The situation
Coldcard had the product and the technical reputation. Bitcoin-only firmware, air-gapped signing, a feature set serious custody people had recommended for years, and documentation deep enough to answer almost anything a current owner might ask.
What sat outside that was a content layer that preceded ownership. Someone deciding whether to hold their own private keys, or trying to understand what air-gapped signing protects against, had to learn those things somewhere else. In practice that meant exchanges, who have a structural interest in customers not self-custodying, or multi-crypto hardware companies, who cannot make a bitcoin-only argument because they are positioned for the opposite.
The conversation that leads a reasonable person toward self-custody was happening on other companies’ terms. Bringing it in-house meant owning the explanation and competing for the searches that come before a purchase.
The main constraint was scope against time. This was a contract engagement over three months, and the plan called for the output of a small content team.
The strategy
Conduct a strategic audit before writing. I ran a keyword and competitor analysis with Ahrefs integrated with AI, looking for where competitors were winning, where they were thin, and which terms combined useful volume with beatable competition. That separated the easy wins from the searches worth fighting for, and it decided the shape of the programme: roughly 45 Learn articles, six comparison pages, and seven guides.
Then came the part that AI cannot do. For each piece I wrote the outline by hand: what the reader already believes, which misconception has to be dismantled before anything else lands, how long it should run, where the section breaks fall, which analogy should carry the concept, and how someone who already knows the material can get out early. That outline was the strategic thinking that shaped the entire project.
AI carried the drafting against those outlines, and I read and rewrote until each piece was production ready. Human review sat at both ends rather than only at the end, because a bad outline produces a fluent article that is subtly wrong, and that is harder to catch than a clumsy one.
Selected work
A “PSBT” (partially signed bitcoin transaction) is a portable file format that separates the job of building a transaction from the job of signing it.
You can think of a PSBT like a contract prepared by a law firm and sent to a client for signature. The firm drafts every detail and packages it to be sent to the client. The client reads it, signs it, and sends it back. In Bitcoin, the coordinator is the firm and the signing device is the signatory.
BIP174 defines six formal roles for the participants in a PSBT workflow. In practice, most setups compress these into two participants.
Three different readers arrive at this same page. Someone comparing wallets needs one sentence, someone setting up a first device needs an analogy to grasp the concept, and someone building a multisig quorum needs all six roles and the difference between BIP174 and BIP370.
The structure lets each of them leave at the right moment, which is why the analogy sits second rather than first. Opening with it would patronise the third reader, and the piece would lose the audience that actually buys the device.
The UTXO data in the per-input section is the security-critical part. The signing device uses these values to calculate and display the actual amount being spent and the transaction fee. Without UTXO data, the device would see only abstract input references and could not show the user a meaningful spend amount. A malicious coordinator could inflate the fee by misrepresenting input values.
The signing device is not taking the coordinator’s word for what is being spent.
This paragraph makes the case for verifying on the device’s own screen without naming a competitor, and it works because an informed reader reaches the conclusion before any product is mentioned. The alternative draft made the same point as a product claim and was cut, because this audience discounts vendor claims but does not discount mechanics.
Audit. Keyword and competitor analysis to find where demand sat, who held it, and what was winnable. Ahrefs for the data, AI for the pattern-finding, both pushed hard on being critical rather than agreeable.
Plan. The audit turned into a programme: which pieces are acquisition, which are onboarding, which exist to be cited by an AI assistant, and how they link to each other.
Outline, by hand. Reader’s starting belief, the misconception to dismantle, length, section breaks, the analogy, the exit point for experts.
Draft with AI, against the outline and a fixed set of editorial standards: question-form headings, key takeaways, no unhedged security claims, no competitor named without a source.
Read and rewrite. Every piece went through me until it was publishable. The AI rejects were usually fluent, but structurally lazy, following the outline while skipping the part that made the concept land.
The failure mode of this is a library that reads well, but teaches nothing. AI models will happily produce a competent article from a shallow outline, but will lack signal and an educator’s discernment in the output.
This is why the outline stage stayed manual and took the largest share of the work. The drafting stage is commoditized and quick with AI, and pretending otherwise is how content teams ship volume that nobody reads.
Outcome
A content team’s quarterly output, delivered in three months by one contractor:
- More than 45 Learn articles across bitcoin fundamentals, how bitcoin works, self-custody, hardware wallets, seed phrases and advanced topics.
- Six comparison pages against the competitors buyers actually shortlist.
- Seven guide walkthroughs, covered on their own page.
- An information architecture and a set of editorial standards the library still follows: question-form headings, key takeaways, sourced claims, and a stated review date on every page.
The library is interlinked, structured for retrieval by search engines and AI assistants, and sequenced so a newcomer can start anywhere and find the next thing they need. It gives the company somewhere to send a prospective customer who is not ready to buy, and it moves the argument for self-custody onto ground Coldcard controls rather than an exchange’s blog.
Written as a content consultant for Coldcard, March to June 2026.