career-hub-build

career-hub-build — Interview-Driven Enrichment

Enrich a career hub via conversational interview, one category at a time.

When Claude Invokes This

Typical prompts:

• "build roles" / "build projects" / "build capabilities" / "build foundations"
• "fill out my role history"
• "let's update my career hub"

Design Philosophy

Hub-first, rendering-second. The hub is the source of truth; CVs, LinkedIn, and GitHub profile pages are views onto it. Every public claim traces to an entry in the hub.

Draft-from-known-content, ask-only-gap-questions. When build enriches a stub, Claude drafts from existing hub content and cross-references, and asks the user only gap questions — never redundant questions. Fast, low-friction enrichment.

Every claim verifiable. No unverifiable adjectives ("expert", "pioneered"). Scale numbers and verification methods live in a.foundations/biographical-facts.md.

Prerequisites

• cwd is a compliant hub (see ../references/hub-detection.md). Refuse otherwise with a pointer to career-hub-init.

Behaviour

Step 1 — Confirm hub

Hub detected at <cwd>.
- a.foundations: <4 files, X complete / Y stubs>
- b.history: <N roles, X active / Y draft>
- c.projects: <N projects, X active / Y draft>
- d.capabilities: <N cards, X active / Y skeleton>
- g.backlog: <N open threads>

Proceed with build <arg>? (y/n)

On explicit y, proceed.

Step 2 — Select target unit(s)

If invoked with a category argument (build roles, build projects, build capabilities, build foundations):

• Enumerate files in that category directory.
• Ask the user the order: chronological, reverse-chronological, user-picks, auto-next-stub (next status: draft file in chronological order).

If invoked bare (build):

• Present a menu:

  Which category?
  1. foundations (4 files, 2 stubs)
  2. roles (12 files, 5 stubs)
  3. projects (9 files, 3 stubs)
  4. capabilities (6 skeletons)
  

• Wait for selection, then continue as above.

Step 3 — For each unit, the draft-then-gap-questions loop

For each file selected:

1. Read existing content plus frontmatter, plus cross-references (linked roles, linked projects, linked capabilities, a.foundations/biographical-facts.md).
2. Draft a concise proposed fill-in for each of the category's body sections (Scope, Outcomes, Verification, Related, etc.).
3. Ask gap questions only — never ask what is already known or derivable from cross-references. Target 2-5 gap questions per unit for a quick pass.
4. On user response, edit the file in place:
   • Bump last_updated in frontmatter.
   • Flip status: draft → active if the unit is now substantively complete.
   • Cross-link to related files using relative paths.
5. Offer confidential sibling if candid content surfaces: "This includes sensitive political / failure / proprietary content. Create a <filename>.private.md sibling for candid notes?" — on yes, create the sibling with the confidential template.
6. Log deferrals — if a claim needs external info (reference name to confirm, conference event name, etc.), create a g.backlog/<slug>.md entry per the backlog item template, and register it in g.backlog/README.md index.
7. Handle splits — if the user flags that a single file should be split (e.g. a founder role covering two concurrent ventures), create the split files and update cross-references + README.md entries.

Step 4 — Update downstream

After editing any role / project / capability file, also update:

• README.md — if the file's row caption changed (scope, active/draft status, headline).
• a.foundations/biographical-facts.md — if a numerical claim, tenure, or scope has changed.
• Cross-linked files — if a name, slug, or date changed.

Step 5 — Commit at natural breakpoints

One commit per completed unit, OR one commit per category pass — operator's choice, default to one commit per unit.

Commit message format:

content(<category>): enrich <slug> — <brief summary>

Example:

content(b.history): enrich 2020-2021-nab-ad-optimisation — scope, outcomes, team structure

Interview Posture

Draft from known content. Ask only gap questions.

The skill's differentiator. Don't ask what you can infer. Don't ask what cross-references already state. Don't run a fixed question list.

For each unit, after reading, your first message to the user is:

**<file name>** — draft from known content:

[short proposed fill-in for each body section, marked where gaps exist]

Gap questions:
1. <specific thing you can't infer>
2. <specific thing you can't infer>
3. <specific thing you can't infer>

Or skip ahead and answer in one message if easy.

Keep gap questions to 3-5 per unit on a quick pass. Offer deep-dive mode on request for high-CV-weight items.

Mode: Quick vs Deep

Default is quick pass — fast gap-filling, target complete unit in one exchange where possible.

Deep dive on request — more thorough probing, additional questions about politics / proprietary details (offer .private.md sibling), follow-up on stories that would make good case studies.

Refuses

• cwd not a hub (per hub-detection.md).
• Unit already status: active unless user explicitly asks to re-interview.

Does Not

• Render to external artefacts (that is publish's job).
• Enforce a fixed question list.
• Re-interview active files without explicit ask.
• Populate categories it wasn't asked to populate.