| name | writing-contextual-todos |
| description | Use when creating docs/todo entries, follow-up work records, or post-validation TODOs that future humans or agents must understand without chat history |
Writing Contextual TODOs
Overview
Write TODO docs as handoff artifacts, not reminders. A future reader should understand what happened, why the follow-up matters, what is already proven, and what must be true before the TODO can close.
When to Use
Use this when asked to record follow-up work after validation, debugging, investigation, implementation, or a boundary discovery.
Do not use it for one-line personal reminders or issue titles that already have full context elsewhere.
Core Pattern
- State the reader standard before writing: understandable without chat context, necessary, concise.
- Gather facts: original goal, verified behavior, invalid evidence, boundary gaps, source-of-truth decisions, links to detailed evidence.
- Create or update
docs/todo/<date>-<topic>.md.
- Include acceptance standards that can fail, including negative cases such as wrong project, manual side effects, or missing observability.
- Have a subagent review against the reader standard and project boundaries.
- Patch the TODO from review feedback, then run a formatting/diff check.
TODO Shape
Use these sections unless the repo has a stronger local template:
Writing Principles: who the doc is for and what “good” means.
Context: the story in a few paragraphs, not a transcript.
What Was Verified: proven behavior only.
Boundary Problems: each gap with why it matters.
Follow-up Work: concrete implementation slices.
Acceptance Standard: observable pass/fail criteria.
Problems Encountered: traps future agents must not repeat.
Source Attribution
Separate product behavior from operator behavior.
- Product behavior: code/runtime/hooks/tools executed by the system being validated.
- Operator behavior: manual shell, direct API calls, manual GitHub/Linear edits, or cleanup.
Manual actions can prepare or inspect; they must not be counted as acceptance evidence.
Reviewer Prompt
Ask an independent subagent:
Review docs/todo/<file>. Check: can someone without chat context understand the goal, why it matters, what was proven, what remains, source attribution, architecture boundaries, actionability, and brevity? Return Critical / Important / Minor / Verdict. Do not edit files.
Common Mistakes
| Mistake | Fix |
|---|
| Vague TODO like “productize this” | Name the exact missing behavior and acceptance criteria. |
| No source of truth | Record where project/repo/issue identity comes from. |
| Manual actions counted as success | Mark them invalid or operator-only. |
| Narrative dump | Link detailed evidence; keep TODO focused on future work. |
| Skipping reviewer | Review catches context gaps the writer no longer sees. |