plumb /plŭm/
Set Claude true to how you work
Claude forgets your rules, deletes before asking, and writes like a brochure. Plumbward fixes all three.
One-time purchase. 14-day refund. Works with Claude Code, Desktop and Cowork.
the setup interview writes plumbward-profile.md
voice, guardrails, context
One composed session, not a recording. The dry-run files and ages come from our published eval fixture.
That session is composed. This one is not.
Forty-five seconds from a real session on 3 September, captions burned in, nothing retyped. It is the setup interview trying to capture a voice.
- It asks for three to five things you wrote yourself. Not AI-drafted, or it would learn the voice you are trying to escape.
- It searches the drive and refuses what it finds. Forty files, almost all written with Claude. Those would teach it the AI voice back.
- Told to use chat messages anyway, it does, and says so. One real message and four commands. The profile is marked partial, with the reason, where you will see it later.
- The draft still catches the real habits. No greeting, the caveat in brackets at the end, the number two where a "to" would go.
Out of the box, Claude drifts.
These are the five that show up again and again in daily work. Every one has a named fix, and you can watch it work in the session above.
| The common failures | Our fix |
|---|---|
| Deletes and overwrites first, apologizes after. | dry-run Asks for the full list and your approval before anything is deleted, and shows its reasoning per file. Measured: 0 of 30 runs changed files first; bare Claude, 30 of 30. |
| Long sessions lose your rules mid-project when context is compacted away. | context-guard Your guardrails live in a checkpoint file and are re-loaded when a new session starts. Machine-enforced in Claude Code. |
| Burns through usage limits re-reading files and padding answers. | token-diet The wasteful habits are named and cut. |
| Writes in fluent AI-brochure voice that you then rewrite by hand. | voice Captured from real samples of your writing, then applied to every draft. |
| Invents plausible-looking citations and states recollection as fact. | verify Separates what was checked from what is recalled, and flags the unchecked, per item. |
Six Custom Made Skills. One profile. Machine enforcement in Claude Code.
Every skill reads a profile built from your answers and your writing. The same download behaves differently on your machine than on anyone else's, because it is shaped by you.
setup
A guided interview, once. It asks how you work, what Claude must never do without asking, who you write for, and collects real samples of your writing. It writes your profile to your machine (nothing is sent to us) and every other skill reads it from then on. A copied zip without the interview is just text files.
voice
Stop rewriting every draft by hand. It learns your cadence, formality and the words you'd never use from 3 to 5 real samples, then strips the tells that make text read as generated.
verify
Never publish a citation nobody checked. Each item is marked verified, recalled or uncertain, and it refuses to print what it could not confirm.
dry-run
Nothing gets deleted behind your back. You see the full list first, oldest to newest, traps flagged, and it waits for your yes.
context-guard
Long sessions stop losing the plot. Your guardrails are checkpointed and reloaded, so a compacted session picks up where it left off.
token-diet
Stop paying for padding and re-reads. Five ranked cost wins, plus a full tactics reference for the habits that quietly burn your limits.
hooks Claude Code
The enforcement layer: your guardrails are re-injected the moment context is compacted, and destructive commands are gated behind a confirmation. Machine enforcement, not a promise in a text file. It is the one part a copied zip cannot fake.
Interview once. Everything reads it.
Install the kit
Two install paths, a few lines each: upload the .plugin in Cowork or Desktop, or add the marketplace folder in Claude Code.
Say “set up Plumbward”
The interview runs in your own Claude session. It ends by writing
plumbward-profile.md into your working folder: your voice, your
guardrails, your context, in plain text you can read and edit. A redacted sample:
# plumbward-profile.md voice: short sentences · no exclamation points · sign-off "Cheers" never-say: "I hope this finds you well" · "leverage" ·████████guardrails: never delete without the list · never email██████context: agency owner · clients███████████· billing in day rates
Work normally
Every skill reads the profile before it acts. Ask for a draft and it comes back in your voice. Ask for a cleanup and you get the list first. Start a new session and your rules load with it.
What it costs you
Every config kit taxes your context. Ours is published, per skill.
A kit that hides its own overhead is selling you a feeling. These are the CLI's own projections for the installed kit, published as-is, with the unflattering number first, because that's the habit this product is built on.
| context-guard | ~150 always-on · ~1.6k when it fires |
| setup (fires once, ever) | ~100 always-on · ~2.8k at interview |
| verify | ~100 always-on · ~1.6k when it fires |
| voice | ~100 always-on · ~2k when it fires |
| dry-run | ~90 always-on · ~1.5k when it fires |
| token-diet | ~90 always-on · ~1.5k when it fires |
| hooks, the enforcement layer | zero model context |
Projections from claude plugin details, reported verbatim; the CLI qualifies them as estimates. Whether the kit saves more than it costs is a separate question, answered below.
One price. No subscription required.
Buy it once. Lifetime updates on this version line, a 14-day no-questions refund, and no subscription to cancel.
Lite
Two skills, no interview
$0
Free. Email only, no card.
- dry-run: the full list before anything is deleted
- token-diet: the habits that burn your limits, named
- ~196 tokens always-on
- Claude Code, Desktop and Cowork
Also on GitHub, no email needed. Upgrade to the kit any time.
Integration Kit
Integrates Claude to your personality
$79
One-time. Yours for good.
- The setup interview: your profile, built from your material
- voice, verify and context-guard, all reading your profile
- dry-run & token-diet, profile-aware
- Hooks enforcement layer (Claude Code)
- Lifetime updates on the v0.2 line
- 14-day no-questions refund
Checkout by Polar. 14-day no-questions refund.
What we've measured, and what we haven't.
This category sells adjectives. We'd rather show you the ledger, including the entries that don't flatter us.
Measured
- That it asks before it deletes. Ten cleanup scenarios, three runs each, same model, same tools. Bare Claude changed files before showing a list in 30 of 30 runs. With the kit, 0 of 30, and it stopped to ask in every run. Measured 25 September 2026. The full results.
- That the kit activates unprompted. In our test harness, the kit's skills fired in 20 of 20 sessions where it was installed, and zero where it wasn't, read from the raw session traces.
- The kit's own overhead. ~630 tokens always-on, roughly 1.5k to 2.8k when a skill fires, hooks free. Published above.
- That our test rig is fair. Both test arms ran identical tool surfaces. We threw out an earlier pilot of our own when we found they hadn't.
Not measured, so not claimed
- A token-savings figure. None exists yet. Until it does, no savings number appears anywhere on this page.
- That the kit out-scores a bare model on citation discipline. Twice tested, twice no difference on the grader. A 20-topic single-run test in August found none. A re-run on 25 September 2026, two runs per topic and a stronger judge, passed both arms in 40 of 40 runs, because both saw the blocked web tools and hedged. What did differ: the kit printed 36 print-ready citations to bare Claude's 63, left 3 of them unmarked to bare Claude's 23, and drew 0 dissenting judge votes to bare Claude's 7. Suggestive, not a claim. The full results.
- Anyone's productivity multiplier. You will not find an “×” on this page.
Why publish the misses? Because the product's whole premise is that Claude should flag what it can't verify instead of inventing it. A storefront that didn't hold itself to the same rule would be selling something it doesn't practice.
FAQ
Isn't vanilla Claude fine?
For short, low-stakes work, often yes. The kit earns its keep where the defaults fail: long sessions that forget your rules, cleanups that delete before asking, drafts in a voice that isn't yours, citations that were never checked. If none of those have bitten you, you may not need this.
Isn't this all free on GitHub somewhere?
Generic versions of some of it, yes, and our own Lite tier is free: dry-run and token-diet, no interview needed, from the pricing card above. What you can't download from a list is the interview: the kit's value is created on your machine, from your writing and your rules, when setup runs. A pile of skill files without your profile is the commodity; the profile is the product.
Does it actually work, or is this placebo?
Here's exactly where the evidence stands. Measured: the kit's overhead (published above); that its skills reliably activate, 20 of 20 installed sessions in our harness, verified from raw traces; and, on 25 September 2026, that it asks before it deletes. Across ten cleanup scenarios run three times each, bare Claude changed files before showing a list in 30 of 30 runs and the kit in 0 of 30 (the full results). Not shown: an advantage on citation discipline. Two tests, August and September, found no difference on the grader; the kit printed fewer unverifiable citations and marked more of them, which is suggestive and not a claim (the full results). Whether that's worth the price is a judgment we'd rather you make with the real numbers in front of you. There's a 14-day refund if you judge wrong.
What do updates cover, and what if Anthropic ships all this natively?
Updates: every update to the v0.2 line is included in what you paid, forever: fixes, compatibility work as the platform changes, improvements to the six skills. No renewal, no re-purchase for this line. Whether a future major version costs anything, and what existing buyers get if one ships, isn't decided yet, so we won't promise it here. What you bought keeps working and keeps getting its updates either way.
If the platform absorbs it: possible, platforms do. Two parts are hard to absorb: your profile (it's built from your material, and it's a file you keep), and the interview that builds it. If Claude someday does natively everything the kit does, you'll still have both, and the kit will have been a one-time purchase.
What happens to my data?
Nothing is sent to us. No profile, no writing samples, no sessions. The interview runs inside your own Claude session and writes plumbward-profile.md to your working folder: plain text you can open, edit, or delete. Your conversations with Claude remain governed by your Claude provider's terms, exactly as they are today.
Which Claude products does it work with?
Cowork and Claude Desktop (upload the .plugin file), and the Claude Code CLI (add the bundled marketplace folder). The hooks enforcement layer is Claude Code only; skills work everywhere. Plumbward is an independent product; it works with Claude but isn't affiliated with Anthropic.
What does it not do?
It doesn't make Claude smarter or raise your usage limits. The paid kit's personalization doesn't work without running the interview once. It can't physically prevent every mistake: skills guide the model strongly, and only the Claude Code hooks are hard enforcement. And it won't give you a productivity multiplier, because we haven't measured one.
What if it's not for me?
Email uzayrhai@gmail.com within 14 days for a full refund. No form, no argument. Everything the kit produced from your material stays yours.