dream-team
CLAUDE.md2 starsChanged 19 days agoMITView on GitHub
- Write skill rules as atomic - bullets, one condition and one action each; never as prose paragraphs. - Add a skill or agent rule only for a failure seen in a real run or eval; never for a hypothetical one. - Prefer editing or deleting an existing rule over adding a new one. - Before editing a SKILL.md or agents/*.md description:, apply each of these: - Write it in the third person: "Writes …", not "Write …". - Say…
In this repository16 other files it gives its agents.
- ask-teamSkill
- assemble-metaSkill
- brainstormSkill
- build-issueSkill
- cut-milestonesSkill
- frictionSkill
- milestone-closeSkill
- qa-reviewSkill
- review-benchSkill
- scope-commitSkill
- scope-issueSkill
- scope-mrSkill
- scope-specSkill
- setup-dream-teamSkill
- ship-itSkill
- triage-issueSkill
The file
Show its rules and text
We split this CLAUDE.md into the rules it gives an agent
24 rules. About 0 are common, 8 are distinctive, and 16 are very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈8
Distinctive · estimate
This author's own choices. Worth a look.
16
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Very specific: line 3
Write skill rules as atomic - bullets, one condition and one action each.
- Distinctive: line 5
Add a skill or agent rule only for a failure seen in a real run or eval.
- Distinctive: line 7
Prefer editing or deleting an existing rule over adding a new one.
- Very specific: line 8
Before editing a SKILL.md or agents/*.md description:, apply each of these:
- Distinctive: line 9
Write it in the third person: "Writes …", not "Write …".
- Distinctive: line 12
Keep it to 1,024 characters, with no XML tags.
- Very specific: line 13
Before adding a skill or agent, or changing a SKILL.md's other sections or files or an agents/*.md prompt, read …
- Distinctive: line 17
Keep rules out of README.md.
- Very specific: line 18
Keep standing contributor facts the code and paths don't show in MAINTAINING.md, one home each.
- Very specific: line 20
Update a MAINTAINING.md fact in the MR that changes it.
- Very specific: line 21
Update the skills/ask-team/WORKFLOW.md diagram in the same MR as a change to the flow it draws.
- Very specific: line 22
Put a change's detail the code doesn't show in the commit, a spec or a topic doc, never in a SKILL.md or agent prompt.
- Very specific: line 28
Run a skills/*/evals/ runner with --plugin-dir ..
- Very specific: line 34
Run the hook tests as python3 <file> each, never python3 -m unittest.
- Very specific: line 36
Call a plugin workflow by its prefixed name (dream-team:<name>), never by scriptPath.
- Very specific: line 38
Before editing CHANGELOG.md, read references/changelog-style.md.
- Very specific: line 43
Commit skills/*/evals/ (cases, fixtures, runner) except skills/scope-*/evals/.
- Very specific: line 45
Commit top-level evals/ except evals/results/.
- Very specific: line 46
Run the top-level evals/ cases with claude plugin eval .
- Distinctive: line 50
Before changing who does a step in a skill, read the skill's top-level role rules (e.g.
- Very specific: line 54
Check skill triggering with skills/scope-commit/evals/run_trigger.sh (unnamed prompts).
- Distinctive: line 59
Keep research findings mined from local session data (workflow journals, run counts) out of the repo and its issues.
- Distinctive: line 66
Before filing a follow-up issue, list the milestone's open issues.
- Very specific: line 68
Start an interactive session in this repo without --plugin-dir ..
≈ “Common” is an estimate: at least 1 in 5 of the 27,641 CLAUDE.md files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
CLAUDE.md
- Write skill rules as atomic
-bullets, one condition and one action each; never as prose paragraphs. - Add a skill or agent rule only for a failure seen in a real run or eval; never for a hypothetical one.
- Prefer editing or deleting an existing rule over adding a new one.
- Before editing a
SKILL.mdoragents/*.mddescription:, apply each of these:- Write it in the third person: "Writes …", not "Write …".
- Say what it does first, then when to use it.
- Name the key terms a request or command for it would contain.
- Keep it to 1,024 characters, with no XML tags.
- Before adding a skill or agent, or changing a
SKILL.md's other sections or files or anagents/*.mdprompt, read https://platform.claude.com/docs/en/agents-and-tools/agent-skills/best-practices.md and follow it. - Fix one failure with one rule; never add a rule for adjacent cases the issue did not hit.
- Keep rules out of
README.md; it points at the skills instead of restating them. - Keep standing contributor facts the code and paths don't show in
MAINTAINING.md, one home each. - Update a
MAINTAINING.mdfact in the MR that changes it. - Update the
skills/ask-team/WORKFLOW.mddiagram in the same MR as a change to the flow it draws. - Put a change's detail the code doesn't show in the commit, a spec or a topic doc, never in a
SKILL.mdor agent prompt. - Name a milestone
YY.MM, its target month. - Version milestones release in flow mode.
- From 2026-12-15 to the 26.12 due date, merge only
bugissues and changes that do not alter skill or agent behaviour (docs, evals, CI). - Run a
skills/*/evals/runner with--plugin-dir .; without the flag it measures the installed plugin cache, not the working tree. - In
hooks/test_body_cap.py, match an over-cap deny on-word ceiling, notceiling; the skill text the hook quotes back contains the bare word. - Judge a
scope-mrchange on in-repo runs too (REPO=<checkout>inrun_baseline.sh); diff-only runs lack the context real sessions read and hide what the skill does there. - Run the hook tests as
python3 <file>each, neverpython3 -m unittest; they are plain scripts and unittest exits 5 with NO TESTS RAN. - Call a plugin workflow by its prefixed name (
dream-team:<name>), never byscriptPath; skill text does not expand${CLAUDE_PLUGIN_ROOT}; a plugin agent body does. - Before editing
CHANGELOG.md, readreferences/changelog-style.md. - Fence a chat-output layout in a skill, so the model reproduces its markdown; show example skill output in docs or evals as rendered markdown, where a fence hides the bold labels judged.
- Lay out a skill's chat output for a monospace terminal: tight
- **Label:**bullets, no blank lines between items; paragraph spacing wastes fixed-height lines. - Commit
skills/*/evals/(cases, fixtures, runner) exceptskills/scope-*/evals/; those andruns/are ignored. - Commit top-level
evals/exceptevals/results/. - Run the top-level
evals/cases withclaude plugin eval . --trust-plugin --model opus --threshold 0.6. - Give a small or pass-through skill or workflow phase stub-agent tests only; add a paid headless-claude eval runner only for a large skill or a failure seen in a real run.
- Before changing who does a step in a skill, read the skill's top-level role rules (e.g. build-issue's "You do not write the change" edit list) and update them in the same change.
- A change to a
SKILL.md, itsreference/files or anagents/*.mdprompt → check every other line of those files for a contradiction with the change. - Check skill triggering with
skills/scope-commit/evals/run_trigger.sh(unnamed prompts); the other evals name the skill in the prompt and cannot see a description change. Keep its fixture repo consistent with each prompt, or the agent stops before loading any skill. - In a headless
claude -prun, pipe the prompt on stdin or put it before--allowed-tools; that flag is variadic and swallows a trailing prompt ("Input must be provided"). - Keep research findings mined from local session data (workflow journals, run counts) out of the repo and its issues; report them in chat only, even when build-issue's research tier says to post a findings comment.
- Delegate grading of eval reports against a ledger to a subagent (Agent tool); the main session only reads the scores and the marks file.
- Cite an outside pull request in commits, PRs, issues and notes in plain words ("unbound pull
request 1525"), never
#1525; GitHub links#Nto this repo. - Before filing a follow-up issue, list the milestone's open issues; a matching one gets a comment with the new evidence, not a new issue (#231 duplicated #226).
- Start an interactive session in this repo without
--plugin-dir .; with it, every Write and Edit here prompts even in accept-edits mode. Keep the flag for headless evals. - Give a headless
pi -prun without piped input< /dev/null; under a Claude Bash call it waits forever on the harness stdin socket, 0% CPU, empty output.
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

