| name | job-intelligence-workflow |
| description | Run the repository's vacancy collection, match analysis, vacancy-status management, and vacancy-specific application preparation without project code calling the OpenAI Platform API. Use for Job Intelligence scheduled runs, processing pending vacancies, publishing Codex-produced drafts, manually changing vacancy status, and orchestrating the mandatory final independent vacancy-catalog generation. |
Job Intelligence Workflow
Obey AGENTS.md first. Never call the OpenAI Platform API from repository code. Use the model already selected for the active Codex task and never claim to have switched models from inside the task.
Choose one mode
Before choosing a mode, read prompts/job-intelligence-workflow.md. It is the shared
execution contract for interactive and scheduled launchers; this skill supplies the
tool sequencing and safety boundaries around that contract.
Before reading vacancy or candidate evidence, researching, or drafting, perform the
one-time Git preflight from the shared contract. Resolve a behind or diverged branch
and any unexpected tracked changes in workflow output paths before model-dependent
work. Preserve unrelated user changes. Do not repeat the fetch during the same run.
- For collection, run
python run.py all, regenerate the registry index, and report source failures separately.
- For analysis, follow
references/analyze.md and process a sealed batch of up to 15
pending vacancies.
- For a manually supplied vacancy, use
$manual-vacancy-application. Analyze its newly
published registry directory directly; do not route that one-vacancy run through
triage, pending analyze all, or the scheduled sealed queue.
- For application packages, follow
references/prepare.md with workflow prepare and
process one to 10 fresh vacancies explicitly selected by the user through vacancy IDs
or registry directories. Generate the full package by default; when the user
explicitly names exactly one document, use the matching --document mode. Use
$write-cover-letter for every selected letter and never expand the selection to
all.
- For a user-requested status change, run
python run.py status <vacancy-id-or-directory> <status>. Never change status without an explicit user request. Preserve the complete history through the command.
Read config/codex-workflows.yaml before model-dependent work. Select the workflow's
default model profile unless the launcher explicitly supplies --model-profile; the
selected Scheduled Task or chat must use the corresponding model and reasoning level.
Pass the workflow and optional model profile to the deterministic publisher; it derives
the only allowed model label from policy. If the requested model is unavailable in the
current Codex surface, tell the user and do not publish under that profile.
Common rules
- For preparation, read only the explicitly selected batch, the configured candidate
source files, and the relevant prompt. Handle one selected vacancy at a time and do
not compare vacancies or reuse vacancy-specific research, keywords, or draft content.
For scheduled analysis, read only the sealed input pack and the batch prompt. For a
manual vacancy, read only that selected directory, its configured candidate evidence,
and
prompts/vacancy-match.md. In either mode, do not compare vacancies or read
another vacancy's artifacts.
- Write model-produced drafts only under
.codex-work/; the directory is ignored by Git.
- When a cover letter is selected, require
$write-cover-letter from the highest installed version of
agent-plugins@valentin-agent-plugins available in the active task during
preparation. Stop if the active task cannot load it; do not recreate the retired
inline drafting flow.
- Publish through
run.py so schema validation, hashes, atomic writes, DOCX conversion, and cache metadata remain deterministic.
- For default full-package preparation, use the two-wave orchestration and exclusive
file ownership in
references/prepare.md. For an explicit single-document request,
run only the roles and handoffs required by that document. The main agent alone
finalizes selected drafts, performs the applicable claim check, validates, and
publishes.
- After the selected drafts for a vacancy are complete, run
python run.py validate-application <job-directory-or-vacancy-id> --input <draft-directory> [--document <document>] once as the prepublication check. Omit
--document for the default full package. Publish only after it succeeds. If
validation fails, fix only its cause and rerun the validator. Do not edit generated
cache metadata by hand.
- Never submit applications or contact employers.
Mandatory final catalog step
After every successful collection, analysis, preparation, or manual status change, use $generate-vacancy-catalog and run its deterministic command as a separate operating-system process. Include its result in the final report. Do not import or call the catalog generator in-process.
Mandatory Git finalization
After the catalog process, run the relevant tests and API-prohibition scan exactly once,
then inspect the full diff, stage all added, changed, and deleted project files with
git add -A, commit once, and push the current branch to origin. Repeat only the
specific failed check after correcting its cause; do not rerun the entire workflow or
full check suite without a failure. For a Codex-authored commit, inspect the staged diff
and write a natural, human-written subject that names the run's actual result, using a
useful count or vacancy context when relevant. Do not reuse or randomly select from the
GitHub Actions templates, and do not use a generic update data, update files, workflow changes, or automated update subject. Never stage ignored secrets or local work
files. If the tree is unchanged, skip the commit and push. End the report with a
changelog derived from the commit plus the commit hash and push result. Do not open a
pull request unless explicitly requested.
