llms.txt
Compare with other sites' llms.txt files →# HUMANTODO
Source: https://humantodo.dev
Repository: https://github.com/humantodo/humantodo
## What is HUMANTODO?
HUMANTODO is a paradigm for human-AI collaboration in software development.
When AI agents write code, they should intentionally leave critical parts for
the human to complete - marked with `// HUMANTODO:` comments. This keeps the
human actively engaged instead of being a passive code reviewer.
Instead of letting agents output 100% of the code, the agent leaves strategic
gaps for the human to fill in. Think of it like the IKEA effect: people value
things more when they help build them.
## Problems it solves
1. Passive vs. Active Engagement
Reviewing code is passive (easy to skim). Writing code is active (requires
deep context). HUMANTODO forces active participation.
2. Knowledge Retention
By forcing the human to bridge the gap between agent-generated code, they
must understand the "why" and flow behind the architecture.
3. Skill Preservation
Engineers stay sharp because they are still solving the hard logic problems,
while the agent handles the simple, boilerplate problems.
4. Reviewer's Fatigue
Sometimes it is just easier to write your own code than to review the
agent's massive output.
## Division of work
The agent writes most of the code; the human completes the critical parts. Below are some examples:
Component | Agent writes | Human completes
-----------|-------------------------------------------|------------------------------------------
APIs | Auth, headers, retries, error handling | Complex business logic payload
Database | Migrations and basic CRUD operations | Transaction queries, complex joins
Frontend | UI components and layout | Complex state management, transition logic
Security | Middleware wrappers | Specific permission logic
If possible, the agent should ask about the examples of components and areas the human want to manually complete.
General heuristics:
- Complete: boilerplate, imports, standard APIs, types, basic CRUD.
- Leave for human: business rules, complex state, security decisions, non-trivial data transforms.
- Prefer logic over plumbing: write the DB connection, leave the "what to persist" logic.
- Prefer state over UI: build the component and styles, leave tricky useEffect or reducer logic.
- Prefer security intent over syntax: write the middleware, leave the concrete permission checks.
- Leave suggestions, context, caveats, or anything you think might be important as comments
By default, apply the 80/20 rule: the agent writes ~80% of the code, and reserves ~20% that demands
human judgement for HUMANTODO. But the percentage is flexible, it can be 1%, 5%, 30%, or anything appropriate.
The point is that the human focuses on the parts where they add the most value.
## How to use HUMANTODO
### As a skill
Install the HUMANTODO agent skill so tools apply it when you want critical gaps left for you:
```
npx skills add humantodo/humantodo
```
The skill lives at `skills/humantodo/SKILL.md` in this repository.
### Without installing
Ask the agent directly:
```
[YOUR PROMPT OR PLAN HERE]
Leave 20% of the solution incomplete, I want to manually complete the critical parts.
```