Onboard

/onboard: build the bank from the user's raw material

Turn a pile of raw sources into the structured bank the tailoring pipeline runs on. Code ingests, an agent structures, and you host the one conversational part: a short interview. The user's answers and their sources are the only places facts come from.

Voice. Describe what is happening to the user's career material in their terms: "I've read your three files", "your two resumes disagree on when you left Acme", "saved, your experience bank is ready". Give a file path where the user acts on it (where their bank lives, where to drop files). One short progress line per stage.

0. Cold-start guard

If bank/experience_bank.md already exists, stop: this command builds a bank where none exists, and running it against a populated bank would clobber curated work. Tell the user what exists and ask how they want to proceed (they can move the current bank aside first if a rebuild is really intended).

1. Resolve the sources

bank/sources/ is the drop box AND the permanent home: everything the user gives us lands there, forever, unedited. From $ARGUMENTS (and, if empty, the conversation):

• Path arguments (files or a folder) → expand a folder to the supported files inside it (pdf, md, txt, tex), then:

  python3 ops/bank_ingest.py bank <file ...>
  

• No arguments → ingest whatever is already in the drop box:

  python3 ops/bank_ingest.py bank
  

  Exit 4 with nothing ingested → the box is empty: tell the user their options (drop files into bank/sources/ and rerun, give paths, or paste their resume text right here).
• Pasted text (in the argument or earlier in this conversation) → Write it VERBATIM to bank/sources/pasted-<YYYY-MM-DD>.md (one copy, change nothing), then run the no-arguments form.

Invite more than one source: old resume versions, a CV, notes; the union holds more true material than any single file, and this is the moment to collect it.

Then read bank/.ingest/manifest.json and act on per-source status: low_yield, not_text, or extract_failed means that source is unusable as text; name the file, say what happened, and offer the paste route for it. Proceed when at least one source is usable.

2. Structure: spawn the extract agent

Spawn the bank-extract agent (Agent tool, subagent_type: "bank-extract") with a prompt giving the concrete paths: the bank directory (bank/), the manifest (bank/.ingest/manifest.json), and the format contract (spec/bank_format.md). It writes bank/experience_bank.md, bank/profile.json, and bank/.ingest/gaps.md, and returns a one-line summary. Relay that summary. Then log the operation:

## [YYYY-MM-DD] extract | <roles> roles, <units> units, <agenda size> gaps

appended to bank/log.md.

3. The interview: interactive, every question skippable

Read bank/.ingest/gaps.md. Select what to ask: every Tier-A item (record blockers), then Tier B, C, D in order until the total reaches 8; drop the rest. Lead with one plain sentence of what you found and why a few questions are worth their time (e.g. "Your resumes are in. A few things only you can answer would make the result stronger; skip anything freely.").

Ask via the AskUserQuestion tool, in batches of up to 4 questions per call, Tier-A first. The lead-in message explains the mechanics once: answers are typed into the Other field; the listed choices are ways to skip, so each question's text is the question alone. How to build each question:

• A conflict between the user's own sources (two dates, two titles): the options ARE the conflicting values, each labeled with where it came from ("August 2019 (your 2022 resume)"). The user clicks the right one, or types the real answer under Other.
• A public-record fact (an institution's location, a credential's issuing body): research it (model knowledge or a quick web search) and offer the candidates as options to confirm, each labeled with what it refers to ("San Marcos, CA (Cal State San Marcos)"). Offer candidates only when confident; otherwise ask openly. The recorded fact's source is the user's confirmation, never the research itself, which is also why ambiguity (two campuses with the name) is presented as options, not resolved silently.
• Every personal question (a scale, an outcome, a decision, a date of their own employment): the answer arrives through the built-in Other free-text field. The two listed options are the two ways to skip, "Skip: no such number/detail exists" and "Skip for now: could dig it up later", because they mean different things for the bank (a dead end vs. a note to revisit). Personal facts never carry suggested values: a proposed number, outcome, or tool name in an option is putting words in the user's mouth.

Rules for the answers:

• The user's answer is testimony: fold it in verbatim, their numbers, their words. An answer that hedges ("maybe 30 or so services") is recorded with its hedge.
• A record-layer answer (a date, a title, a location) updates BOTH bank/profile.json and the matching header line in bank/experience_bank.md; the two must round-trip.
• A narrative answer (scale, outcome, decision reasoning) is a targeted Edit to the unit the agenda item names: fill its thin field (Scale, Results, a new Design contributions entry). Everything else in the unit stays byte-for-byte.
• A skipped question stays a gap: leave the unit thin and the profile field null. Mark each item in gaps.md as answered, skipped (no such detail exists), or deferred (worth revisiting) so the file records what remains and why.
• One follow-up round is fine if an answer opens something clearly valuable; past that, stop; the bank grows later, this is not an interrogation.

Then log it: ## [YYYY-MM-DD] interview | <n> answered, <m> skipped appended to bank/log.md.

4. Application defaults: one optional offer

If the interview left profile.json without an application block (the sources may have volunteered one), make ONE offer: saving application-form defaults (work authorization / sponsorship, salary expectation, notice period, relocation, remote/hybrid/onsite preference) means application forms never have to re-ask. One AskUserQuestion batch at most: closed items (sponsorship, work preference) as options, free-text items via Other, recorded verbatim into the application block per spec/bank_format.md. Decline = move on; ask nothing item-by-item.

5. Integrity check

Run python3 ops/bank_lint.py bank. An ERROR is yours to fix before moving on: after a fold it is usually a header line that drifted from profile.json; correct the file and run the check again until no errors remain. Warnings are known gaps, already the user's choice (a skipped date, a unit with no stack): carry them into the closing summary in plain terms rather than fixing anything.

6. Index and close

Invoke the bank-map skill to build bank/evidence_map.md from the new bank. Then close with: where everything lives (bank/experience_bank.md, bank/profile.json, the sources in bank/sources/), what stayed thin (the skipped items in bank/.ingest/gaps.md, as things worth adding when the user has the numbers), and that /tailor <posting> is now ready to run.