Engineering Dialogue Skill Mining

Engineering Dialogue Skill Mining

Use this skill to turn prior conversations into reusable engineering-operation skills. The goal is not to dump chat history; it is to extract repeatable patterns with evidence, then write concise skills that another Codex instance can actually use.

Scope And Privacy

• Read only locations the user authorizes or the current environment permits.
• Start with the current workspace, then Codex session indexes such as .codex/session_index.jsonl, sessions, and archived_sessions when authorized.
• Extract user requests, assistant final answers, tool calls, and command evidence.
• Do not include unrelated private chat content, account secrets, license keys, or full irrelevant transcripts.
• Preserve paths and thread IDs in an audit report when useful, but keep skills themselves free of personal details.

Search Workflow

1. Build keyword groups.

   • Software: CAD, AutoCAD, DWG, DXF, SolidWorks, SLDPRT, SLDASM, Ansys, Fluent, Workbench, MATLAB, Revit, BIM, COMSOL.
   • Engineering work: cost estimating, quantity takeoff, drawings, BOQ, budget, rendering, model, modeling.
   • Chinese query equivalents: gongcheng zaojia, suanliang, tuzhi, qingdan, yusuan, xiaoguotu, xuanran, moxing, jianmo.
   • Operations: open, export, convert, save, inspect, run, command, operation, desktop, computer-use.

2. Search efficiently.

   • Use rg --files and rg -n for text files.
   • Exclude huge Base64/image/cache files that drown real evidence.
   • Parse session indexes to find thread titles and JSONL paths.
   • For each candidate thread, extract only relevant user_message, assistant process/final messages, and function/exec calls.

3. Classify threads.

   • Keep threads that include real engineering-software operations or reusable procedural lessons.
   • Drop false positives such as generic PDF/report generation unless they involve CAD/engineering software workflows.
   • Group by software and operation type.

Skill Extraction Heuristics

Mark a workflow as worth turning into a skill when it has at least two of:

• It solved a real local software/file problem.
• It used a robust discovery sequence.
• It corrected a previous overclaim or failure mode.
• It produced a verifiable artifact.
• It contains menus, commands, file extensions, or validation checks likely to recur.
• It reveals a safety boundary, such as legal activation or not claiming unverified CAD output.

Output Structure

Produce:

• An audit Markdown report with thread ID, title, path, user request summary, key operations, final result, and reusable lesson.
• A shortlist of candidate skills, each with trigger conditions and the effective operation pattern.
• Separate skill folders when the user asks to formalize them.

Writing Skills From Dialogues

When creating skills:

• Keep frontmatter descriptions broad enough to trigger on Chinese and English requests.
• Put "when to use" information in the YAML description.
• Keep the body procedural and short.
• Remove personal paths, chat fragments, and one-off project names unless they are needed as examples.
• Include validation and final-response shape so future agents do not repeat vague or unverified delivery.

Final Response Shape

Return:

• Number of threads reviewed and kept.
• Skills created or updated.
• Where the skill folders are saved.
• Any important limitations, such as garbled encoding or missing session files.