The req / reply communication method
One req / reply round
Below is the full content of this Rule in the kit, ready to use.
.claude/rules/req-reply-protocol.md Download# Rule: req / reply communication protocol
Purpose: keep "what the human says" and "what the AI answers" in two files, with line numbers, so they can be looked up later and do not vanish when a session is cleared.
## Files
- `req.md`: I write requests here (casual is fine, one per line).
- `reply.md`: you write answers here. Only APPEND new blocks at the end of the file; never rewrite, merge or delete old blocks.
## Trigger
When I say "req updated" or "after req line N": first read `req.md` and check it has enough lines (if not, say "the file may not be saved"), then answer.
## reply block format
```
══════════════════════════════
❓ [date time]
Original:
1. (paste the req text verbatim)
2. (paste the req text verbatim)
Reorganised questions (mapped to the original)
1. (original [1][2]) the reorganised question
───────────────────────────────
Answers
1. (original [1][2]) title
Answer (plain text, no tables or Markdown marks, because I read the raw text with line numbers)
Status: [1] ✅ [2] ⏳ waiting for confirmation
📌 req lines X-X → reply line XX ✅
```
- The original must be verbatim; "the gist" is not allowed.
- The reorganised-questions mapping must always be there. If one number has several sub-points, map and answer each one; never merge them and drop one.
- `reply line XX` is the line number of this block's `══` line; check it with a search after writing, never write "here".
- In the terminal output only: one line of confirmation + `✅ req lines X-X → reply line XX`.
The req / reply idea (T00's way of communicating)
The problem with an ordinary chat window: a long answer scrolls away and cannot be found the next day; when you ask several things at once the AI easily skips one; and long answers also eat the conversation's context. T00 writes both the "question" and the "answer" into files:
| File | Purpose |
|---|---|
temp01_req.md | The questions and requests you write; afterwards tell the AI "updated after line N" |
temp03_reply.md | The AI's answers, appended at the end only, each with time and line number |
temp04_pending.md | Open items waiting for your confirmation |
Three key rules:
- Copy the original text verbatim into the answer: later, looking only at the answer file, you still know what was asked.
- There must be a "reorganised question ↔ original" mapping table: when one number holds several small points, map and answer each one, so that "it looks like one thing" does not cause a merge and a missed answer.
- Answers are append-only; old ones are never changed: history is not overwritten; the terminal only prints "updated (line N)" and does not repeat the content.
Benefits: long answers do not use conversation context; every entry has a line number for lookup; "missed answers" can be found with the mapping table. Cost: one extra round trip, so it suits projects with many requests and a long life, not one-sentence questions.