You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
Org rules: subagent cost, parallel cap and verification, from Anthropic's prompting guidance #78
The org rules become claudeMd in the managed settings, so every Claude Code session in the org reads them: people's own sessions as well as agent runs. Raghav, 2026-11-11: the agents and the managed settings should follow Anthropic's current prompting guidance (the best-practices page and the Fable 5.2, Opus 5.5, Sonnet 5.5 and Haiku 5.5 pages).
Three points from that guidance apply to every session, whatever model a person picks:
Subagents can be a large part of what a session costs. Forks run on the session's own model, so delegate only for parallel or isolated work.
Verify with a real check before reporting work done.
Mention extra work at the end instead of doing it.
The rules say some of this already ("Done means…", "no drive-by refactors"), but nothing about subagent cost. Each person's model and effort stay their own choice, so the rules set no model.
Acceptance criteria
rules/org-rules.md gains a short section on working as an agent: when a subagent is worth its cost, at most 2 at once unless the task says otherwise, and verification by a real check. Each point says why, with no shouting. The section names no private repository.
Existing rules read for wording the guidance warns against: absolutes with no reason, and vague approval categories. Only what is needed is changed.
bin/generate regenerates managed/settings.json. bash tests/run.sh and bin/generate --check pass.
Raghav pastes the new managed/settings.json into the Claude org console (owner step), and a new session shows the section.
#79 merged and closed this issue; its description doesn't mark these 4 items met, so they stay unticked:
rules/org-rules.md gains a short section on working as an agent: when a subagent is worth its cost, at most 2 at once unless the task says otherwise, and verification by a real check. Each point says why, with no shouting. The section names no private repository.
Existing rules read for wording the guidance warns against: absolutes with no reason, and vague approval categories. Only what is needed is changed.
bin/generate regenerates managed/settings.json. bash tests/run.sh and bin/generate --check pass.
Raghav pastes the new managed/settings.json into the Claude org console (owner step), and a new session shows the section.
(board-sync, hq#1238)
Activity
realRaghavGupta commentedon Nov 11, 2026
PR: #79. Subagents section added, Changes bullet sharpened, managed/settings.json regenerated; bin/generate --check current, tests/run.sh 776/786 (same 10 Mac-only failures as main). Owner step after merge: paste managed/settings.json into the Claude org console.
setkernel-ops commentedon Nov 11, 2026
#79 merged and closed this issue; its description doesn't mark these 4 items met, so they stay unticked:
rules/org-rules.mdgains a short section on working as an agent: when a subagent is worth its cost, at most 2 at once unless the task says otherwise, and verification by a real check. Each point says why, with no shouting. The section names no private repository.bin/generateregeneratesmanaged/settings.json.bash tests/run.shandbin/generate --checkpass.managed/settings.jsoninto the Claude org console (owner step), and a new session shows the section.(board-sync, hq#1238)