A standard for setting up and operating Clairvoyance Workspaces where AI Staff do sustained work.
It exists because most Workspace problems are not intelligence problems — they are memory, orientation and verification problems, and those are fixable with structure.
If you have a Clairvoyance Staff member you trust to read and write files on your machine, paste this to them verbatim. You stay in the approval seat; you do not have to read the Standard.
Adopt the Clairvoyance Workspace Standard on this instance, from
https://github.com/bubomortis/clairvoyance-workspace-starter
Treat IMPLEMENTATION.md in that repository as the AUTHORITATIVE procedure: read it in
full and follow §2. Observe these rules:
1. CHECK WHAT IS ALREADY HERE, BEFORE DERIVING ANYTHING. Search this instance for
existing workspace guidance under ANY name — a criteria card, a workspace standard
or SOP, house conventions, an earlier or adapted copy of this Standard, rules
already living in library.md or AGENTS.md, or conventions recorded in a README,
decision record or pinned note. Tell me what you searched and where, so I can judge
whether "nothing found" is trustworthy.
If you find ANYTHING, STOP. Do NOT merge, refine, update or overwrite it on your
own judgement. Tell me: what you found and where; whether it descends from this
Standard or is independent; whether it is actually followed and has been revised, or
is a stub nobody uses; where the two substantively disagree; and what replacing it
would cost me. Then give me three options with costs, and wait for my answer:
(a) KEEP MINE — adopt nothing, or lift individual criteria into what I already have.
(b) MERGE, ADDITIVE ONLY — add only what mine lacks; change, reword and remove
nothing that is already there. If the two genuinely contradict, do NOT resolve
it: leave mine standing, record the conflict, and bring it to me separately.
(c) REPLACE — only if I say so explicitly, in those words. PRESERVE the old
artifact; do not delete it.
Do not treat my interest in this Standard as approval to replace what I already
have. If you find nothing, say so plainly and continue.
2. Read STANDARD.md in full once before compressing it. Do not work from the README.
3. Produce the Criteria Card, save it as a note titled exactly "Workspace Criteria
Card", and route it from library.md. Create library.md if it does not exist.
4. Then STOP. Show me the card and tell me in one paragraph what you dropped and why.
5. Do not change, create, or reorganise any Workspace until I approve the card. When I
pick a Workspace you will score it and propose — with the cost of each item and an
explicit list of what you are NOT recommending. I approve item by item.
6. Report every file you create or modify, by path.
How to tell it worked: the card is far shorter than the Standard, it names things it deliberately excluded, and the first plan you get back contains at least one "not applicable here." If everything came back as a gap to close, send it back — the flex guidance did not survive compression.
You want your Staff to just do this → IMPLEMENTATION.md Hand it to a Staff member. They read the Standard, derive a compact working copy tuned to your setup, and apply it. You approve; you do not have to study anything.
You want to understand, challenge, or extend it → STANDARD.md The canon. 18 principles, 7 procedures, a 35-criterion rubric plus 6 external-tool criteria — with the evidence, the counter-evidence, and a section listing what was rejected during verification.
You want to check the evidence → CITATIONS.md Every source behind the Standard, with a confidence mark per entry, the claims that have no external citation, and the places the Standard overreaches what its sources actually measured.
You already adopted an earlier version → CHANGELOG.md
Current release: v1.10. Entries marked breaking for adopters change a file you were previously
told to create — in v1.3, CLAUDE.md became AGENTS.md. Entries marked tightens conformance
mean an audit that passed under the previous version may not pass now — v1.5, v1.7, v1.8, v1.9 and
v1.10 all do. v1.9 adds two Charter criteria (C17, C18), so an audit that passed under v1.8 is
incomplete against v1.9 rather than invalidated by it. v1.10 re-sources C1a/R5: tuning is the
claim the evidence supports, and provenance survives only as a narrow rule against shipping
generated output unmodified. R5's pass set moves in both directions — a file nobody has revised
now fails it however carefully it was typed, and a generated file you have tuned against real
failures now passes. Re-score R5; do not carry a v1.9 verdict forward.
You want a Staff member built for this work → Workspace Architect · Clairvoyance Engineer The Architect sets up the place. The Engineer builds the things that run in it — Exhibits, CAMs and Extensions, Binder and MCP integrations, and scheduled automation.
The two personas are maintained as standalone files so you can copy them without extracting them from a code block:
Save either to <your Clairvoyance data directory>/neurons/personas/ — the file, not the
personas/ folder, since discovery is flat and a nested copy fails silently — then confirm it
appears in the Knowledge Base dropdown when you create a Staff member. The build runbooks above
cover hiring, model tier, and a first assignment that verifies the setup rather than assuming it.
Their evidence bases differ, and you should weigh them differently. The Engineer has been in production use by five Staff members across four workspaces for several weeks. The Architect is new: at time of writing it has been exercised once, deliberately, to check that it behaves as documented. Treat the Architect as the less-proven of the two and read its first plan closely.
Disclosure: the authors of this repository run on the Engineer persona published here. That is the same familiarity bias the Standard asks you to record, so it is recorded.
A compliance checklist. Most Workspaces legitimately need a minority of the criteria — the Standard says so repeatedly and includes a section on when to ignore it. If it ever tells you to build seven files for a two-day project, it is being misapplied.
Built from structured surveys of five production Workspaces, plus published research that was content-verified — sources fetched and read, not just confirmed to exist. That process refuted part of the first draft; the correction is recorded in the Standard rather than quietly fixed.