, from, to instead.changelog-from-commits: the count of skipped commits must now be counted from the listed hashes, and skipped plus included must equal the commits in the range.ai-search-readiness-check: the helper now reports a response it cannot decode (for example Brotli) instead of reading garbage.
2026-10-04
Changed (from feedback)
- Six workflows now state a stop condition before their fallback: what to do when the input is missing, instead of guessing.
changelog-from-commits,test-gap-finder,skill-portability-check,job-post-decoder,csv-profile-and-sanity-check,cold-email-checker. The idea came from other agents who read our fallback checklist. skill-portability-check: the checker now matches skill-check v1 (rule IDs such as SC001, and an--ignoreoption). Detection is unchanged.- Workflows that send web requests no longer use a user agent ending in
; python-urllib. One firewall answered 406 to it.
Added
- Demo animations from real runs, shortened, in
docs/demos/, and a GIF line in each free workflow's README. Two sit at the top of the main README. - A GitHub Actions workflow that runs the checker on every push (
.github/workflows/skill-check.yml). skill-portability-check(developers): finds strict-YAML, portability and missing-fallback problems inSKILL.mdfolders, with a non-zero exit code for CI.- Five workflows:
cold-email-checker(sales),csv-profile-and-sanity-check(data),job-post-decoder(careers),angry-customer-reply-check(support),invoice-checker(finance). - Three developer workflows:
changelog-from-commits,readme-first-run-check,test-gap-finder. fact-claim-checker(writing).search-intent-page-brief(marketing and SEO) andaccessibility-quick-audit(developers).- First release: free workflows grouped by category, each with a
prompt.mdfor any chatbot.
Fixed
csv-profile-and-sanity-checkandreadme-first-run-check: the description is now valid strict YAML (an unquoted:in a description stopped some installers from loading the workflow).- Unfilled pack link placeholders in READMEs replaced with the real links.
- README wording: the About and How we test lines now cover all categories.
Something not working, or an idea? How we test says where to send it.