interview-for-personalizing
First step of bootstrapping a new workbench. Through a divergence-first, gap-based interview, draw out the user's persona (language, slug style, required issue elements, PR skeleton, label scheme, harvest fields, gate toggles) and draft it into a temporary .persona/ scratch that generate-workbench consumes. Use this whenever the user wants to set up, create, or personalize a workbench, or talks about "their own rules / conventions" for one — even if they don't say the word "interview". Precedes generate-workbench; together they are the bootstrap flow.
- License 许可
- MIT
- 最近更新
- 2026-07-18
- Stars
- ★ 0
- Forks
- 0
01 前言 Front Matter
- name
- interview-for-personalizing
- description
- First step of bootstrapping a new workbench. Through a divergence-first, gap-based interview, draw out the user's persona (language, slug style, required issue elements, PR skeleton, label scheme, harvest fields, gate toggles) and draft it into a temporary .persona/ scratch that generate-workbench consumes. Use this whenever the user wants to set up, create, or personalize a workbench, or talks about "their own rules / conventions" for one — even if they don't say the word "interview". Precedes generate-workbench; together they are the bootstrap flow.
03 SKILL.md 正文
interview-for-personalizing
Bootstrap step 1. Draw out the persona and draft it into a temporary .persona/.
The next step, generate-workbench, reads .persona/, bakes it into the user's repo,
then discards it.
What is persona? Only taste the machine does not enforce. The mechanism (disposable/accumulate, task meta rules, log format, docs wiki structure, …) is framework-fixed (
AGENTS.core.md), so don't ask about it. Ask only persona cells.
Where the defaults come from
This skill runs from the installed workbench-kit plugin, before a user workbench
exists — so don't look for defaults in the cwd. Read the bundled starting point at
${CLAUDE_PLUGIN_ROOT}/scaffold/AGENTS.overlay.md (Codex: the plugin's
scaffold/AGENTS.overlay.md). Show it as "here's the default — keep it?" and let the
user override only what they care about.
persona cells (= what to ask)
| Cell | What to ask | Default |
|---|---|---|
| language | language of operational output (prose, commits, issue/PR, your docs) | English |
| slug style | word count, case, separator | lowercase kebab, 2–4 words |
| required issue elements | what to enforce in a ticket | Background·Goal·Done·Out-of-scope |
| PR skeleton | PR body sections, AI callout | 5 sections + callout |
| label scheme | priority system, type/area use | priority only, human-confirmed |
| harvest fields | decision/lesson/runbook format | templates default |
| gate toggles | where to stop and ask | PR / cleanup / absorption all on |
language governs only operational output. The framework (skills, AGENTS.core) stays English; filenames/identifiers stay stable. Default is English; capture the user's choice (e.g.,
ko,ja).
Flow
The discipline is the same as ticket-incubate's: don't interrogate — diverge first, then ask only what's missing. Asking all 7 cells up front makes the user feel processed and produces shallow answers; letting them talk surfaces most cells for free.
-
Diverge. Learn how the user works (tools, taste, what their projects are). Show the default profile and ask "this is the default — good as-is, or anything off?"
-
Fast path — all defaults. If the user says "just use the defaults" (or clearly has no strong preferences), skip the interview: copy the defaults into
.persona/, confirm once, and hand off. Don't manufacture questions. -
Transition signal. Otherwise, once you can draft 2+ cells from the conversation, or the user signals "let's build it," move to the gap interview.
-
Gap interview — empty cells only, one at a time. Don't re-ask cells already surfaced; just confirm ("so, Korean for output — right?"). Skip cells the default covers. One question per turn.
-
Confirm. Show the assembled persona (the 7 cells, defaults + overrides) and get an explicit OK before writing — the user should see what they're about to generate from, just like ticket-incubate shows the ticket before creating it.
-
Write
.persona/(see schema/examples below), then hand off togenerate-workbench. Don't keep digging — details can be refined after the workbench exists.
.persona/ schema + examples (the contract with generate)
Output is temporary (dot rule — infra scratch; generate consumes then deletes it):
.persona/
overlay.md # filled AGENTS.overlay.md — same sections as scaffold default
codebases.yaml # (optional) initial work targets — may be omitted
meta.yaml # where/how generate should place the workbench
.persona/overlay.md — the scaffold default with the user's overrides applied.
Only the changed cells need differ; keep the rest as default. Example (a Korean-output
user who otherwise took defaults):
## Language
- Operational output language: **Korean** (ko). Framework stays English; filenames stable.
## slug style
- lowercase kebab, 2–4 words.
## Required issue elements
- Background · Goal · Done criteria · Out of scope.
... (remaining sections unchanged from default) ...
.persona/meta.yaml — tells generate where to create the workbench:
# exactly one of `repo` or `path`
repo: owner/my-workbench # create/use this GitHub repo
# path: ~/dev/my-workbench # or a local path
language: ko # mirrors overlay Language (generate reads it directly)
.persona/codebases.yaml (optional) — initial targets, same format as the
framework's codebases.yaml:
my-app: https://github.com/owner/my-app.git
Boundaries
- Don't ask about the mechanism (core) — only taste (overlay).
- Output goes only to the disposable
.persona/— no user repo is created yet (that's generate). - Leaving mid-way is fine — re-invoke to resume from defaults.
04 同仓库更多 Skills
ticket-incubate
Use to turn an idea that has no issue/ticket yet into an actionable GitHub issue. Grow a vague notion by divergence, narrow only the gaps via a deficit-based interview, and create the issue once the four required elements (background, goal, done criteria, out-of-scope) are filled. Triggers on "I have an idea", "let's make this a ticket", "create an issue", "interview me", or describing something to make/fix/improve without an issue number, even if the user never says the word "ticket". It is the front of the workbench lifecycle (idea -> ticket-incubate -> issue -> task-start). Distinct from brainstorming/design skills: those work out the how of what to build now; this captures the what of later work as a well-honed ticket.
task-tickets
Use to scan open workbench tickets joined with task branches and lifecycle events into unstarted / claim-attempted / in-progress / awaiting-PR / cleaned-up states. Read-only.
task-submit
Use when the user asks to submit, wrap up, or extract from a task, or when work has reached a point to open a PR.
task-status
Use to see local and remote task status at a glance and diagnose drift (unpushed, stale status, missed cleanup). When you want to know where everything stands before starting or submitting, or after a multi-device handoff.
task-start
Use when the user asks to start a task, pick up an issue, or begin a new attempt, or hands over an issue number to start work.
task-done
Use when the user asks to clean up or discard a task, or to clear the workspace after a PR is merged.