vibe coding
verified

one command, from any project folder

1. ○finds your project and its GitHub remote, and reuses the PAT you set once
2. ○brings up the firewall and proves it – outbound on a short allowlist, verified before work starts
3. ○signs Claude Code in with your Claude subscription – no API key
4. ○ready in seconds – the container's already built, the PAT's already on file

once per repo, not every day

1. ○detects your project's GitHub remote, then asks once for a fine-grained PAT – scoped to that one repo
2. ○builds the sandboxed container image (first launch only – reused in seconds after that)
3. ○saves the token and reuses the image, so every launch after this one is instant

once every 90 days, in seconds

1. ○finds the repo and rotates its stored token in place – nothing else to update
2. ○same advice as first launch – set an expiry, and 90 days is a good default
3. ○takes effect on the next vibe launch – no rebuild needed

nobody marks their own homework

1. ○a planner drafts the spec – acceptance criteria plus a model plan (which tier does what)
2. ○a spec critic attacks the spec before any code – then you approve it
3. ○a builder writes the code to the locked spec
4. ○an independent tester writes and runs the tests – it never sees the builder's diff
5. ○fail means fix and re-test – the tests are immutable, claims don't count
6. ○a final evaluator reads the raw test log itself – default-fail on ambiguity

no arguments – it finds its own work

1. ○reads your TODO.md backlog and picks the first bounded open item
2. ○announces the pick and rings the bell – you get a window to redirect it
3. ○runs the work unattended, acting as you would – a hard-escalate list stops what needs your hands
4. ○commits locally with a full audit trail – never pushes; you review, you push

burn the rest of your session productively

1. ○loops the solo run, item after item
2. ○questions park to your notes – it never stops to wait
3. ○commits keep landing while you sleep – locally, nothing pushed
4. ○rolls into a fresh credit window when one runs out
5. ○stops itself at the perfection gate

traffic off the allowlist

1. ○every outbound connection is checked against a short allowlist
2. ○GitHub, npm, Anthropic and a few named extras are allowed – the rest is rejected at the network layer
3. ○and it fails closed – if the allowlist can't be built, nothing gets out

the content guard reads every diff first

1. ○staged diffs and commit messages scanned before every commit
2. ○secrets and personal data both stop the commit – secrets as BLOCK, personal data as WARN
3. ○pushes re-scan the outgoing range for secrets before anything leaves

out through the one-repo PAT

1. ○the pre-push guard re-scans the outgoing range for secrets on the way out
2. ○pushes with the repo's own fine-grained PAT – one repo per token is the whole blast radius
3. ○docs and CHANGELOG travel in the same commit

what this month cost

1. ○month-to-date tokens per model, across every vibe session on this Mac – plus estimated credit spend

lessons that travel

1. ○captures a cross-project lesson to your learning library – the write needs your yes, every time

container → Mac clipboard

1. ○copies Claude's last code block to your Mac clipboard – via a watched scratch file

token thrift on demand

1. ○lean mode – no subagents, terse replies, cheaper-model suggestions; /feast turns it back