Job search
Build a ChatGPT Project for Your Job Search: Keep Facts, Roles, and Drafts Organized
Organize reviewed candidate facts, role-specific requirements and draft versions in a ChatGPT project, then check each claim before using the result.
A useful ChatGPT project for your job search starts with a small, reviewed candidate fact library, then gives each role its own conversation and draft history. Keep confirmed experience separate from job requirements and unanswered questions. Before using any generated application material, check its claims against your records and the current posting.
Use the project to catch an outdated claim before it enters another application. You can browse recent roles on LandOffer.ai, then bring a selected posting into your own review process. LandOffer.ai publishes this guide; using either service does not establish that a role fits or that a draft is accurate.

Set up the project around evidence
Create a ChatGPT project for this search and give it a recognizable name, such as “Maya — backend search — October 2026.” In the project's sources, add only the candidate material you intend to use across roles. Put the drafting rules in project instructions. Start a separate chat inside the project for each selected vacancy.
OpenAI's current Projects and chats documentation describes shared uploaded files, instructions and connected sources. It distinguishes the project's Chats and Sources sections. It also says an ordinary ChatGPT project does not directly access a folder on your computer. Writing a local pathname into a prompt is therefore not a reliable way to provide its contents.
For this workflow, use deliberately supplied project sources and role-specific attachments. Local folder projects and Codex have different access arrangements. Do not assume instructions about those environments apply to a ChatGPT project on the web.
Use a readable source file with short headings and stable fact IDs. Keep a reviewed original outside ChatGPT as well. If an upload is unavailable or rejected, work from a smaller relevant excerpt and identify its version in the chat. Before retrying an archive, check which source names and passages the chat can actually identify.
Worked example: Maya's candidate fact library
Everything in both examples is fictional. No candidate records were uploaded and no ChatGPT account or project was configured for this guide. These are completed editorial examples, rather than observed model outputs or an accuracy benchmark.
Maya's prepared source is candidate-facts-v1 — reviewed October 22, 2026. Her review note says the following entries agree with her fictional resume, personal project notes and employment dates; those records are invented for this example.
| ID | Reviewed fact | Boundary for drafting |
|---|---|---|
| F01 | Backend engineer at Cedar Systems, June 2022–September 2026 | No invented promotion or current employment |
| F02 | Built Python services using PostgreSQL for an internal support workflow | Do not turn internal use into a public product |
| F03 | Owned a retry handler and its tests within a four-person team | Team delivery does not establish team leadership |
| F04 | Helped investigate two production incidents | No claim of owning the on-call program |
| F05 | Maintained a personal Docker-based deployment demo | Personal practice, not production Kubernetes experience |
| F06 | Seeks US remote backend roles and can work Pacific hours | Employer location eligibility still needs checking |
| U01 | No reviewed latency, revenue or cost-saving figure | Omit numerical impact until Maya supplies evidence |
| U02 | No reviewed Kubernetes work or management responsibility | Flag these gaps rather than borrowing posting language |
The library is intentionally selective. Maya leaves out her home address, reference contacts, private compensation documents and confidential incident details. The internal workflow description contains no customer names or employer code. Her contact information can be added later in the final document she reviews, without making it shared context for every drafting chat.
She also prepares these complete project instructions:
Help me prepare accurate job-search drafts. Use the candidate-facts version named in each request. Preserve employment dates, distinguish my work from team work, and cite fact IDs beside proposed claims in the review copy. Job descriptions describe employer requirements; they are not evidence of my experience. Mark missing information as unknown and ask a focused question when it changes the draft. Do not invent metrics, tools, titles or eligibility. Produce drafts for my review; do not submit applications or send messages.
These instructions express the desired behavior. They are not a technical guarantee that every answer follows them. Maya still checks each draft. OpenAI's prompting guidance supports naming the intended result, relevant sources and important boundaries, then reviewing the response.
Maya's first review request is concrete:
Read candidate-facts-v1. Before drafting anything, list F01–F06 and U01–U02, identify conflicting statements, and tell me which entries you could not locate. Do not treat an absent fact as false or supplied. Wait for my review of this inventory.
If the inventory drops U02 or changes “helped investigate” to “led incident response,” she corrects that discrepancy before requesting a resume paragraph. When reviewing the next draft, check those same ownership and unknown-information distinctions again.
Keep each role's requirements inside its own chat
Maya opens two chats within the same project. She chooses names that carry the role and outcome: R17 — Harbor API — resume review and R18 — Juniper Platform — gap review. These names are her convention, not special ChatGPT commands.
Each chat begins with a short role record: employer, role title, posting URL or supplied text, capture date, requirements and unresolved conditions. A posting can change or close. Preserve its captured version without describing it as permanently current.
Keep job descriptions out of the common candidate file so employer requirements stay separate from your experience. Otherwise, a requirement such as “manage five engineers” can become easy to mistake for something the candidate has done. The shared library should describe Maya; the role record should describe the employer's request.
The official project guide supports distinct chats for distinct outcomes. That organization helps retrieval, but it does not certify that every response retrieves every relevant item. Ask each chat to name its actual inputs before drafting.
Worked example: two role chats and one controlled correction
Maya's fictional R17 record captures a Python API Engineer posting on October 22. It asks for Python, PostgreSQL, service testing and Kubernetes familiarity. US remote eligibility is not specified in the supplied excerpt. The requested output is a short resume summary using candidate-facts-v1.
Her complete request is:
For R17 only, compare its four requirements with candidate-facts-v1. Draft a two-sentence summary supported by F01–F05. Keep Kubernetes as an experience gap, flag the missing location eligibility, and show the fact IDs in a separate review note. Do not add performance metrics. Label the result R17-summary-v1, awaiting candidate review.
Her prepared, bounded R17-summary-v1 reads:
Backend engineer with experience building Python services and PostgreSQL workflows for internal users. Owned a retry handler and its tests within a four-person engineering team, with additional experience helping investigate production incidents.
The review note maps the first sentence to F01–F02 and the second to F03–F04. Kubernetes and remote eligibility remain open questions. The summary does not imply that Maya satisfies every requirement.
For R18, the fictional Platform Engineer posting requests production Kubernetes operations, incident response and a Pacific-hours overlap. Maya captures it separately on October 22. Her prepared gap decision is: Pacific-hours preference aligns with F06; F04 provides limited incident-investigation evidence; production Kubernetes ownership is unsupported. She defers a tailored application and saves this question for her own research: “Does the role require immediate ownership of the cluster, or is there a supervised transition?”

Then Maya reviews her notes and corrects one candidate fact. F03 in candidate-facts-v2, reviewed October 23 in this fictional sequence, becomes: “Implemented tests for the retry handler; another engineer designed and owned the handler.” She records the change explicitly and supplies v2 as the current source. The earlier ownership sentence must no longer be reused.
Her revision request is:
Use candidate-facts-v2 for R17. Replace the F03 ownership claim in R17-summary-v1; preserve supported details from F01, F02 and F04. Return R17-summary-v2, a change note, and any affected R18 conclusions. Do not describe the earlier draft as approved or automatically rewrite unrelated material.
The completed R17-summary-v2 is:
Backend engineer with experience building Python services and PostgreSQL workflows for internal users. Implemented tests for a retry handler within a four-person engineering team and helped investigate production incidents.
Its change note says: “Removed handler ownership; F03 now supports testing contribution only. R18's Kubernetes gap remains unresolved. Recheck any other draft using F03 before reuse.”
Maya's resulting register makes the decision visible:
| Item | Current input | Saved output and review decision |
|---|---|---|
| R17 summary | Candidate facts v2; R17 captured October 22 | Summary v2 reviewed for these claims; v1 superseded |
| R18 gap review | Candidate facts v2; R18 captured October 22 | Application deferred; Kubernetes ownership unresolved |
| Shared facts | Candidate facts v2, reviewed October 23 | F03 corrected; earlier versions retained outside active inputs |
An offline check of this fictional example confirmed its fact IDs, version references and changed ownership wording. It does not demonstrate that ChatGPT detects a correction automatically. Maya remains responsible for finding the affected drafts and reviewing the revised text.
Make memory, limits and privacy explicit
Project organization and memory are different controls. OpenAI's Memories documentation distinguishes ChatGPT web memory from local Codex memory and points web users to Settings > Personalization. Do not rely on a promise that everything in every prior chat will be recalled, or that creating a project alone establishes a particular memory-isolation boundary.
For important work, name the current fact file and role record again. If the answer cannot identify them, supply the relevant text before continuing. “Use what you remember” is too vague when an old ownership claim has already been corrected.
No universal project file count, upload allowance or plan-specific memory configuration was verified for this guide. Check your account’s project upload controls, any displayed limit message and available memory options before organizing an archive or buying more access. Keep this workflow small enough to inspect even when available features differ.
Likewise, a project name is not a privacy setting. Review your account or workspace's applicable data, retention and sharing controls before adding personal material. This guide does not establish your account's training setting or who can access a shared workspace. Begin with a sanitized candidate summary and public posting text; leave confidential employer material and other people's personal information out.
Close the loop with a reviewed file
Keep a short register outside the conversation: role ID, source version, draft version, review date and next action. Distinguish “generated,” “candidate reviewed” and “submitted.” A polished paragraph remains a draft until you approve its factual content and check the employer's instructions.
If a conversation grows hard to follow, save a short handoff note before opening a fresh role chat. Include the current source version, the latest reviewed paragraph, rejected claims and open questions. Ask the new chat to confirm that inventory. Renaming a conversation or continuing it later does not itself prove that the earlier decisions remain in working context.
When producing a document, request a distinct filename and open the result. Check dates, contact details, links, page breaks and the final text against the reviewed version. OpenAI's file guidance recommends inspecting the generated file and identifying the specific passage or element when requesting a revision.
For your next selected role, browse recent roles on LandOffer.ai, save the posting's date and requirements, and start one project chat with your reviewed candidate-facts version. Before using that draft, trace each experience claim to a reviewed fact and choose the next question for any unresolved requirement.
Sources and Further Reading
- OpenAI: Projects and chats — web instructions, shared sources and folder-access boundary
- OpenAI: Prompting — useful sources, boundaries and review
- OpenAI: Memories — web and local memory controls
- OpenAI: Work with files — review and revise generated deliverables
Public documentation checked October 22, 2026. Examples, filenames and review conventions are original fictional teaching material. No authenticated project, upload, memory, privacy or plan-limit test was performed.



