Clearstack
Claude Code skill packs. No bloat.
Plain-text memory & continuity. Stop re-explaining every session.
€29 · one-time
It was told nothing. This is what it said.
One question, typed into a project the session had never seen. No setup, no follow-up, no second attempt. The clip is an excerpt because the full answer does not fit on a screen. All of it is below.
Nobody asked it about any of this:- It warns you off a change to read_rows that has already been made and reverted twice, because a previous session wrote that down.
- It will not say the tests pass. It tried to run them, was blocked for approval, and reported that instead of guessing.
- It finds a line in the project's own notes that contradicts the code, and says which one wins.
- It picks up the task left open two days earlier.That is the product. Three markdown files your agent writes and reads.
You do not have to trust the clip. Install, run /memory-init once, do a normal session's work, then open a fresh session and ask it:What is this project, how do I run the tests, and what should I do next?If it cannot answer, you have lost ten minutes, and you found that out in one session rather than after a month of using it.
What is in the zip
Nine files, about 14 KB, all plain markdown you can read before you run anything.- catchup cold-boots a session from your memory files
- checkpoint records what a future session needs to know
- safe-compact saves state before your context gets compacted
- memory-init one-time setup, reads your repo and writes the files
- decide locks a decision and the reasoning behind it
- prune removes stale claims, with your approvalPlus a boot file, an install guide and the licence.Requires Claude Code. If you are not using Claude Code, this does nothing for you.
Try two of the six skills first, free./plugin marketplace add clearstackhq/clearstack-lite
/plugin install memory-continuity-lite@clearstackGet the free two-skill pack on GitHub
€29 · one-time
What it does not do
- It does not make the model smarter and it does not write your code.
- It is not a vector database and it is not agent orchestration. It is six skills and three markdown files.
- On a big codebase the memory file grows past one screen, which is the exact problem it exists to prevent. That is why /prune ships with it.
- The clip above is a real captured session replayed at reading speed, not a screen recording. We say so on the clip itself.
- The licence is one per person, while the design is team friendly because memory/ commits to git. That is a real tension and we would rather name it.
The full answer, unedited. Nothing cut, nothing reworded, first and only run.
Booted from memory/. I read PROJECT.md, DECISIONS.md, and the whole of SESSIONS.md (one entry), then checked each claim against the files.The brief- What this is — csvpeek, a stdlib-only CLI that prints a CSV's row count, column count, and filled-vs-blank tally per column. csvpeek/reader.py parses and counts, csvpeek/cli.py prints. No dependencies, no packaging config, not a git repo. Confirmed against all three source files.
- Where the last session stopped — 2026-07-30, a memory-init run. It created the memory files and read the project into PROJECT.md. It never ran the suite.
- The next open action — run python -m unittest discover -s tests and record the result. Still open: I attempted it just now and the command was blocked for approval, the same wall the last session hit. So "the tests pass" remains unverified for the second session running.
- The trap touching that action — read_rows reads bytes and strips the UTF-8 BOM by hand. That is locked in DECISIONS.md and has been "simplified" to a plain open() twice. test_bom_does_not_corrupt_the_first_header is the only guard, and it only guards if someone actually runs the suite — which is exactly what has not happened.
- Note that disagreed with the code — one. PROJECT.md:24 says "For CI the suite runs under pytest tests/ -q." There is no CI in this project: no .github/, no .gitlab-ci.yml, no pyproject.toml/setup.cfg/tox.ini, and it is not a git repository. The command itself would work (the test file inserts the project root on sys.path at import), but the "for CI" framing describes something that does not exist here.How to run the tests, from the project root:python -m unittest discover -s testsThat matches tests/test_reader.py:5's own docstring. Four tests, plain unittest, nothing to install. Note the suite covers reader.py only — cli.py has no tests at all.What I'd do next, in order:1. Approve the test run. Everything else is guesswork until the suite is green — and the BOM regression guard is unexercised. Re-run the command above and let it through, then checkpoint the result.
2. Fix README.md. Both defects are still there, verified a moment ago: line 9 gives python csvpeek/cli.py data.csv, which cli.py's own module docstring says cannot work, and line 14's Tests section is the bare word TODO despite a working documented command.
3. Correct the CI line in PROJECT.md to say what is true — that pytest tests/ -q also works, with no CI wired up.The undocumented exit codes in main() are real but minor; they're 2 for no-args and 0 otherwise. I've edited nothing. Say the word on which of these to take.
Support: [email protected]Personal use licence. One buyer, install on your own machines. No resale or redistribution.Payments and receipts are handled by Lemon Squeezy.Clearstack is an independent product. It is not affiliated with, endorsed by, or sponsored by Anthropic. Claude and Claude Code are trademarks of Anthropic.