meta-job-search-pipeline

meta-job-search-pipeline

Self-improver persona meta-skill. Handles 4 modes via an llm_classify router — TAILOR_NEW (the default, pastes-JD-gets-application-pack flow), INTERVIEW_PREP, COMPARE_ROLES, and STATUS_DIGEST. Each mode unlocks only its relevant steps via when: conditions on the classifier output, so a single composition handles all four without forking the DAG into separate skills.

Composition philosophy — multi-skill bundled orchestration

This meta-skill uses only OpenSquilla-bundled atomic skills plus the five built-in step kinds — no external dependencies. The point of a meta-skill is to orchestrate multiple skills, so this DAG calls into 7 distinct bundled atomic skills, each at the right point in the pipeline:

Skill	Step(s)	Role in the DAG
multi-search-engine	web_research	Web research per target company
deep-research	deep_research	Extra-context round for INTERVIEW_PREP only
memory	recall_company, store_pack	Cross-session memory of past company research and prior application packs — recalled before web research, stored after deliverable
pptx	interview_deck	Generate an interview-prep slide deck when INTERVIEW_PREP and the user mentions "deck" / "slides" / "幻灯片"
xlsx	tracker_xlsx	Export the application ledger as a spreadsheet when STATUS_DIGEST
docx	export_docx	Optional final-deliverable export when the user picks EXPORT_DOCX: YES

Step kinds used: llm_chat, llm_classify, user_input, skill_exec, agent.

What got dropped from the original 17-step design

The original draft had two ClawHub-shaped skill_exec steps that turned out to be unnecessary:

• A skill_exec: lead-enrichment step that produced a structured company brief. The current enrich_company step (an llm_chat reading outputs.web_research) covers exactly the same contract with no external dependency. The original was the substitute path; it's now the primary path.
• A skill_exec: notion-api-skill step that POSTed the application pack to Notion. The deliverable is the markdown emitted by deliver_jobpack; the user copies it wherever they want. Convenience does not justify the dependency.

Honest limitations

• No application-ledger persistence. STATUS_DIGEST mode is paste-driven: the user pastes their current ledger every turn. Once the proposed state: primitive ships, the ledger can persist across turns automatically.
• No auto-apply. The skill produces text for the user to send; there is no LinkedIn / job-board posting integration. This is a deliberate read-only design.
• COMPARE_ROLES is text-based. Without a foreach primitive, the matrix is one llm_chat call with multiple roles in the same prompt; per-role isolation would need foreach.
• Interview prep depth. INTERVIEW_PREP mode runs one deep-research round; a multi-round interview-loop would benefit from cross-turn state.