agentic-delegation

Name: agentic-delegation
Author: Raishin

17stars3forks

How to install agentic-delegation?

raishin/vanguard-frontier-agentic/agentic-delegation

$npx -y skills add Raishin/vanguard-frontier-agentic --skill agentic-delegation

Installs into the current project.

Summary

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.

Files · 1

View on GitHub

SKILL.md

1	# Agentic Delegation
2
3	## Doctrine
4
5	Most tasks in this repo decompose into cheap, parallelizable work plus a small amount of
6	work that genuinely needs the orchestrator's judgment. Default to delegating the former.
7	Before doing multi-step work yourself, ask: can a cheaper model do this step just as well?
8
9	## a) Exploration and reconnaissance → Haiku
10
11	- Use the `Explore` agent type with `model: haiku` for read-only reconnaissance: locating
12	files, grepping for symbols, mapping call sites, summarizing existing structure.
13	- Scope each Explore task tightly — one question, one area of the tree. Do not send an
14	Explore agent an open-ended "understand the whole system" ask; split it into targeted
15	sweeps instead.
16	- Require file:line citations in every finding. A report without exact paths and line
17	numbers is not actionable — re-run it with a tighter prompt rather than accepting it.
18
19	## b) Bulk writing → Sonnet
20
21	- Route bulk writing — docs, guides, boilerplate, test scaffolding, repetitive multi-file
22	edits — to Sonnet subagents.
23	- Give each writing task a precise spec: exact file paths to create or edit, the content
24	shape expected, and which repo conventions to mirror (frontmatter shape, heading
25	structure, existing tone).
26	- Hard-constrain every writing delegate:
27	- Files it may touch — list them explicitly; nothing outside that list.
28	- Linters/gates it must pass — e.g. `npx markdownlint-cli2`, `codespell`, or the
29	schema/validation gate relevant to the files it is touching.
30	- No commits — delegates write files; only the orchestrator commits.
31
32	## Orchestrator requirements
33
34	- Haiku must never be the orchestrator. It explores and runs gates; it does not plan,
35	decompose, or accept work.
36	- When Sonnet is the orchestrator, run it at high reasoning effort at minimum — use the
37	harness's maximum-thinking mode where available. Planning and delegation quality degrade
38	below that, and a weak plan wastes every delegate downstream.
39	- The model split in this skill is unchanged by who orchestrates: even a Sonnet orchestrator
40	routes bulk writing to Sonnet subagents — the benefit is keeping the orchestrator's context
41	clean for judgment, not just the per-token price.
42
43	## c) What the orchestrator keeps
44
45	Never delegate:
46
47	- Architecture and design decisions (schema shapes, scope boundaries, precedence rules).
48	- Security-sensitive code (auth, secrets handling, trust-boundary logic).
49	- Surgical edits to load-bearing logic (validation gates, schemas, catalog generators).
50	- Final verification and the commit itself.
51
52	## d) Every delegate gets
53
54	- Exact file paths — absolute, not "somewhere in docs/".
55	- Acceptance criteria — what "done" looks like, stated concretely and checkably.
56	- An explicit "do NOT" list — files not to touch, commands not to run (no
57	`npm run validate`, no `cargo test`, no `git commit` inside a delegate unless
58	explicitly asked to run them for verification).
59
60	## e) Verify before accepting
61
62	- Run the repo's own gates on delegate output before treating it as done: `npm run
63	validate`, `cargo test` (for `tools/vfa-tui`), `npx markdownlint-cli2`, `codespell`.
64	- A delegate's self-report is not verification — read the diff, run the gate, then accept.
65
66	## Workflow templates
67
68	Three reusable orchestration shapes cover most multi-step tasks in this repo. Reach for one of
69	these before inventing a bespoke delegation plan.
70
71	### a) Recon sweep
72
73	Parallel Haiku `Explore` agents, one question each, citations required.
74
75	- Split the open-ended question into narrow, independent sub-questions — one per agent, one
76	area of the tree each.
77	- Launch all Explore agents in the same message so they run in parallel, not sequentially.
78	- Require file:line citations in every finding, same as section (a) above.
79	- When to use — you don't yet know where something lives, or need a map of an unfamiliar
80	area before deciding what to change.
81	- Hard constraints — read-only; Explore agents may not `Edit`/`Write`. No commits. If a
82	sweep comes back thin or off-target, re-run it with a tighter prompt rather than accepting a
83	vague report.
84
85	### b) Spec-driven implementation
86
87	Orchestrator writes an exact file-scoped spec, Sonnet implements, orchestrator reviews the diff
88	and runs decisive verification before accepting.
89
90	- Orchestrator writes the spec first: exact file paths, the content/code shape expected, which
91	repo conventions to mirror, and acceptance criteria stated concretely.
92	- Delegate the spec verbatim to a Sonnet subagent — do not compress it to a one-line ask; a
93	vague handoff produces a vague implementation.
94	- Orchestrator

People who liked this also viewed

6 picks

Type

agentic-delegation

17stars3forks

View on GitHub

How to install agentic-delegation?

raishin/vanguard-frontier-agentic/agentic-delegation

$npx -y skills add Raishin/vanguard-frontier-agentic --skill agentic-delegation

Installs into the current project.

Summary

Files · 1

View on GitHub

SKILL.md

1	# Agentic Delegation
2
3	## Doctrine
4
5	Most tasks in this repo decompose into cheap, parallelizable work plus a small amount of
6	work that genuinely needs the orchestrator's judgment. Default to delegating the former.
7	Before doing multi-step work yourself, ask: can a cheaper model do this step just as well?
8
9	## a) Exploration and reconnaissance → Haiku
10
11	- Use the `Explore` agent type with `model: haiku` for read-only reconnaissance: locating
12	files, grepping for symbols, mapping call sites, summarizing existing structure.
13	- Scope each Explore task tightly — one question, one area of the tree. Do not send an
14	Explore agent an open-ended "understand the whole system" ask; split it into targeted
15	sweeps instead.
16	- Require file:line citations in every finding. A report without exact paths and line
17	numbers is not actionable — re-run it with a tighter prompt rather than accepting it.
18
19	## b) Bulk writing → Sonnet
20
21	- Route bulk writing — docs, guides, boilerplate, test scaffolding, repetitive multi-file
22	edits — to Sonnet subagents.
23	- Give each writing task a precise spec: exact file paths to create or edit, the content
24	shape expected, and which repo conventions to mirror (frontmatter shape, heading
25	structure, existing tone).
26	- Hard-constrain every writing delegate:
27	- Files it may touch — list them explicitly; nothing outside that list.
28	- Linters/gates it must pass — e.g. `npx markdownlint-cli2`, `codespell`, or the
29	schema/validation gate relevant to the files it is touching.
30	- No commits — delegates write files; only the orchestrator commits.
31
32	## Orchestrator requirements
33
34	- Haiku must never be the orchestrator. It explores and runs gates; it does not plan,
35	decompose, or accept work.
36	- When Sonnet is the orchestrator, run it at high reasoning effort at minimum — use the
37	harness's maximum-thinking mode where available. Planning and delegation quality degrade
38	below that, and a weak plan wastes every delegate downstream.
39	- The model split in this skill is unchanged by who orchestrates: even a Sonnet orchestrator
40	routes bulk writing to Sonnet subagents — the benefit is keeping the orchestrator's context
41	clean for judgment, not just the per-token price.
42
43	## c) What the orchestrator keeps
44
45	Never delegate:
46
47	- Architecture and design decisions (schema shapes, scope boundaries, precedence rules).
48	- Security-sensitive code (auth, secrets handling, trust-boundary logic).
49	- Surgical edits to load-bearing logic (validation gates, schemas, catalog generators).
50	- Final verification and the commit itself.
51
52	## d) Every delegate gets
53
54	- Exact file paths — absolute, not "somewhere in docs/".
55	- Acceptance criteria — what "done" looks like, stated concretely and checkably.
56	- An explicit "do NOT" list — files not to touch, commands not to run (no
57	`npm run validate`, no `cargo test`, no `git commit` inside a delegate unless
58	explicitly asked to run them for verification).
59
60	## e) Verify before accepting
61
62	- Run the repo's own gates on delegate output before treating it as done: `npm run
63	validate`, `cargo test` (for `tools/vfa-tui`), `npx markdownlint-cli2`, `codespell`.
64	- A delegate's self-report is not verification — read the diff, run the gate, then accept.
65
66	## Workflow templates
67
68	Three reusable orchestration shapes cover most multi-step tasks in this repo. Reach for one of
69	these before inventing a bespoke delegation plan.
70
71	### a) Recon sweep
72
73	Parallel Haiku `Explore` agents, one question each, citations required.
74
75	- Split the open-ended question into narrow, independent sub-questions — one per agent, one
76	area of the tree each.
77	- Launch all Explore agents in the same message so they run in parallel, not sequentially.
78	- Require file:line citations in every finding, same as section (a) above.
79	- When to use — you don't yet know where something lives, or need a map of an unfamiliar
80	area before deciding what to change.
81	- Hard constraints — read-only; Explore agents may not `Edit`/`Write`. No commits. If a
82	sweep comes back thin or off-target, re-run it with a tighter prompt rather than accepting a
83	vague report.
84
85	### b) Spec-driven implementation
86
87	Orchestrator writes an exact file-scoped spec, Sonnet implements, orchestrator reviews the diff
88	and runs decisive verification before accepting.
89
90	- Orchestrator writes the spec first: exact file paths, the content/code shape expected, which
91	repo conventions to mirror, and acceptance criteria stated concretely.
92	- Delegate the spec verbatim to a Sonnet subagent — do not compress it to a one-line ask; a
93	vague handoff produces a vague implementation.
94	- Orchestrator

People who liked this also viewed

6 picks

Type