// HUMANTODO

Install as a Skill

Install HUMANTODO as an agent skill:

npx skills add humantodo/humantodo

No installation required

If you don’t want to install anything, just ask your agent to leave the code incomplete.

[YOUR PROMPT OR PLAN HERE]

Leave 20% of the solution incomplete, I want to manually complete the critical parts.

discuss onHN

More developers are adopting agents (LLMs that use tools in a loop to complete tasks) to generate code. Some claim it is no longer necessary to write code12345. Some say production stability is decreasing across the industry678. Some, well, just miss writing code910.

This resulted in widespread confusion about the role of the human1112. Developers can still write code without AI, but we are seeing a trend in the industry where AI generates the code, and humans are becoming the passive code reviewers1314. If we stop hand-writing code, how will models be trained on novel code and ideas1516?

Today, it seems that the role of the human is either higher-level orchestration171819, or whatever percentage left by the autonomous/agency level of their tools:

Claude CodeOpenCodeCodex

Opinionated Agents

50%

No matter which type of agents we choose, instead of letting them output 100% of the code2021, what if the agent leaves some of them blank for the human to complete22?

Like the TODO comments, we use HUMANTODO as a paradigm on how to make the human participation more active. By intentionally telling the agent to leave the gaps, we aim to solve:

ProblemDescription
Passive vs. Active Engagement23

Reviewing code is passive (easy to skim). Writing code is active (requires deep context).

Knowledge Retention2425

By forcing the human to bridge the gap between agent generated code, they must understand the why and flow behind the architecture.

Skill Preservation26

Engineers stay sharp because they are still solving the hard logic problems, while the agent handles the simple, boilerplate problems.

Reviewer’s Fatigue27

Sometimes, it is just easier to write your own code than to review the agent’s massive output.

For example, we can define within the prompts that the agent should leave 20% of the code incomplete. It can be any parts, maybe the most critical parts, maybe where we know the AI tends to be verbose, maybe it is a particular technology we want to refresh our memory on, or maybe halfway through our manual intervention, we realise it’s actually unnecessary complexity28293031.

Examples of different types of code for the LLM and human to complete:

ComponentAgentHuman
APIsAuth, headers, retries, error handling, etc.Map the complex business logic payload.
DatabaseCreate migrations and basic CRUD operations.

Write specific transaction queries or complex joins.

FrontendBuild the UI components and layout.

Handle the complex state management or transition logic.

SecurityWrite the middleware wrappers.Define the specific permission logic.

The goal of an autonomous future is efficiency and convenience. But this can be counterproductive to learning32, where it often has to be slow, manual, and controlled33.

Two people sitting down to build a piece of IKEA furniture

HUMANTODO is the IKEA effect, but for coding in the AI era.

FAQ

The agent can mock the test cases, and if the human does their part correctly, the system will work. This creates a contract-based development flow where the agent defines the interface and the human provides the implementation.

Tweet/Bluesky and share your thoughts.

Footnotes

  1. https://www.reddit.com/r/AskProgramming/comments/1jcwvwl/anthropics_ceo_says_that_in_3_to_6_months_ai_will/ ↩

  2. https://www.itpro.com/technology/artificial-intelligence/sundar-pichai-says-more-than-25-percent-of-googles-code-is-now-generated-by-ai-and-its-a-big-hint-at-the-future-of-software-development ↩

  3. https://www.cnbc.com/2025/04/29/satya-nadella-says-as-much-as-30percent-of-microsoft-code-is-written-by-ai.html ↩

  4. https://www.techspot.com/news/107411-microsoft-cto-predicts-ai-generate-95-percent-code.html ↩

  5. https://techcrunch.com/2026/02/12/spotify-says-its-best-developers-havent-written-a-line-of-code-since-december-thanks-to-ai/ ↩

  6. https://arstechnica.com/ai/2026/03/after-outages-amazon-to-make-senior-engineers-sign-off-on-ai-assisted-changes/ ↩

  7. https://mrshu.github.io/github-statuses/ ↩

  8. https://news.ycombinator.com/item?id=47229839 ↩

  9. https://nolanlawson.com/2026/02/07/we-mourn-our-craft/ ↩

  10. https://www.jernesto.com/articles/thinking_hard ↩

  11. https://youtu.be/pzkwn3hu1Cc ↩

  12. https://addyosmani.com/blog/next-two-years/ ↩

  13. https://leodemoura.github.io/blog/2026/02/28/when-ai-writes-the-worlds-software.html ↩

  14. https://youtu.be/5DP0az1q_8M?t=458 ↩

  15. https://x.com/naval/status/2028314493206585471 ↩

  16. https://antirez.com/news/157 ↩

  17. https://simonwillison.net/guides/agentic-engineering-patterns/ ↩

  18. https://blog.katanaquant.com/p/your-llm-doesnt-write-correct-code ↩

  19. https://boristane.com/blog/how-i-use-claude-code/ ↩

  20. https://acko.net/blog/the-l-in-llm-stands-for-lying/ ↩

  21. https://x.com/mntruell/status/2026736314272591924 ↩

  22. https://www.antifound.com/posts/codegen-is-not-productivity/ ↩

  23. https://www.marginalia.nu/log/a_132_ai_bores/ ↩

  24. https://www.rockoder.com/beyondthecode/cognitive-debt-when-velocity-exceeds-comprehension/ ↩

  25. https://www.media.mit.edu/publications/your-brain-on-chatgpt/ ↩

  26. https://addyosmani.com/blog/comprehension-debt/ ↩

  27. https://www.ivanturkovic.com/2026/02/25/ai-made-writing-code-easier-engineering-harder/ ↩

  28. https://news.ycombinator.com/item?id=36380711 ↩

  29. https://neilmadden.blog/2024/12/03/the-square-roots-of-all-evil/ ↩

  30. https://terriblesoftware.org/2026/03/03/nobody-gets-promoted-for-simplicity/ ↩

  31. https://youtu.be/ZJEnQOsMtsU ↩

  32. https://apenwarr.ca/log/20260316 ↩

  33. https://jamon.dev/night-shift ↩