TODO Workflow

Each project keeps its tasks in a root TODO.md that its owner shares with the coding agents. The /next-task command works the file and the todo-md skill defines its format, in Claude Code, OpenCode and VS Code Copilot alike (see Skills for how both reach each tool).

Piece	Source	Role
todo-md skill	config/skills/todo-md/SKILL.md	Format of TODO.md, applied to every edit of the file
/next-task command	config/commands/next-task.md	Picks, implements, verifies and records tasks

Sections

Section	Holds	Worked by
🔴 In progress	The task an agent is on, or stopped on (BLOCKED)	/next-task, first
🟡 To do	Approved tasks, most important first	/next-task
🖐 Manual — do it yourself, in order	Steps only the owner can take, each with its success signal	the owner
🟢 Ideas / backlog	Ideas not approved yet	nobody
🤖 Agent recommendations	Agent suggestions waiting for triage	the owner triages
🔒 Security audit	Agent security findings waiting for triage	the owner triages
✅ Done	Finished tasks and steps, dated, newest first	—

A task reads - [ ] <TYPE>: <description>, where <TYPE> is the upper-cased Conventional Commit type it will ship under (FEAT, FIX, CHORE, …, never REFACTOR). The owner approves an item by moving it to 🟡 To do, or by naming it to /next-task; agents never approve one on their own.

Running tasks

Invocation	Effect
/next-task	Works the next task, then stops
/next-task all	Works every open task of 🔴 and 🟡, one at a time
/next-task 3	The same, for the next three tasks
/next-task <words>	Works the open item matching the words, outside 🖐 Manual

For each task, the agent:

1. moves it to 🔴 In progress, so an interrupted run leaves a trace;
2. implements it, with its tests and the docs it makes stale;
3. runs the project's checks — a check it cannot run becomes a 🖐 Manual step, never a pass;
4. moves it to ✅ Done with the date and how it was verified;
5. proposes a Conventional Commit message and the files to stage, without committing.

A task it cannot finish stays in 🔴 In progress, marked BLOCKED YYYY-MM-DD: <what is needed>, and the agent stops to ask. A batch stops at its first blocker; later runs skip a BLOCKED task until the owner removes the marker or names the task.

After the task — once at the end of a batch — the agent adds at most about five grounded recommendations and five security findings for triage. A batch ends with a commit plan: one commit per task, in order.

English only

Everything written to TODO.md is in English, whatever the language of the conversation, and so are commit messages (commit-types skill). A TODO.md still in the former French layout (🔴 En cours, 🟡 À faire, ✅ Fait, …) is migrated on its first edit — headings renamed, content translated — as a change of its own.