ClaudeMoneyTaxesAI score · 65/100AI-assisted edit · no obvious prompt-injection riskX

Filing 7 Tax Returns in a Week with Claude Code: Agent Self-Review + GitHub Evidence Chain + Human Signature

Input: Multi-year, multi-state tax documents across multiple taxpayers → Process: Claude Code targeted self-review, GitHub audit trail → Output: Final returns ready for signature

ExperiencedStepsWorkflowTip
Who
US taxpayers with complex filing scenarios (multiple people, years, states), comfortable with code and GitHub, willing to deeply engage in review
You get
Completed 7 returns/amended returns in one week; agent error rate dropped to ~5 issues (1 material); author no longer hires human tax preparers
You need
Claude Code · GitHub

Constraints

Requires: Claude Code desktop app, GitHub repository, historical tax documents (W-2, 1099, prior returns), US tax basics. Author is a US taxpayer filing 1040 and schedules. No prompt needed — just follow the steps. High-risk output requires human verification; sign only after judging evidence sufficiency.

What they did

Author lastmjs, self-described as an 'agentic engineer,' is a US taxpayer with extremely rigorous tax review standards (historically finding 5-10 errors per professionally prepared return, 1-3 material). He experimented with AI-assisted tax preparation since 2022, but early models made basic arithmetic errors that destroyed trust. Around late 2025, GPT-5.2 Codex became reliable enough to complete a complex AAR (Administrative Adjustment Request) in 24 hours, which he signed. In March 2026, he used GPT-5.3/5.4 Codex for a full partnership return with complete PDF filling. In July 2026, a personal return. In the week of October 2026, using Opus 5.5 in the Claude Code desktop app, he completed 7 returns/amended returns across multiple people, years, and states.

How it went

Author reports: Opus 5.5 made only about 5 mistakes, with just 1 material and the rest extremely nitpicky. By late in the week, he trusted the agent enough to stop line-by-line review for most schedules, relying instead on the agent's multiple targeted self-reviews and judging their evidence. The agent also found errors from previous professional preparers. All work, decisions, and evidence were kept in GitHub repos; after signing, files were automatically organized, committed, pushed, tagged, and released with certified mail tracking. The author states he no longer plans to pay human preparers and calls this one of the most amazing technological experiences of his life. (author_report)

Worth knowing

【Historical progression】Author emphasizes 2022-2025 AI tax prep was nearly unusable due to arithmetic errors; GPT-5.2 Codex in late 2025 marked a reliability inflection point.【Not tax advice】This is a personal experience report, not professional tax advice; author is a US taxpayer dealing with federal and multi-state filings.【Signature liability】Author repeatedly stresses that his signature means legal accountability for quality, so human evidence judgment is non-negotiable.【Prior preparer error】October 2025: AI caught a preparer's failure to account for BBA (Bipartisan Budget Act) restrictions, which would have led to impermissible amended returns instead of required AARs.【Cost reference】~$40,000-$50,000 spent on human preparers over 6 years; partnership returns previously cost ~$1,000-$3,000.【GitHub automation】Repo auto-tracks certified mail receipt numbers and tags/releases final versions, but specific automation scripts are not public.【Model versions】Opus 5.5 used for this batch; historical uses of GPT-5.2/5.3/5.4 Codex noted.【Limitations】No public repo, no reproducible prompts, no technical implementation details disclosed; reproduction requires building your own GitHub workflow and possessing tax domain knowledge.

Steps

  1. Create a GitHub repo and organize tax documents

    Create a dedicated GitHub repository. Upload historical tax documents (W-2, 1099, prior returns, previous preparer files) organized by taxpayer, year, and state. Ensure Claude Code can access this repo as persistent context.

  2. Load documents in Claude Code and start filing

    Open the Claude Code desktop app pointing at the GitHub repo. Have Opus read all materials and begin processing each return or amended return. Explicitly state the multi-taxpayer, multi-year, multi-state scope.

  3. Guide the agent through multiple targeted self-reviews

    Actively direct Opus to perform multiple rounds of targeted, intense review across specific schedules and line items. Require the agent to provide evidence for key decisions rather than accepting conclusions blindly. The author later retained line-by-line spot checks for the 1040 main form, while delegating schedules to higher-level evidence review.

  4. Review evidence, sign, and auto-archive

    Human judges whether the agent's evidence is sufficient. Upon confirmation, sign. Send final documents back to the repo where automation handles: file organization → commit → push → track certified mail receipt numbers → tag and release the final verified version.

Prompt you can copy

Reference prompt · Editorial suggestion。Confirm the destination, input variables and output format before use.

This block is the prompt. Ads are never placed here.

No prompt needed — just follow the steps.

No prompt needed. The core skill is directing the agent's review focus and judging the quality of its evidence.

Source

用 Claude Code（Opus）一周完成 7 份报税/修正申报：决策与证据存 GitHub 仓库，让 Agent 多轮定向自审，本人判断证据后签字 · lastmjs · X · 2026-10-16T13:18:24.000Z

Comments & feedback

Tried this method? Share results, problems or improvements. Comments appear after moderation. Please avoid sensitive information.

Loading…

    Related