agentic-delegation
Skill23 starsChanged 4 months agoApache-2.0View on GitHub
Delegate exploration sweeps to Haiku subagents and bulk writing to Sonnet subagents while the orchestrator keeps architecture, security-sensitive edits, and commits; use at the start of any multi-step task in this repo to minimize token spend by delegating to cheaper models.
In this repository15 other files it gives its agents.
- agents/AGENTS.mdAGENTS.md
- AGENTS.mdAGENTS.md
- catalog/AGENTS.mdAGENTS.md
- docs/AGENTS.mdAGENTS.md
- mcp/AGENTS.mdAGENTS.md
- rules/AGENTS.mdAGENTS.md
- schemas/AGENTS.mdAGENTS.md
- skills/AGENTS.mdAGENTS.md
- templates/AGENTS.mdAGENTS.md
- tests/AGENTS.mdAGENTS.md
- CLAUDE.mdCLAUDE.md
- .github/copilot-instructions.mdCopilot instructions
- definition-of-doneSkill
- model-registry-refreshSkill
- pr-babysitSkill
The file
We split this skill into the rules it gives an agent
4 rules. About 0 are common, 1 is distinctive, and 3 are very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈1
Distinctive · estimate
This author's own choices. Worth a look.
3
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Distinctive: line 13
Before doing multi-step work yourself, ask: can a cheaper model do this step just as well?
- Very specific: line 17
Use the Explore agent type with model: haiku for read-only reconnaissance: locating files, grepping for symbols, mapping call sites, …
- Very specific: line 68
Run the repo's own gates on delegate output before treating it as done: npm run validate, cargo test (for tools/vfa-tui), npx …
- Very specific: line 102
Run the gate(s) relevant to the touched files (schema validation, npm run validate, cargo test, linters) and treat a pass as necessary, not …
≈ “Common” is an estimate: at least 1 in 5 of the 372,429 skill 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
Agentic Delegation
Doctrine
Most tasks in this repo decompose into cheap, parallelizable work plus a small amount of work that genuinely needs the orchestrator's judgment. Default to delegating the former. Before doing multi-step work yourself, ask: can a cheaper model do this step just as well?
a) Exploration and reconnaissance → Haiku
- Use the
Exploreagent type withmodel: haikufor read-only reconnaissance: locating files, grepping for symbols, mapping call sites, summarizing existing structure. - Scope each Explore task tightly — one question, one area of the tree. Do not send an Explore agent an open-ended "understand the whole system" ask; split it into targeted sweeps instead.
- Require file:line citations in every finding. A report without exact paths and line numbers is not actionable — re-run it with a tighter prompt rather than accepting it.
b) Bulk writing → Sonnet
- Route bulk writing — docs, guides, boilerplate, test scaffolding, repetitive multi-file edits — to Sonnet subagents.
- Give each writing task a precise spec: exact file paths to create or edit, the content shape expected, and which repo conventions to mirror (frontmatter shape, heading structure, existing tone).
- Hard-constrain every writing delegate:
- Files it may touch — list them explicitly; nothing outside that list.
- Linters/gates it must pass — e.g.
npx markdownlint-cli2,codespell, or the schema/validation gate relevant to the files it is touching. - No commits — delegates write files; only the orchestrator commits.
Orchestrator requirements
- Haiku must never be the orchestrator. It explores and runs gates; it does not plan, decompose, or accept work.
- When Sonnet is the orchestrator, run it at high reasoning effort at minimum — use the harness's maximum-thinking mode where available. Planning and delegation quality degrade below that, and a weak plan wastes every delegate downstream.
- The model split in this skill is unchanged by who orchestrates: even a Sonnet orchestrator routes bulk writing to Sonnet subagents — the benefit is keeping the orchestrator's context clean for judgment, not just the per-token price.
c) What the orchestrator keeps
Never delegate:
- Architecture and design decisions (schema shapes, scope boundaries, precedence rules).
- Security-sensitive code (auth, secrets handling, trust-boundary logic).
- Surgical edits to load-bearing logic (validation gates, schemas, catalog generators).
- Final verification and the commit itself.
d) Every delegate gets
- Exact file paths — absolute, not "somewhere in docs/".
- Acceptance criteria — what "done" looks like, stated concretely and checkably.
- An explicit "do NOT" list — files not to touch, commands not to run (no
npm run validate, nocargo test, nogit commitinside a delegate unless explicitly asked to run them for verification).
e) Verify before accepting
- Run the repo's own gates on delegate output before treating it as done:
npm run validate,cargo test(fortools/vfa-tui),npx markdownlint-cli2,codespell. - A delegate's self-report is not verification — read the diff, run the gate, then accept.
Workflow templates
Three reusable orchestration shapes cover most multi-step tasks in this repo. Reach for one of these before inventing a bespoke delegation plan.
a) Recon sweep
Parallel Haiku Explore agents, one question each, citations required.
- Split the open-ended question into narrow, independent sub-questions — one per agent, one area of the tree each.
- Launch all Explore agents in the same message so they run in parallel, not sequentially.
- Require file:line citations in every finding, same as section (a) above.
- When to use — you don't yet know where something lives, or need a map of an unfamiliar area before deciding what to change.
- Hard constraints — read-only; Explore agents may not
Edit/Write. No commits. If a sweep comes back thin or off-target, re-run it with a tighter prompt rather than accepting a vague report.
b) Spec-driven implementation
Orchestrator writes an exact file-scoped spec, Sonnet implements, orchestrator reviews the diff and runs decisive verification before accepting.
- Orchestrator writes the spec first: exact file paths, the content/code shape expected, which repo conventions to mirror, and acceptance criteria stated concretely.
- Delegate the spec verbatim to a Sonnet subagent — do not compress it to a one-line ask; a vague handoff produces a vague implementation.
- Orchestrator reads the resulting diff in full before running any gate — do not skip straight to "did the gate pass."
- Run the gate(s) relevant to the touched files (schema validation,
npm run validate,cargo test, linters) and treat a pass as necessary, not sufficient, for acceptance. - When to use — the shape of the change is fully known up front (new file, defined edit to an existing one) and doesn't require architectural judgment mid-implementation.
- Hard constraints — files it may touch: exactly the list in the spec, nothing else. No commits — the orchestrator commits after review.
c) Gate run
Haiku runs the full repo gate suite and reports pass/fail with raw failure output.
- Delegate to Haiku:
cargo fmt --check,cargo clippy -- -D warnings,cargo test(fortools/vfa-tui),npm run validate,codespell,npx markdownlint-cli2, thennpm run asset-integrity:writelast, only after every other gate is green. Regenerating integrity before other generators finish stales the manifest — see the ordering caveat inCLAUDE.md/AGENTS.md. - Require raw failure output verbatim in the report — not a paraphrase like "some tests failed." The orchestrator needs the actual error to decide the next move.
- When to use — verifying a change is ready before the orchestrator reviews/commits, or a periodic health check with no code changes attached.
- Hard constraints — this is a read/verify pass: the only file it may write is
catalog/asset-integrity.jsonviaasset-integrity:write, and only after all other gates pass. No other edits. No commits — report results back to the orchestrator, who decides whether to fix, re-run, or commit.
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.

