Every post in the Metered series, from what are you counting to paying for outcomes, leaned on the same file: the public usage-consumption skill for Salesforce Revenue Cloud. I quoted it as a source for how the platform behaves, and always as the repo's account rather than anything I'd run. The file itself, and the forty-odd around it, turned out to be a more interesting artefact than most of what they describe.
A skill, in Anthropic's own definition, "is a directory containing a SKILL.md file", and the format exists so that "anyone can now specialize their agents with composable capabilities by capturing and sharing their procedural knowledge." Anthropic published it as an open standard on December 18, 2025. Its neighbour, AGENTS.md, pitches itself as "a README for agents", says it is "used by over 60k open-source projects", and is now "stewarded by the Agentic AI Foundation under the Linux Foundation." Both are plain markdown that an agent reads before it does anything, which matters more than any of their syntax.
A notebook with release numbers
The skills come from rlm-base-dev, a CumulusCI build project for Revenue Cloud, and the same project was republished on September 9, 2026 under the SalesforceLabs organisation as revenue-cloud-foundations, a fresh repository and not a GitHub fork. The original now carries 47 skill directories, running from billing setup to Robot tests.
The usage ones read less like documentation than like the notes of someone who kept breaking the product and wrote down what broke. Rules carry a date and the error text that established them: a commitment product's policy record "may carry ONLY a UsageCommitmentPolicy", because setting either period field fails with a quoted INVALID_INPUT message, and the entry ends "Live-verified 2026-07-24." Further down the same file is a retraction made in place. An object table had listed two policy lookups on TransactionUsageEntitlement until 2026-08-14, and the correction says that "was an authoring error, not a 264 removal, so do not 'restore' them and do not describe them as removed in 264." The evidence is a count: the object was described on a 262 org and a fresh 264 org, "40 fields → 36, exactly four removed", none added, and the two lookups absent on both. A wiki page would have quietly fixed the table; this one keeps the mistake on the page because the next reader is likely to make it again.
What the documentation is for
The skill that routes agents to Salesforce's documentation is careful about what that documentation does. Its table sends object and field shape to the developer guide and "what a configuration value means" to the help pages' one-line definitions. For runtime behaviour, "drawdown order, what each bucket decrements by, why a rated summary is empty", the answer column says "Neither. Establish by live verification." Then the instruction: "Do not infer runtime behaviour from a field description in either snapshot", and when you establish a rule, record it "so the next agent does not have to rediscover it."
I don't read that as a complaint about Salesforce's docs. A reference guide describes fields because fields are what a vendor can commit to across releases; drawdown order and the interaction between three policies are emergent behaviour that a vendor rarely puts in writing on any platform, and in my experience the knowledge has always lived with whoever had shipped the thing and been bitten by it. What's new is a file format that gives that person a reason to write it down.
Distrusting its own paperwork
The habit extends to the repo's own metadata. Its skill manifest carries a last_verified stamp, and above it a warning that the stamp "means someone RESOLVED every path in this file against the working tree, not that they read it." It then records that an earlier stamp "did not survive that test": two skills were unlisted, some referenced filenames "had never existed", and one documented corpus pointed at a directory "that has never been committed on any branch."
Some of the knowledge has gone further and become executable. The usage skill's verification page lists "18 checks, no org needed" that encode platform rules as offline tests: a tier entry with no adjustment "rates as zero", tier bounds must stay identical across currencies while the rates inside them are converted. The last check, docs_state_the_real_count, fails if the page or AGENTS.md advertises the wrong number of invariants, and the page notes that "it counts itself." Below the list sits a check retired at release 264, with the reason and "Do not re-add it." The instructions for adding a check ask for a docstring stating "what the platform does when the invariant is violated", because "that docstring is why the next reader trusts the check."
Why anyone bothers
Runtime knowledge has been written down before, in runbooks, wiki pages and handover documents, and mostly it rotted, because a page that's wrong costs nothing until a person happens to read it months later. A skill is read at the start of every agent session, and the agent acts on what it says, so a wrong line turns into a failed run within days and someone has to fix it; that's where the dates, the retractions and the self-counting test come from. I see a small version of it in the repository for this site, whose agent instructions record that emdash 0.4.0 "defaults --url to localhost", so an exported environment variable "is never read", and that its login token refresh "is broken in 0.4.0." Both were learned by hitting them, and both are written down because something reads that file every session.
Where it stops being a spec
As with any notebook, the findings are true of the org and release they were made on (262, 264, a "fresh" org on a given day), and they stay true only while someone keeps running the experiments. The SalesforceLabs copy shows what happens otherwise: it hasn't been pushed since September 16, while the original has taken 363 commits since, and the Labs verification page still says "17 checks." Its manifest still carries the warning about stale stamps, under a last_verified of July 26. The more official-looking copy is the older one, and nothing in it says so.
There's also the question of who stands behind it. A developer guide is the vendor's commitment; a skill in a project repository is one maintainer's record of what happened, and it's only as good as their last test run. If the most precise account of how an enterprise platform behaves at runtime is now being written in files meant for agents, should the vendor's documentation team be maintaining them, or only reading them?
No comments yet