← sources

I haven't re-onboarded Claude Code after /clear in 10 days. Markdown/Issues memory rots because nothing enforces its lifecycle, so I built a CLI that refuses bad writes.

source: reddit (reddit:ClaudeCode)published 2026-09-21 07:05ingested 2026-09-22 02:40

retained text

20 years of web dev, first time posting something I built. For the last 10 days I've run my day job, and the development of this tool, through the tool itself.



The first thing that disappeared was the re-onboarding ritual after every /clear: "here's the issue we're on, the process is in this Markdown file, last time we finished this PR, next is X." I haven't typed that once in 10 days.



How Claude Code is used: as the project lead, not the typist. My Claude Code session usually acts as product owner. It triages requests, writes each requirement and its acceptance criteria into gy, hands implementation to a second agent (often opencode, with Codex as a second reviewer), then measures the result before recording it. I never run gy myself. gy ships as a CLI plus agent skills; the entry skill tells Claude Code to start from \`gy handover\` / \`gy next\`, and that is the whole onboarding. I /clear any of the agents whenever I feel like it.



The bigger relief: early decisions used to come back as late-project pain, and I'm usually the cause. I make a wrong call, or state a need that contradicts an earlier one.



Real example. Early on I wrote down, as a principle, that letting this tool affect how I run git had been a mistake. Four days later colleagues wanted team sharing, and the only zero-cost option was a git remote. Claude Code listed "this narrows your own principle" as a cost, quoted the sentence, and recorded the new decision as narrowing just that sentence: the record stays out of the project repo, a separate machine-only repo is allowed. The rest of the old decision stayed in force, and any fresh session that opens it sees that one sentence marked as retracted, with a pointer to what replaced it.



I don't watch how it finds these. What I see is the old sentence, the proposed change, and a decision to make.



What I learned, as two opinions:



1. Markdown wikis and GitHub Issues are too free. Labels, parent/child links, an LLM-maintained wiki: all structure, none of it enforces what is still a proposal, what was agreed, what was replaced, and what must exist before something counts as done. Without that it becomes debt, and correctness depends on the agent remembering the convention.



2. GitHub Issues is too slow a loop for agents. The overhead per decision is sized for humans.



So gy is a small CLI where agents record needs, questions, decisions, requirements and acceptance criteria, and invalid writes are refused when they are made:



```
$ gy criterion satisfy ac-a80e --evidence "tests pass"  
ac-a80e is not covered by an approved requirement (targets); req add "<title>" --need <N> --targets ac-a80e

```


Replacing a decision means quoting the passage that loses effect, and the quote is checked against the old text.



What it doesn't do: it checks the record, not your code, and it doesn't detect contradictions by itself. It forces the fix to be one sentence wide. I still have to decide goals and needs. That part didn't go away, it's just the only part left.



Local, Rust, open source, `cargo install gy`. Team sharing (a dedicated git repo) is experimental.



[https://github.com/aq2bq/gy](https://github.com/aq2bq/gy)



If you've kept a Markdown- or Issues-based memory alive across /clear for more than a few weeks: what broke first?

document fetched 2026-09-22 00:40 · canonical https://reddit.com/r/ClaudeCode/comments/1wm5iiy/i_havent_reonboarded_claude_code_after_clear_in