| name | manual-vacancy-application |
| description | Intake and analyze one manually sourced vacancy, and prepare its Job Intelligence application package only after explicit user approval. Use when the user provides a raw vacancy, link, pasted job post, recruiter message, or company careers text; do not infer preparation consent from intake alone. |
Manual Vacancy Application
Obey AGENTS.md first. Never add or restore OpenAI Platform API calls, SDK usage, API keys, OpenAI-compatible model endpoints, or model-proxy code. This skill is a Codex-only orchestration wrapper over deterministic manual intake commands and $job-intelligence-workflow.
Goal
Read prompts/job-intelligence-workflow.md before starting. That file is the shared
repository contract used by this interactive URL flow and by Scheduled Tasks; this
skill owns the manual intake orchestration and vacancy isolation.
Turn one raw/manual vacancy into an analyzed registry entry, and into a prepared
application package only after an explicit user gateway decision:
The default preparation scope is the complete four-document package. If the user
explicitly asks for exactly one document, generate and publish only that document with
the matching --document value and preserve other existing artifacts.
- Publish the extracted vacancy with the deterministic
add-manual command.
- Analyze only the newly published vacancy when no acceptable match already exists.
- If and only if the user explicitly asks to prepare the vacancy, prepare the
application artifacts through
$job-intelligence-workflow, using
$write-cover-letter when cover-letter.md is selected, then publish the deterministic
DOCX-backed package.
Do not bypass deterministic project commands. Do not hand-edit published registry artifacts or cache metadata.
Workflow
- Run the one-time Git preflight in
prompts/job-intelligence-workflow.md before
reading vacancy or candidate evidence. Resolve a behind or diverged branch and
unexpected tracked changes in workflow output paths first; preserve unrelated user
changes. Then read config/codex-workflows.yaml and identify the configured model profiles,
reasoning levels, and prepare_min_score. Use the workflow default or an explicitly
selected --model-profile. You cannot switch the current task model from inside
the repo; if the active task is not using the selected profile, tell the user before
publishing model-dependent results.
- Extract fields from the supplied vacancy text or link, ask only for missing or
ambiguous required evidence, write one manual draft under
.codex-work/manual-job/,
and publish it with python run.py add-manual --input <draft.yaml>.
- Capture the newly published vacancy directory from the
add-manual command output.
Do not rediscover it through a queue or broad registry scan. From this point, keep
the workflow scoped to that single vacancy.
- If the vacancy has no current
match.yaml, analyze only this directory. Read its
vacancy files, the configured candidate evidence, and prompts/vacancy-match.md;
write one result draft under .codex-work/manual-analysis/, then publish it with
python run.py analyze <vacancy-directory> --input <draft.yaml> --workflow analyze --model-profile <selected-profile>.
Do not run triage, pending analyze all, analyze-batch, or any queue command for
this manual flow, and do not read unrelated vacancy directories.
- Treat the published match score only as an eligibility signal, never as consent:
- If score is below
prepare_min_score, do not prepare and report the reason.
- If score is at least
prepare_min_score but the user has not explicitly requested
preparation, stop after intake/analysis and report that user approval is required.
- If score is at least
prepare_min_score and the user explicitly requests
preparation for this vacancy, use workflow prepare.
- Prepare exactly this vacancy with
$job-intelligence-workflow preparation rules.
Use its two-wave orchestration for the default full package; for an explicit
single-document request, run only the necessary roles. In Wave 1, run research, CV/evidence, and
requirements/risks roles in parallel when subagent slots are available. Research
receives meta/job/company plus minimal candidate motivation hooks, not the full CV;
CV/evidence receives the vacancy and configured candidate sources, performs no web
research, and includes a complete proposed CV draft in evidence-map.md;
requirements/risks receives the vacancy and candidate evidence. Each writes only its
exclusive handoff under .codex-work/application/<vacancy-directory>/parts/:
research.md, evidence-map.md, or requirements-risks.md. No Wave 1 role may
publish, run a deterministic project command, or write a final artifact. The main
agent reconciles the handoffs, rejects unsupported claims, and writes the final
cv.md. Limit the overall preparation scope to this vacancy's meta.yaml, job.md,
optional company.md, configured candidate source files, and
prompts/vacancy-application.md; route only the subset assigned to each role.
Research in one pass using the posting plus at most two primary company sources
unless a critical eligibility or company-identity fact remains unresolved.
- For the default full package, start Wave 2 only after the final CV is fixed. Run cover-letter,
interview-preparation, and application-analysis roles in parallel when slots are
available, with exclusive ownership of
cover-letter.md,
interview-preparation.md, and analysis.md. Cover letter receives the vacancy,
final CV, verified research, and only required candidate evidence; interview receives
the vacancy, final CV, requirements/risks, and verified research without browsing
again; analysis receives the vacancy, final CV, and all Wave 1 handoffs. The
cover-letter role must invoke
$write-cover-letter when the selected scope includes a cover letter; stop if that skill is unavailable and never substitute the
retired inline drafting logic. If subagents or enough slots are unavailable, run the
same roles sequentially with the same handoffs, wave boundary, and file ownership.
No role may read the full registry, another vacancy, or inputs it does not need. Do
not claim a model switch inside the active task.
- The main agent performs one cross-file consistency and claim-grounding pass after
Wave 2. Then run the single combined deterministic draft check:
python run.py validate-application <vacancy-directory> --input .codex-work/application/<vacancy-directory> [--document <document>].
After it succeeds, publish once with python run.py prepare <vacancy-directory> --input .codex-work/application/<vacancy-directory> --workflow prepare --model-profile <selected-profile> [--document <document>].
If validation fails, fix only its cause and rerun the validator. If DOCX conversion
fails after validation, fix only that deterministic issue and retry publication.
- Confirm the application directory contains the selected Markdown artifacts, their
derived DOCX artifacts where applicable, and
manifest.yaml; verify unselected
existing artifacts remain unchanged.
- Regenerate the vacancy catalog through
$generate-vacancy-catalog, then run the
required tests and prohibited-API scan exactly once. Inspect the full diff, stage,
commit, and push when repository files changed. Repeat only a specific failed check
after correcting its cause.
Output Rules
Report the vacancy directory, match score, chosen workflow, generated artifact paths,
catalog result, commit hash, and push result. Never submit the application or contact
the employer.
