User Asks

Answering User Questions

Use when the user pastes a question from someone (Discord text, a GitHub issue, a #N reference) or asks you to help answer a person. Goal: a short reply they can paste back, grounded in a citable docs URL, plus a fix suggestion when docs or the API fall short.

Relay signal, easy to miss: third-person pronouns about the asker mean this skill applies, even with no explicit "someone asked me" framing or pasted quote. A message phrased as the user's own technical question can still be a relayed one — "they probably need X", "a user of mine is trying to Y", "I think there's 2 parts to their question." Any "they"/"someone"/"a user of mine" referring to a person other than the one you're talking to is enough on its own to classify this as a relay. Don't wait for an explicit "someone asked me" or a pasted quote before recognizing it.

1. Classify First

• Usage question ("how do I…", "does Gum support…", "why does X happen", "where is Y") → this skill handles it.
• Bug report, feature request, or concrete task → fall through to the normal issue-driven workflow in CLAUDE.local.md. (A usage question can still end in an issue — see step 4 — but that's a docs/API gap we surface, not a defect the user reported.)

For a GitHub issue, read both gh issue view <num> and gh issue view <num> --comments before deciding — comments often reclassify the ask.

Then decide the surface — tool or code. The docs split into the Gum Tool (docs/gum-tool/, the WYSIWYG editor — menus, properties, editor workflows) and Code (docs/code/, runtime C# APIs). The same question usually has a different answer in each, so don't assume. (Tool cues: "in the editor", a menu/property/tab name, a screenshot. Code cues: C#, "in code", a class or method name.)

• Obvious which surface → answer for that one only.
• Not obvious → treat it as possibly both: run two investigations (docs/gum-tool/ and docs/code/) and give two answers.
• Investigation is hard or hinges on what they meant → ask the user to clarify the surface (and specifics) before digging deep.

2. Where to Look: skills orient, docs cite, code confirms

These are roles, not a strict sequence — you'll often touch all three in one pass. The point is what each is for:

• Skills (.claude/skills/) orient — short, flag the gotchas, point you at the right doc page / source file. Skills are internal scaffolding — never link or paste a skill to a user.
• Docs (docs/) cite — the page you give the user. Use docs/SUMMARY.md as the index. This is what produces the URL.
• Code confirms — verifies the real behavior and locates where a doc should exist or where the API is wrong.

If a skill or code answers the "why" but no doc does, that's a docs-gap signal (outcome C below).

3. The Four Outcomes

4. Always Produce

1. A paste-able reply — write it like a quick message a busy human types in Discord, not a document. Lead with the answer, keep it to a few plain sentences, include the docs URL when one exists. Point to the Discord or a GitHub issue only when genuinely useful.
   • Tone: terse and a little lazy, not polished. The failure mode is sounding like AI. So: no bold, headings, or bullet lists in the reply — plain sentences only. No reassurance or warmth filler ("no worries", "you're doing fine", "great question", "so sorry"). No em dashes (they read as AI); use a comma, parentheses, or two sentences. Cut every word that isn't load-bearing — prefer the shortest phrasing that still says it. Always keep code formatting (inline code and fenced blocks).
     • Calibration example. Too wordy: "The EnableWindowsTargeting line you added is harmless, leave it if you want." Right: "You can keep the EnableWindowsTargeting line in still."
   • Deliver it as a file, not inline. Pasting markdown straight from the terminal into Discord/GitHub mangles the formatting. Instead: Write the reply to a .md file (use a temp path like C:\Users\vchel\AppData\Local\Temp\gum-reply-<slug>.md), then open it with Start-Process <path> (Windows start) so the user can copy clean source from their editor/viewer. Keep your terminal response to the working notes + a short ## Summary that links the file — don't duplicate the full reply inline.
2. Confidence + justification for any doc/API fix you propose (rubric below).
3. An issue — only for outcomes C/D, and only after the user agrees. Then gh issue create with a descriptive title (reference related issue numbers like #3163 when relevant). No template or label is required; labels are freeform and often omitted.

Confidence Rubric (for proposed fixes)

• High — must justify: cite the contradicting doc/code, the broken behavior, or the duplicated/missing content. Don't claim High without the evidence.
• Medium — plausible improvement, some judgment involved (wording, placement).
• Low — a hunch worth raising; flag the uncertainty.

Approval Gate

Propose first, act after sign-off. Do not edit docs/, change code, or run gh issue create until the user agrees. (Per the user's standing preference: doc fixes are proposed, then applied on approval.)

Building the Docs URL

Published base: https://docs.flatredball.com/gum/. The path mirrors the docs/ tree:

• docs/code/controls/listbox.md → https://docs.flatredball.com/gum/code/controls/listbox (drop .md)
• A folder's README.md → the folder path itself: docs/gum-tool/setup/README.md → .../gum-tool/setup
• Heading anchor = heading lowercased, spaces→-, punctuation stripped: ## Keyboard and Gamepad Navigation → #keyboard-and-gamepad-navigation

Full example: https://docs.flatredball.com/gum/code/controls/listbox#keyboard-and-gamepad-navigation

Pointers

• Doc-fix mechanics (SUMMARY.md, GitBook syntax, tone, tool-vs-code split) → gum-docs-writing. Don't restate them here.
• Issue-creation / branch / worktree / PR flow → CLAUDE.local.md.
• Domain skills (gum-layout, gum-forms-*, gum-runtime-*, etc.) are your first-pass index in step 2.