agentic-workflows / templates
CLAUDE.md1 starsChanged 6 months agoMITView on GitHub
Delegation is OPT-IN: only on explicit request ("use codex", "have gemini research", "@codex-delegate"). NEVER auto-delegate during normal work. Don't grep the repo or improvise flags — they're below. Do NOT use: <models the user flagged untrusted/never>. CONSTRAINT: dynamic Agent-tool spawns accept only current aliases; to run a subagent on a versioned model, PIN the full id in an agent file. Wrapper subagents (@codex-delegate, @gemini-research, @model-delegate) are pinned to <cheapest proven Claude> — they only craft prompts + summarize.
The file
We split this CLAUDE.md into the rules it gives an agent
2 rules. About 0 are common, 2 are distinctive, and 0 are very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈2
Distinctive · estimate
This author's own choices. Worth a look.
0
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Distinctive: line 22
Do NOT use: <models the user flagged untrusted/never>.
- Distinctive: line 29
When axes conflict for anything that ships: intelligence > taste > cost.
≈ “Common” is an estimate: at least 1 in 5 of the 26,980 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
Model orchestration — right model for the right job
TEMPLATE: fill the <PLACEHOLDERS> from discovery + the user's answers.
Place at ~/.claude/CLAUDE.md (global) or <project>/.claude/CLAUDE.md (scoped).
Delegation is OPT-IN: only on explicit request ("use codex", "have gemini research", "@codex-delegate"). NEVER auto-delegate during normal work. Don't grep the repo or improvise flags — they're below.
Roster (this machine's available + approved models)
Calibrate cost to what the USER actually pays/limits (not list price). State
whether higher means cheaper or pricier and be consistent.
| Model | cost | intel | taste | Reached via | Right job |
| <cheap-bulk model, e.g. GPT via codex> | <n> | <n> | <n> | codex exec | bulk, mechanical, clear-spec impl, data, investigation |
| <cheap proven Claude, e.g. a specific version> | <n> | <n> | <n> | Agent model: (pin full id) | light Claude-flavored work, wrapper orchestration |
<strong Claude> | <n> | <n> | <n> | main loop / Agent model: | hard reasoning, orchestration |
| <premium/taste model> | <n> | <n> | <n> | Agent model: | UI/copy/API design, final polish, deep reviews |
| <big-context model, e.g. Gemini> | — | — | — | gemini -p / llm | research, large-context, web |
| <local model, if any> | free | <n> | <n> | ollama run | private/offline bulk |
Do NOT use: <models the user flagged untrusted/never>. CONSTRAINT: dynamic Agent-tool spawns accept only current aliases; to run a subagent on a versioned model, PIN the full id in an agent file.
How to apply (defaults, not limits)
- Judge the OUTPUT, not the price. If a cheaper model underdelivers, rerun on a stronger one without asking — escalating beats shipping mediocre work.
- When axes conflict for anything that ships: intelligence > taste > cost.
- Bulk / mechanical / clear-spec →
<cheapest capable>. - Light Claude-flavored work →
<cheap proven Claude>. - User-facing (UI, copy, API design) or top-quality → <premium/taste model>.
- Hardest / most ambiguous reasoning →
<highest-intelligence model>. - Reviews of plans/impl → two strong models; optionally one other-provider lens.
- Research / read-everything / web →
<big-context model>.
Verified invocations (external CLIs; close stdin or they hang)
<cheap-bulk>: codex exec -s read-only "<prompt>" </dev/null (workspace-write ONLY when the user authorizes edits · codex exec review)<big-context>: gemini -p "<self-contained query>" </dev/null- other models: llm -m
<model>"<prompt>" </dev/null - local: ollama run
<model>"<prompt>" - Parallel research → one process per topic, redirect each to its own file:
gemini -p "
<bounded query>" </dev/null ><dir>/<topic>.md Bound every prompt to the project's actual architecture + invariants. - Windows cmd uses
<nulinstead of</dev/null.
Mechanics
- Non-Claude models are reached ONLY via their CLIs. In subagents the model param takes Claude models only, so a non-Claude model runs through a thin wrapper that shells out to the CLI and returns the result (see the delegate wrappers).
- Claude models run via the Agent/subagent
modelparameter (alias, or full id when pinned in an agent file).
Wrapper subagents (@codex-delegate, @gemini-research, @model-delegate) are
pinned to <cheapest proven Claude> — they only craft prompts + summarize.
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.