Rechargly Marketing
What this thing is, how it is put together, and how to get real work out of it. Fifteen minutes to read once, then keep it open for the first week.
You already know how to work with Claude. This guide is not about that. It is about what is different when Claude has a memory, a filing system, and keys to your actual tools.
The short version: a chat forgets you the moment you close the tab. This does not. Everything it learns about the market, the audience, your brand, your decisions and the language your customers actually use goes into files that stay. Every session starts from what the last one finished with.
This brain sees exactly what you see. Nothing more. Every connection signs in as you - your Slack, your HubSpot, your Notion, your Google account. It has no back door into anyone else's conversations, and there is no bridge to Matt's setup even though yours was forked from the same chassis.
That is deliberate and permanent. It is written down in ACCESS-BOUNDARIES.md, and it is the reason you can point this at your whole working day without thinking carefully around it.
If it ever seems to know something you would not know, that is a fault. Stop, and tell Matt.
Seven folders. The split between the first two is the important one.
voice/ is read-onlyTwo reasons, and the second is the real one.
It keeps your voice yours. If the brain could edit your writing, then over months your recorded thinking would drift into its phrasing, and the corpus it learns your voice from would slowly become its own voice reflected back at you. That failure is invisible while it happens.
And it makes the brain safe to be blunt with. Somewhere to write a half-formed thought, a frustration, or an unvarnished read on a situation, knowing nothing will rewrite it, soften it, or push it anywhere. That is only true if the rule is absolute, so it is absolute.
Every file in wiki/ is plain Markdown with [[wiki links]] between pages. Point Obsidian at the wiki folder and you get search, backlinks and the graph view over exactly the same files the brain reads and writes. Nothing is duplicated and nothing is locked in - it is the same folder, seen two ways.
Use Obsidian when you want to browse and think. Use Claude Code when you want to ask and produce. Neither is the real one.
One thing to type. Nothing to paste, nothing to configure.
/settle-in
The brain interviews you, so it can stop guessing.
It knows a fair amount about Rechargly and almost nothing about you. That gap is deliberate - a confidently wrong profile is worse than an empty one and much harder to spot later. So the first session closes it properly rather than by inference.
It runs as a conversation, one question at a time, and it writes what it learns into the files as it goes and shows you each line. About twenty minutes. Five things get covered: what marketing actually owns here, what you are trying to do, how you write, how you want to work with it, and what you have brought with you from Cowork.
Two of those are worth knowing about in advance.
It will not ask you to describe your writing style. It asks for three or four things you have actually written instead. Everybody describes the writer they would like to be, and a style description produces drafts that sound like an aspiration. A handful of real emails produces drafts that sound like you.
It will ask what you would find irritating rather than helpful. Answer that one properly and specifically. It gets written down verbatim and it saves weeks of low-grade annoyance.
The session ends with something real rather than a completed profile, most likely a first run of /listen. If a setup session produces only admin, the whole thing reads as overhead, and it is not.
A note on being coached. For the first fortnight the brain will occasionally point out when you are doing something the hard way - a habit that made sense in Cowork and is unnecessary here. One line, once per habit, and then it gets on with what you asked. It stops entirely at the end of August. If it is annoying before then, say so and it stops immediately.
Worth knowing before you start, so you neither rebuild something that exists nor assume something is there that is not. The brain arrived with a fair amount of context already written, and with several deliberate holes.
| Where | What is in it |
|---|---|
wiki/marketing/signal-engine.md | The piece most worth your time. How marketing signal works here: the sources available free today, how to turn an event into a meaning, what to make, and who it gets routed to. Plus what to build first and in what order |
wiki/marketing/audience-data.md | The data asset. Roughly 21,000 accounting-firm records in HubSpot, how well they match for audience building, how to segment them, and what LinkedIn will and will not let you do with a list |
wiki/company/ | Orientation. What Rechargly does and sells to, the tool stack including what has been retired, who's who by role, and the locked terminology |
wiki/protocols/docs-stack.md | Where each kind of output is supposed to land, and why |
These exist as pages but are mostly scaffolding. They were left thin on purpose, because guessing at them would have been worse than admitting they were not known.
memory/user.md is nearly empty. Your remit, your goals, your voice rules, how you want to work, and what you find irritating rather than helpfulwiki/_hot.md is your flags and nobody else's. There is no inherited priority list, no task board handed down, and no plan waiting for you to execute.
One page is worth knowing about: wiki/marketing/parked-ideas.md. A strategy document written before you arrived contained a short section of marketing suggestions - a newsletter, a named-individual advertising pilot, some event support. They are recorded there so they are not lost, explicitly framed as somebody else's parked ideas with no standing. Binning any of them needs no justification. You are the marketing manager; the ideas come from you.
This was forked from a chassis Matt also runs for his own seat. His priorities, his flags, his commercial terms, his account material and anything from a private conversation were all left behind. Not trimmed for tidiness - excluded, because they are not yours and in several cases not appropriate for you to hold.
Read the honesty markers as real. Throughout the wiki you will find [TO CONFIRM], [TO CAPTURE] and [my read, worth checking]. They are not padding. They mark exactly the places where something was inferred rather than known, and they are the difference between a brain that admits its gaps and one that quietly fills them in with something plausible.
Nine, and four of them are one-word habits. That is on purpose. A system with thirty commands has twenty-five nobody remembers, and unused machinery is confusion rather than value.
Run /listen first. Hundreds of call notes are already in HubSpot, written automatically, and nobody has ever mined them. One session produces a list of the phrases your market actually uses - immediately useful to everyone who writes anything, and the fastest way to learn this market's language without sitting through fifty calls.
It starts deliberately timid and climbs one lane at a time.
| Level | What it may do |
|---|---|
| Observer where you start | Reads and reports. Does not act. |
| Advisor after about a week | Drafts anything. You see everything before it goes anywhere. |
| Partner lane by lane | Acts unattended, but only in lanes you have named and it has written down. |
The level lives in memory/soul.md. Move it when you are ready and not before. There is no prize for rushing.
Nothing sends, publishes, schedules or spends until you type an actual token. Not "yes", not "go ahead", not a thumbs-up - a typed token.
| To do this | Type |
|---|---|
| Send an email or message | confirm-send |
| Publish or send anything in HubSpot that reaches an audience | confirm-publish <asset> |
| Write to a Notion page | confirm-notion <page-id> |
| Write to a HubSpot record | confirm-hubspot <record-id> |
| Post to a social channel or start paid spend | confirm-post <channel> |
Reads are always free. You never need to authorise looking at something. The friction is deliberately all on the way out, because in marketing the way out cannot be undone.
The publishing gate matters more here than anywhere else. A marketing email reaches a few thousand people and cannot be recalled. A wrong figure in a live landing page is public. Treat every publish as outbound, every single time, no matter how small it looks.
Worth knowing before it costs you something.
It can produce a fluent, well-structured, completely wrong answer, and it will read exactly like a right one. There is no tell. The defences are that it says where things came from, and that you check anything expensive to be wrong about before it ships.
You will see two kinds of statement. Something drawn from a real record is stated plainly and the source can be produced on request. Something that is the brain's own read carries a visible flag: [my read, worth checking]. If a claim carries neither, that is a bug - say so.
It should never guess one, and it is told not to. If it gives you a figure without a source, do not use it. Ask where it came from.
Automatic meeting transcripts mangle names of people and firms constantly, and those mangled versions end up in the record. Every name headed for anything published gets verified against a real source. This market notices a misspelled partner's name.
The single most dangerous habit in AI-assisted marketing writing is the plausible placeholder - an example statistic, a sample customer quote, a case study sketched in to show the shape. It is meant to be replaced. It very often is not. This brain is instructed to leave a visible gap instead, like [NEED: real figure, source unknown]. If you ever see a specific-sounding statistic you did not supply, treat it as fabricated until proven otherwise.
Run /today even though the cache is empty - it shows you the shape. Then run /listen, which is the one that pays for the whole setup in a single session.
Then spend twenty minutes filling in memory/user.md. It is deliberately near-empty because guessing at you would have been worse than admitting it did not know. The voice section wants three or four things you have actually written rather than a description of your style - everyone describes the writer they would like to be.
Use /new for everything, without thinking about where it goes. Read wiki/marketing/signal-engine.md, which is the piece of thinking most worth your time in here. Watch the six Slack product channels and say nothing - you are learning what normal looks like, and a signal system built before you know the baseline generates false alarms.
Then /process-inbox on Friday, and move it to Advisor if the drafts have been good.
Message Matt. Do not debug a connection on your own - a brain quietly wired to the wrong Notion workspace produces confident, wrong work, which is considerably worse than one that is plainly broken.
The most common fault by a distance: at the Notion consent screen it is easy to pick a personal workspace instead of Rechargly's. Nothing about it looks wrong from the inside. If the brain seems to be reading things that are not quite right, check that first.