Changelog

Added

• Requirements from Jira and Confluence. The plugin ships Atlassian's official remote MCP server (.mcp.json, https://mcp.atlassian.com/v2/mcp). Name an issue key or paste a Confluence link and Casely reads the issue, its acceptance criteria and linked issues like an attached spec. Sign-in runs through Atlassian's OAuth page; no API token is stored or asked for.
• /casely:run (Claude Code). Walks the exported UI cases against a URL through the Playwright MCP server, records pass, fail or blocked per case, saves a screenshot per failure and writes run/run_report.md. API cases stay with Newman.
• CSV export. export_to_xlsx.py --csv writes all_test_cases.csv next to the workbook: UTF-8 with a BOM so Excel and TestRail keep Arabic, Korean and other non-Latin text, standard quoting for commas, quotes and line breaks. Refused when cases carry different columns.
• Export round-trip test (tests/test_export_roundtrip.py). Arabic and Korean fixtures with commas, quotes, an escaped pipe and line breaks go through the XLSX and CSV exports; the test reads both back and compares every cell, the row count and the column count with the untouched fixtures. Suggested in a dev.to comment thread; Arabic coverage follows feedback in #9.
• /casely:bug (Claude Code; on claude.ai and Cowork, ask for a bug report). Writes one bug report in the team's format from a trace or event ID, a monitoring link, pasted logs, or a failed case from /casely:run. Reads the event through any connected monitoring MCP server (GlitchTip, Sentry, the company's own) with read tools only, or from pasted logs when the tester has no access. Masks tokens, emails, phone numbers, user ids and IPs; keeps the trace ID and link; writes unknown instead of guessing; checks Jira for likely duplicates when connected; never files the bug. The team's bug report examples are saved once as bug-style.md.
• Saved team style. After the first run Casely offers to save the style it learned as casely-style.md (.casely/style.md in Claude Code): headers in order, ID scheme and the last ID used, tone, and one example row with real data replaced by placeholders. Later runs read it instead of asking for an example; a changed example needs the user's yes before the file changes.
• Epics and sprints from Jira. Casely finds the stories with one JQL search, lists them with the count and asks which to cover, and offers the issue key as the requirement column.
• Read-only connectors. Casely uses only the read and search tools of a connector and refuses to write to Jira or Confluence. references/connectors.md shows how to restrict the Atlassian server to read and search, and a permissions.deny block for Claude Code that removes the write tools.
• Directory listing fields in plugin.json: icon, documentation, support and privacy links.

Changed

• SKILL.md is shorter: Jira, saved style and Postman details moved into references/, which Casely reads only when a run needs them.
• README: plugin directory install first, an install prompt for Claude Code, the star link points to the repository (the stargazers page returns 404 to signed-out visitors).
• SKILL.md: dropped an unsourced "~40% of a QA engineer's time" claim.

Added

• Casely is now a plugin for Claude Code and Codex. The repository is a plugin marketplace, so you can install it with two commands and update it later:

  # Claude Code
  claude plugin marketplace add JohnWayneeee/casely-qa-skill
  claude plugin install casely@casely
  
  # Codex
  codex plugin marketplace add JohnWayneeee/casely-qa-skill
  codex plugin add casely@casely
  

  New versions arrive with claude plugin update casely@casely. npx skills add keeps working as before.

• .claude-plugin/plugin.json and .claude-plugin/marketplace.json.

Changed

• skill/casely/ moved to skills/casely/, the standard plugin layout. If you keep a local clone or zip the skill yourself, update the path.
• Skill description shortened from 1,036 to 801 characters, under the 1,024-character limit in the Agent Skills specification. The triggers it lists are unchanged.
• marketplace.json moved from the repository root to .claude-plugin/, where Claude Code and Codex look for it.

For claude.ai and Claude desktop

Download casely-v2.3.0.zip below and upload it under Customize → Skills. casely/ sits at the archive root, which is what claude.ai requires. GitHub's own "Source code (zip)" wraps everything one level too deep, so don't use that one.

Added

• Postman collections for API test cases. When the requirements describe an API — endpoints as method + path, an OpenAPI/Swagger file, curl examples, response schemas, status codes, an auth section — Casely now exports the API-level cases as a runnable Postman v2.1 collection alongside the Excel file. It says so in the Phase 3 plan first, so the existing approval gate covers it and nobody receives an artefact they didn't ask for.
• scripts/build_postman_collection.py — assembles per-case JSON request specs into the collection, an environment file holding every variable, and a README covering import, variables, the Collection Runner and a Newman command for CI. Assertions are generated from each case's expected result rather than hand-written, so a typo cannot turn a broken endpoint into a green run.
• Everything environment-specific becomes a variable. {{baseUrl}}, {{authToken}}, entity ids and test data land in the environment file with descriptions and empty values, secrets typed as secrets. The value a case is actually testing stays literal — parameterizing amount: 50001 would hide what the case checks.
• The build refuses unsafe or unrunnable input, the way the Excel exporter already did: a hardcoded host, anything shaped like a JWT or a secret key, a duplicated case id, a missing or invalid field. It names the file and exits non-zero rather than writing a collection that points at production or carries someone's token.
• references/api_collection.md — the signals that decide whether a collection is worth building, the request spec format, request chaining, and the variable and assertion rules.
• Evals extended with the plan-time offer, the no-endpoints case, and the variable rules.

Changed

• Phase 5 is now "Export": Excel always, the Postman collection when the plan promised one.
• Casely never guesses an endpoint from a described flow. A spec that names screens but no paths gets a request for the API docs instead of a collection that 404s on first run.

Casely no longer runs on slash commands. Attach your requirements (and, ideally, an example of test cases your team already writes), say what you need, and Casely handles the rest in one conversation — across Claude Code, claude.ai, and the Claude desktop app.

What changed

• No more commands. /init, /parse, /style, /plan, /generate, /export are gone. Attach files, describe the task, done.
• One approval gate. Casely always stops on the test plan and waits for your OK before writing a single test case.
• Real test design. Equivalence partitioning, boundary values, decision tables, state transitions, error guessing — not a restatement of the requirements as cases.
• Reports gaps in your spec — untestable wording, contradictions, missing error paths — alongside the plan.
• One Excel file, one row per case. Matches how TestRail, Qase, Zephyr and Xray actually import, instead of forty separate files.
• No parser dependency. Claude reads PDF/DOCX/XLSX natively — nothing to install.

Full history in CHANGELOG.md.

Stop manually writing test cases. Claude skill: Requirements → Test Plan → + 100 Test Cases (Manual & Auto) in 5min 🚀

Writing test suites from scratch is the most tedious part of QA. So I built Casely — a virtual QA Lead for Claude (Anthropic Skills).

It doesn't just "hallucinate" tests. It follows a rigorous pipeline: Parse: Deep-reads PDF/DOCX into clean Markdown (using Docling).

Style: Analyzes your test cases to mimic your specific style/columns.

Plan: Segregates requirements into logical modules.

Generate: Cranks out 100+ atomic test cases.

Export: Spits out a .xlsx file ready for TestRail/Xray.

Real stats: 27-page requirement doc → 87 structured cases in 2:32 min.