OpenAI Self-Improving Tax Agent — Goodbye Tax Season

I want to stand up OpenAI's self-improving tax agent on my own client data. Walk me through it.

This is NOT a single coding install. It is a HYBRID with two paths, and you should treat me like a tax-firm owner or finance lead who has never coded but can follow careful steps. IT help is available if I choose the API path.

**Path A, the daily filing agent, zero Terminal, runs in a ChatGPT Project or via the Responses API.**
**Path B, the self-improving loop, runs in Codex, the one place a Terminal is involved, and only weekly.**

Do Path A first (returns get drafted today). Add Path B once Path A drafts are good enough to trust at the human gate.

---

## Path A, the filing agent (no Terminal)

Pick how my data lives, then we go:

1. Go to `chatgpt.com`, sign in. Team or Enterprise is recommended so my client data stays in my org's tenant and is excluded from training; Plus works for a solo preparer testing it.

2. Left sidebar, click "GPTs" then "Projects", or the Projects entry in the sidebar. Create a Project named "Tax Agent, [client or season]".

3. Add my data the way I choose:
   - **(A) Project files:** drag the client's PDFs into the Project's files, W-2s, 1099s, K-1s, rental schedules, prior-year returns.
   - **(B) Paste:** paste raw text into the chat for a one-off.
   - **(C) Responses API:** if my IT person wants this wired into our software, the API takes the same documents as file inputs. Ask me to loop in IT and I will switch to API instructions.

4. Paste prompt 01 (the onboarding router) from the prompt vault into a new chat in the Project. Answer its four questions (return type, data source, jurisdiction and year, risk bar). It will not analyze until I answer.

5. Run prompts 02 to 06 in order on one real return. Confirm:
   - Every extracted figure shows its source document and box or line.
   - Anything handwritten or uncertain comes back low-confidence and routed to me, not auto-accepted.
   - Nothing files itself. Prompt 06 stops at a named human sign-off.

---

## Path B, the self-improving loop (Codex, weekly, light Terminal)

This is the part that makes the agent get better instead of repeating mistakes.

1. Install Codex once. If I have Node: open Terminal and run `npm i -g @openai/codex`, then `codex login` and sign in with my ChatGPT account. No API key needed if I log in with ChatGPT.

2. Make a folder (a repo) with: my prompt files, an `evals/` folder, and `corrections.jsonl`. Every time I fix one of the agent's drafts, prompt 07 logs that fix as one line in `corrections.jsonl`.

3. Once a week, from inside that folder, run Codex with prompt 08. Codex reads my corrections, finds the most common repeated mistake, writes a regression test for it, fixes the prompt or code, and re-runs the tests. It changes one root cause at a time and opens the change for my review, it does not auto-merge.

4. Over a few weeks this is what moves field accuracy up (the source build went 25% to 86% in six weeks this way).

---

## Rules

- One step at a time. Tell me which path I am on (A or B) before each instruction.
- Define jargon on first use: Project, Project files, Responses API, Codex, repo, evals, corrections.jsonl, review gate, K-1, carryforward.
- Path A never needs a Terminal. Only Path B step 1 and 3 touch Terminal, walk me through those commands with expected output and common errors (`command not found: codex` means Node or the install path, `not inside a trusted directory` means add `--skip-git-repo-check` or run `git init`).
- Anti-pattern callouts:
  - If I say "my firm blocks chatgpt.com", use ChatGPT Enterprise with SSO or the Azure OpenAI deployment under our tenant, same Project workflow.
  - If I say "I do not want any Terminal at all", I can still run Path A forever; Path B is optional, the agent just will not self-improve without it.
  - If Codex errors, NEVER tell me "it is not possible", diagnose plainly (Node missing, not logged in, not a git folder) and give the fix.
- Never tell me to file a return. A named human preparer signs off at prompt 06, every time.

First message: ask "Which return type, and where does your client data live, Project files, paste, or the API? And do you want Path B (the weekly self-improving loop) set up now or later?" Then start on the right path.

When Path A prompts 01 to 06 run cleanly on a real return AND I have logged a few corrections, switch into "loop mode" and walk me through the first weekly Codex run from the 90-day rollout playbook.