The procedure for every doc change in common-docs or a repo's CLAUDE.md, FEATURE.md or SKILL.md. Use when creating, editing, moving or deleting a doc, finding duplicate or disagreeing docs, planting a repo pointer, consolidating a node, or on /docs and the daily docs-steward run.
<!-- SYNCED COPY — do not edit here.
Canonical: common-docs/skills/docs/SKILL.md
This file is distributed to every consuming repo by
common-docs/meta/scripts/sync_skills.py. Edit the canonical, run the
sync, and commit each repo. Edits made here are overwritten and lost. -->
Find the subject's node in the domain tree (Domain → Feature →
Sub-feature). The doc goes in that node's folder under systems/, or in a temporary
projects/ folder only while work is tangled across two or more homes.
No node fits: add a feature where its domain clearly holds it (same commit, tell Arman), or
propose the change in chat. Never invent a domain.
Something already covers the subject? Edit that file. Never create a parallel doc.
A repo keeps only rules tied to a code path in that directory (a trap, an import ban, a file
map). Everything with meaning (what, why, status, plans, contracts, vision) lives in the node.
Done when: the target path is a node folder and no other file covers the subject.
2. Edit = whole-document review
Read the whole doc, not the section you came for.
Verify every claim you keep or add against live code, the DB, or git. A doc's own "verified"
is not evidence. What cannot be checked here is marked UNVERIFIABLE — <what would prove it>.
Put the change where it belongs and merge in place: no addenda, no "Update:" blocks, no
history, no changelog. Two lines saying one thing become one.
Cut what a frontier model already knows and what a command or file read reveals.
Never lose a rule: tightening keeps every invariant, path and pointer. Removing a rule is
something you say out loud in your report.
Done when: the doc reads as one voice and every claim in it is currently true.
3. Dedupe and verify (one subject)
Census. Grep the subject's terms across common-docs and every repo it touches; follow
one ring of links. List every file that makes claims about it.
Elect the survivor: the node's file from step 1.
Classify each disagreement: fact vs fact → reality decides, fix the loser; doc vs
Arman's words → his words win, and code that drifted from them is a finding, not a doc fix;
his words vs his words, or a disagreement about meaning → step 6, never resolved by you.
Move, never copy. One source file at a time: merge its unique truth into the survivor,
git rm the source, repoint every inbound reference in every repo, commit all of it in one
command. A file never exists in two places across a commit.
Prove it. For every deleted path, grep all repos plus common-docs: it comes back empty.
Never touch: generated files, a .md that code reads (grep .py/.ts for its name first),
a published package README, inbox/, a repo's .arman/. A spotted duplicate that is not
blocking you is still yours: run this section or dispatch it, never just report it.
4. Clean up as you go
Working notes belong in your project or node; nobody cleans them but you and the daily sweep.
Each time a step finishes, in the same commit:
Status: the step becomes one line on its board or main doc ("P27 cutover: landed").
Resume point: one line ("next: steps 2 and 3"); delete it when the lane closes.
Evidence: say how to rerun it (the command); keep only what a later decision needs.
Censuses, scans, logs and batch-by-batch verification notes are not kept once read.
Arman's words go to the node's VISION.md (§5). Anything for him is shown in the chat.
Clean-up-later notes: working material you must keep for now gets one line at its top saying when and how to cut it ("Once Meet ships, cut to what the spec lacks"). The daily sweep honors and acts on these notes.
Finished work shrinks to a status phrase ("Gmail: live"); finished projects, closed
handoffs and stale plans are deleted, never archived. Git holds history.
5. Arman's words
Any verbatim quote of Arman found outside a VISION.md moves, verbatim and dated, into the
closest node's VISION.md (create it with type: Vision, authority: owner if missing), and
the source line is cut. An agent's paraphrase attributed to him is deleted. Never rewrite,
trim or "fix" existing VISION content. You never set authority: owner on your own writing.
6. Conflicts
A disagreement about intent or a ruling: one or two sentences in
/operations/conflicts.md, then ask Arman in chat. When he answers:
delete the entry, apply the answer, tell him how many remain.
7. Repo pointer lines
Every repo that touches the node gets ONE line, in that feature's FEATURE.md or else the repo
CLAUDE.md, never a stub file and never restated content:
Cross-repo system-of-record: /Users/armanisadeghi/code/common-docs/<path> — read it before touching this feature in ANY repo.
A renamed or deleted doc orphans its pointers: grep every repo for the old path in the same
session.
Repo checks after a doc change: aidream python3 scripts/check_doc_links.py (--all for
package CLAUDE.md, PRINCIPLES.md, FOUND_DEFECTS.md, skills) and
python3 scripts/check_docs_guards.py; matrx-frontend pnpm check:docs-guards and
pnpm check:doc-claims. FOUND_DEFECTS.md IDs: defect ownership § Entry IDs.
8. Finish
Frontmatter with a non-empty type on every new .md; bundle-relative links; the folder's
index.md updated.
python3 meta/scripts/okf_lint.py exits 0. Editing a SKILL.md → also the
skill-authoring rules and
python3 meta/scripts/skill_descriptions.py lint.
Fresh-reader review for guidance you wrote (CLAUDE.md, SKILL.md, STATE, FEATURE, memory):
dispatch one fresh standard agent with only the file paths and
reviewer-brief.md sent verbatim. Apply each CUT,
REWRITE, MOVE; reject one only with a factual reason; answer each ASK from live
evidence; turn a GUARD into code or a feedback item.
Commit by pathspec, git pull --rebase, push. Unpushed docs do not exist.
Daily sweep (the docs-steward schedule)
The sweep cleans what agents left behind while the work is still fresh. Commit message carries
the scorecard (no log file).
Yesterday's changes first: every doc changed in the last 24 hours
(git log --since=24.hours --name-only). For each: apply §4 — finished steps become status
lines, regenerable evidence goes, resume points for closed lanes go, Arman's words move to
VISION.md, duplicates merge (§3). Keep anything a live lane, an open decision or a running
script needs; when unsure, keep it and list it in the commit message.
Large binaries and source data added that day go to file storage and are linked
(how).
Instructions that make agents write logs: if a skill, brief or plan told an agent to
keep a log, evidence dump or per-event record, fix the instruction to §4, not just the file.
okf_lint.py to zero; sync_skills.py --check.
Inbox: each new item is dispositioned per /inbox/README.md; nothing
in inbox/ is ever deleted.
Admin debug system wiring for the floating AdminIndicator's live debug data and Copy Full Context. Use when adding debug visibility to a route or feature, wiring useDebugContext or the debug panel, capturing console errors, or enabling the copy-context workflow for a page.
Compact two-icon Copy / Copy-for-AI controls (components/agent-copy). Use when adding copy buttons to a row, card, list, or record; merging duplicate Copy/JSON/AI controls; continuing the copy rollout; or writing a Copy-for-AI payload. NOT for markdown content actions (use rich-document-actions).
Disclosing a surface's existing fixed AI jobs in the shell's top Agents menu. Use when a page, panel, overlay, or window already runs a mandate behind a button, assist, automatic action, or mode; when the agent-disclosure guard names a file; or during a surface check.
The watch-fix-rerun method for making an existing platform agent or automated process efficient and correct: baseline its ledger, run one unit yourself, fix the class behind every wasted call, rerun, record. Use when asked to improve, optimize, watch, or 'make efficient' an agent, a sandbox session, a sync job, or any recurring automated process.
Redux state for agent execution and firing agent shortcuts. Use when editing features/agents/redux/, building agent UI, creating a conversation, touching assembleRequest or NDJSON stream state, wiring agent-state selectors, adding a per-conversation capability, or triggering a shortcut from a button, menu, or code (useShortcutTrigger, launchAgentExecution).
The provision for a call site: the exhaustive menu of values that place in the code can realistically produce. Use before creating or fixing a mandate or agent that reasons about more than its own input, when deciding whether an agent can answer at all with what it is sent, or when its output 'looks right' but nobody checked it could be. NOT for building the agent (use create-agent).