Loop
A session-based working method for coding agents: plan in thin complete slices, build one at a time, verify by observing real behaviour, hand off cleanly. Skills + agents + 2 hooks for Claude Code.
npx -y skills add cbdreamer11/CB-loop-kit-claude-plugin --skill loopAssembled from the repository path, not quoted from the project. Check it against their README if it does not work.
2 things to look at
- 13 days oldThe repository was created 13 days ago. New is not bad, but a brand new repository carrying a familiar-sounding name is the shape a typosquat arrives in, and there has been no time for anyone else to find a problem with it.
- 6 stars6 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.
What its author says it does
Copied from the file, not written here
The build loop - locate yourself, investigate the real code, build one slice completely, verify it for real, fix and re-verify, then close and hand off. Use for any session that is implementing something from the plan.
SKILL.md
3.6 KB, 872 tokens by cl100k_base, as published. Nobody here has run it
The loop
Nothing is done until it has been observed working. A green build, "I read the code and it looks right", and "this should work" are not verification. The only valid signal is observed behaviour.
Run these six steps for every item. Do not skip 0 and do not skip 5.
0 · Locate yourself
- Read
.loop/STATE.md(what is done, what is next, what is parked) and.loop/GOTCHAS.md(traps this project has already sprung). git log --oneline -15andgit status. Anything already committed is not rebuilt. A lot of what a document calls missing is already built and just not visible — confirm against the code before redoing it.- Uncommitted files that are not yours may belong to a parallel session. Do not
touch them, and do not reuse a shared counter they may have taken (see the shared
resource in
.loop/VERIFY.md— read its true state with the command declared there). - The next item is the first unchecked
[ ]in.loop/STATE.md. You do not need new instructions to continue.
1 · Investigate
Open the real files you are about to change. Map what exists, what is editable,
what is missing. Never work from memory of an API or a schema.
If a question is genuinely open — or the item touches money, permissions, auth, or
schema — run the loop-council skill before writing code, not after.
2 · Build, completely
One item at a time, finished. If completing it requires one more thing, that thing gets built too: it is not reported as a leftover and it is not offered as a question. An item is complete only if it would be safe to leave exactly as it is, forever — no dead end, no control that does not do what it says, no setting with no effect.
3 · Verify for real
Run the slots in .loop/VERIFY.md that apply. The rule for each: it must produce an
artifact or an observation — a rendered page, a row, a file, a screenshot, a log
line that only exists if the new code ran. Forbidden as evidence: an exit code, an
HTTP 200, and "the build passed". Those prove the code compiled, not that it works.
4 · If it fails, fix it and verify again
Repeat 2 → 3 → 4 until it actually works. This is the loop. Do not move on with something unverified, and do not describe a failure as a partial success.
5 · Close
- Commit on the working branch, staging only your own files (never
git add -Awhen someone else has uncommitted work). Never push to the protected branch. - Update
.loop/STATE.md: mark the item[x]and add one line — what you verified and how. Record anything you could not verify asGAP: <why>. - If you were blocked, write
PARKED: <why> + exactly what is missingand move to the next item. Park, never abandon, and never get stuck on one item. - Add to
.loop/DECISIONS.mdif you chose between real alternatives, and to.loop/GOTCHAS.mdif something fooled you — the next session should not fall for it. - Report the delta honestly: what is verified, what is a GAP, what is blocked on the owner. Then continue with the next item. Do not ask "am I done, shall I continue?".
When to actually ask the owner
Only for what only they can unblock: a key or an access you do not have, moving real money, a business/legal/pricing decision, or publishing. Everything technical you decide yourself. "Do you want me to test it?" is never a valid question — if it matters, you already tested it.
What ships with it
Read from the repository
Just SKILL.md. No reference files, no scripts.