name: mdthis description: Generate a markdown file from the current conversation context based on a short directive. Use when the user wants to capture discussion, findings, changes, or decisions as a markdown document. Not for writing or editing project documentation (README, docs/, ADRs) or any file the user wants at a specific repo path — mdthis only produces one-off timestamped artifacts in local/, distilled from the conversation. argument-hint: <what to write, e.g. "jira ticket", "PR comment", "latest findings">

mdthis — Conversation-to-Markdown Skill

You have full access to the current conversation context. The user wants you to produce a markdown file based on everything discussed so far, shaped by their directive.

Directive

$ARGUMENTS

Instructions

1. Infer the format from the directive. Match the tone, structure, and conventions of the implied document type:

   • "jira ticket" / "ticket" → Title, Description, Acceptance Criteria, Technical Notes
   • "PR comment" / "PR description" → Summary, Changes, Test Plan
   • "latest findings" / "findings" → structured summary of discoveries, conclusions, open questions
   • "outline" / "summary" → hierarchical overview of what was discussed
   • "runbook" / "playbook" → step-by-step operational instructions
   • "bug report" → Steps to Reproduce, Expected vs Actual, Root Cause, Fix
   • "meeting notes" / "status update" → agenda items, decisions, action items
   • "doc" / "documentation" → technical documentation for the topic discussed
   • Anything else → use your best judgment to match the intent

2. Pull content from the conversation. Do NOT ask the user to re-explain. Everything you need is in the conversation history. Focus on:

   • Key decisions and their rationale
   • Technical details, code changes, file paths
   • Problems identified and solutions applied
   • Open questions or next steps

3. Write the file. Use the Write tool to save the markdown file.

   • Filename: YYYY-MM-DD-HHMMSS-<descriptor>.md. Get the timestamp from the shell — run date +%Y-%m-%d-%H%M%S — never guess it from context. Descriptor = format word plus the topic, kebab-case, ≤6 tokens; take the topic from the directive if it names one, otherwise from the conversation's dominant/most-recent thread (e.g. 2026-03-02-231500-questdb-wal-findings.md, 2026-03-02-231500-auth-refactor-pr.md).
   • Location: Always write to local/ at the root of where Claude was launched in (see Prerequisites & Output).

4. Tell the user the filename and a one-line summary of what you wrote. Keep it brief.

Quality Rules

• Be concise. Match the density the format calls for — a Jira ticket is tighter than a findings doc. Paste-destination formats (tickets, PR comments) should be paste-ready: roughly 40 lines or fewer.
• Never copy secret values (credentials, tokens, API keys, connection strings, private keys) from the conversation into the file; replace each with <REDACTED> or a reference to where it lives (e.g. ~/.claude/jira.env). Query results and technical details are fine — redact secret values only.
• Use the conversation's actual technical details, not generic placeholders.
• Do not pad with boilerplate or filler sections.
• If the directive is ambiguous about scope (e.g., "latest findings" but there were many topics), focus on the most recent/primary thread of discussion.
• Do not include meta-commentary about the skill itself in the output file.

Directive Patterns

• (empty) — If the directive is blank, treat it as bare summary: apply the scoping rules below, use summary as the filename descriptor, and state the chosen scope in the document's intro.
• Bare format name (findings, doc, runbook, ticket, notes) — scope to the conversation's substantive content. If there's a single dominant topic, use it. If the conversation is multi-topic, scope to the last 3-5 exchanges and state the scope choice in the document's intro — e.g. "Scoped to the most recent thread: database migration. Re-run with findings on auth refactor for the earlier one."
• Qualified directive (format word + topic phrase, e.g. findings on cache layer, PR description for auth refactor) — scope to the subset of the conversation that discusses the named topic; ignore unrelated tangents even if they were the most recent exchanges.

Prerequisites & Output

• Output dir = <launch-root>/local/. Determine the launch-root from the session context: in Claude Code it is stated verbatim in the <env> block as "Working directory" — use that value. If no such env info exists, fall back to the nearest ancestor of the current cwd containing .git/ and write to <that>/local/. Report the resolved absolute path in your reply.
• Create local/ (mkdir -p) if it's missing. On filename collision, append -2, -3, … to the descriptor until the name is free; never overwrite.
• If the launch-root is a git repo and git check-ignore -q local fails, note in your one-line reply that local/ is not gitignored so the user can decide.