Services
A documentation audit at a fixed price of $2,000, covering one repository, one tracker, and one release. No write-back, no chatbot, nothing autonomous.
Savoie Studio reconciles a tracker, a repository, and the published documentation, and hands you the rows where those three sources disagree. It is read-only reconciliation: nothing is written back to your systems, and the decision about what to change stays with you.
Documentation audit — the first engagement
Fixed price: $2,000. You know the number before any work begins, and it does not move.
An audit covers one repository, one work tracker, one documentation source, and one release or date window. You choose that scope in writing before the run starts.
What you receive:
- A written report, in Markdown or JSON, listing the work that shipped and never reached the documentation.
- A corroboration summary showing how many closed tickets and merged changes were found, how many could be matched to each other, and how much each documentation source contributed.
- Every finding with the claim it makes, the evidence behind it, the reasoning that connects them, and a proposed action.
- An explicit list of any check that did not run, along with the reason it was skipped. A skipped check is never reported as a clean result.
- A roadmap of one to two pages describing what to fix first.
- A 60-minute walkthrough of the report.
What the audit does not do. The structural pass sends nothing outside your network. It uses no AI model, requires nothing to be installed, and reads only the metadata around a change — commit messages, ticket titles, and page text — rather than the source code inside it. It will tell you that a merged change never reached your changelog. It will not tell you whether a function matches its docstring.
Delivery takes ten business days from the day all three sources are confirmed reachable.
Release-note drafts
Fixed price: $5,500. This work is only available after an audit has shown that a gap exists.
Drafts are built from work that has already shipped, meaning corroborated tickets and merged changes, and never from the gaps themselves. Your terminology rules are applied as the draft is written, and mechanical checks look for unsupported CVE references and empty sections before anyone reads it.
A person reviews every draft, and nothing is published automatically.
This is not "AI writes the release notes." Jira and GitHub already draft from their own records, and they do it well. The difference is that these drafts are built only from work that can be corroborated across all three systems at once.
Governed drafting
For longer documentation, a four-stage pipeline plans the work, audits that plan before anything is written, produces a draft, and then verifies the draft against the original request. The verification stage uses mechanical checks that require no AI model, and a person makes the final decision at the end of every run.
This is a companion to the audit rather than an extension of it. An audit that finds nothing does not trigger a drafting run.
Quoted per engagement.
Documentation as code, and CI
I can propose and implement static-site toolchains, GitHub Actions templates, and terminology linting inside your build pipeline. This work complements an audit rather than replacing it, and it is not usually where a first engagement should start.
Quoted per engagement.
What the audit can read
| Supported today | |
|---|---|
| Repository | GitHub, GitLab |
| Tracker | Jira, or the built-in issues in either forge |
| Documentation | Confluence, MkDocs, Docusaurus, ReadTheDocs, Wiki.js, or a crawled website |
What is not supported. SharePoint, ServiceNow, DOORS, and GitBook appear in a lot of tool catalogs, but I have no working reader for any of them, so I will not take that work. Documentation sites built entirely in JavaScript cannot be read either, because the crawler does not execute JavaScript. Documentation in languages other than English has never been measured, so I make no claims about it. If your stack appears in this paragraph, I will tell you on the first call rather than in week three.
What leaves your network
The audit runs six checks. Four of them are structural: they compare identifiers — ticket keys, commit references, the text on a page — and count what does not line up. No AI model is involved in those four, and nothing leaves your network.
The other two cannot be answered by matching. They require reading a page and judging whether it still describes what actually shipped, and that judgment needs an AI model. Those calls go to Anthropic's Claude.
Two things constrain them. They are off by default and switched on for a single run at a time, and every finding they produce is checked against the page it cites before it reaches your report.
| Privacy tier | What leaves your network |
|---|---|
airgapped |
Nothing at all. No AI model is called. |
local |
Nothing. Your own inference server handles it. |
private_cloud |
Nothing leaves your accounts. Runs in your own Bedrock, Vertex, or Azure tenancy. |
hosted |
Anthropic's API, and only after you have approved it in writing. |
A team under a strict privacy constraint can pilot the structural half of the audit without opening a procurement conversation. At $2,000, a first engagement usually falls below the threshold that would start one.
What I will not sell you
- Autonomous documentation. Every finding is a review queue for a person, not an action the tooling takes.
- Write-back to Jira, GitHub, or your documentation site. Nothing I run has a write path to your systems.
- A bot that opens tickets, or a nightly scheduled run. A person starts every run.
- A chatbot over your documentation. That is a different product, and I do not build it.
- An adapter I cannot run today. I will not promise support for a platform in order to win the work.
I would rather you know all of this now than discover it in week three.
How this was tested, and on what
Everything above describes what a client engagement covers. The evidence that it works is separate, and it is deliberately public.
Before selling any of this, I ran it against nine public open-source projects — Caddy, Hugo, Kafka, Homebrew, Docusaurus, ReadTheDocs, GitLab Runner, MkDocs, and Wiki.js — plus a single-release run on Bootstrap and a Notion workspace read as a documentation source.
Those runs are written up as case studies on this site. They are demonstrations of the method rather than accounts of work done for a client, and each one says so at the top. No client engagement will be anonymized and sold back to you as a case study. The trade-off is that the numbers come from other people's projects; the benefit is that you can check every one of them yourself.
Who this is for
Mid-size SaaS companies, developer-tool vendors, API-first businesses, and regulated-industry teams of roughly 50 to 500 people, shipping faster than their documentation can follow.
A first look, at no cost
Send a link to a public repository. I will run a structural pass across its public commit metadata — no source code, nothing installed, and no AI model involved — and send back a short report of what shipped but never reached the documentation.