AI Guides › Playbooks
By Nigel Guy · 7 min read
Most people use ChatGPT for a job search the way they use a search engine with manners: paste a CV, ask "what jobs suit me?", get a confident list, and start again next week with nothing carried over. It feels productive because the answers are fluent. It fails because a job search is a weekly process with a memory, and a chat window has neither a schedule nor a record of what you already decided.
ChatGPT Work is built for the other shape of problem. OpenAI describes it as an agent that can take action across your apps and files, stay with a project for hours, and run Scheduled Tasks. That is a different tool from a chatbot, so it needs a different way of being briefed.
The rule: give the agent a standing brief and a ledger, let it do the gathering and drafting every week, and keep every decision that touches another person for yourself.
OpenAI launched ChatGPT Work on 9 July 2026, alongside its GPT-5.6 model. According to the launch post, it can gather information across connected apps, produce finished sheets, slides, docs and web apps, and break long projects into steps. It follows the same usage structure as Codex, so heavier tasks use more of your plan's included usage. Check the plan page for the current £ price and limits before you commit; they change.
Things worth knowing before you build a routine:
I could not verify any claim that Work is a purpose-built "job-search" product. It is a general work agent; the job hunt is a good use of it, not a built-in feature.
A ledger is one spreadsheet or doc the agent owns and you read. It has four tabs, and every weekly run updates them.
| Tab | What it holds | Who edits it |
|---|---|---|
| Brief | Target roles, location, pay floor, deal-breakers, CV facts | You, rarely |
| Pipeline | One row per role: link, source, fit score, status, next step, date | Agent proposes, you decide status |
| Evidence | For each role, the lines of the advert that support the fit score | Agent |
| Drafts | Tailored CV points and messages, each marked "not sent" | Agent drafts, you send |
The point of the Evidence tab is that every fit score has a quote or link next to it. A score with no source is a guess.
Spend twenty minutes on this by hand. List the job titles you would accept, the commute or remote limit, the lowest salary you would take (in £), three deal-breakers, and the facts on your CV that are true and provable. Decide which employers are off limits, such as your current one.
Open Plugins in the sidebar and install only what the job needs: a spreadsheet or document home for the ledger, and, if you want it to read recruiter replies, your email. Resist connecting everything. Each connection is a door you then have to supervise. Your CV contains personal data, so check what you are comfortable uploading to a Library that other tasks can draw on.
You are a careful job-search assistant working for me, a [YOUR_PROFESSION] based in [YOUR_LOCATION].
Goal: build a Job-Hunt Ledger in [SPREADSHEET_NAME] with four tabs: Brief, Pipeline, Evidence, Drafts.
Inputs I will give you: my CV [ATTACHED_CV], target roles [TARGET_ROLES], pay floor in pounds [PAY_FLOOR], work pattern [REMOTE_HYBRID_ONSITE], deal-breakers [DEAL_BREAKERS], and employers to exclude [EXCLUDED_EMPLOYERS].
Steps:
1. Put my inputs into the Brief tab exactly as given. If anything is missing or ambiguous, ask me before continuing; do not guess.
2. From the CV, list only the facts that are stated in it. Do not add skills, dates or achievements I have not written.
3. Search [JOB_SOURCES] for roles that match the Brief. For each, add a Pipeline row with the link, employer, location, stated pay (or "not stated"), and date seen.
4. Score fit from 1 to 5 and, in the Evidence tab, quote the advert lines that justify the score. If you cannot quote evidence, mark the score "unverified".
5. Stop and show me the first ten rows for review.
Constraints: do not apply, send messages, create accounts or pay for anything. Ask for my approval before any action outside the ledger. Never evade a site's blocks or CAPTCHAs.
Before answering, check that every row has a link, a date and evidence, and list any rows that do not.
Fill in the square brackets. Read the first ten rows against the adverts yourself; this is where you find out whether the agent understands your Brief.
Once the first ten rows are right, ask Work to turn it into a Scheduled Task, for example every Monday morning. Standalone scheduled tasks start from a saved prompt each run, so the prompt must stand alone.
Every [DAY] at [TIME], open the Job-Hunt Ledger [SPREADSHEET_NAME] and read the Brief tab first.
1. Search [JOB_SOURCES] for roles posted since the last run date in the Pipeline tab.
2. Add new roles as rows, skipping duplicates and excluded employers.
3. Mark any existing role whose advert has closed or changed, with the date.
4. For the five highest-fit new roles, draft a short list of CV points that reuse only facts in the Brief tab, and a two-paragraph message. Put them in Drafts, marked "not sent".
5. Write a summary of no more than 150 words: new roles, closed roles, anything that needs my decision.
Do not send, apply or reply to anyone. If a source blocks you or asks for a login, note it in the summary and carry on. If the Brief is unclear, ask me instead of assuming.
Before finishing, check that no draft contains a claim that is not in the Brief tab.
Monday: the summary arrives. You spend about thirty minutes reading it, opening the five best adverts yourself, and changing Pipeline statuses to "apply", "park" or "reject". Midweek: you edit and send the drafts you chose, from your own account. Friday: you update the Brief tab if your view has shifted, for example after a recruiter call. The agent's job is the gathering and first drafts; yours is the judgement and every send button.
Skip anything that promises to "apply to hundreds of jobs for you". Skip paying for a bolt-on service when a spreadsheet and a weekly task do the job. Skip rewriting your whole CV every week; tailor only the top few. Skip connecting your main inbox until the ledger has worked for two weeks without surprises.