claude-handoff

Name: claude-handoff
Author: PaulRBerg

68stars3forks

How to install claude-handoff?

paulrberg/agent-skills/claude-handoff

$npx -y skills add PaulRBerg/agent-skills --skill claude-handoff

Installs into the current project.

Summary

Orchestrate one to five Sonnet subagents to implement an approved Claude Code plan.

Files · 1

View on GitHub

SKILL.md

1	# Claude Handoff
2
3	Plan in Claude Code with the session's planning model, then hand the approved implementation to one to five Sonnet
4	subagents in the sequence the task requires. Subagents run in-session through the Agent tool with the host session's
5	permissions, so they can write anywhere the session can. Use this skill to reserve the expensive planning model for
6	thinking, orchestration, and verification while cheaper Sonnet agents do the implementation.
7
8	## Contract
9
10	- Run only after the user explicitly invokes this skill in Plan mode. If Plan mode is not active, ask the user to switch
11	and stop.
12	- Claude owns discovery, decisions, the implementation plan, and agent orchestration. Do not consult subagents while
13	planning.
14	- Each subagent implements its assigned part of the approved plan. It may inspect, edit, and validate, but must not
15	redesign the solution or return another plan.
16	- Use the smallest effective team. One agent remains valid; use additional agents only when decomposition materially
17	improves latency, correctness, or verification. Never exceed five agents total across the handoff.
18	- Keep Claude's implementation work to orchestration, integrity checks, failure handling, and the conditional polish
19	pass.
20
21	Use `$ARGUMENTS` as the task when present; otherwise use the active user request.
22
23	## Plan Phase
24
25	Produce a decision-complete plan with this section:
26
27	```markdown
28	## Claude Handoff
29
30	- Strategy: `<sequential\|parallel\|hybrid>`
31	- Agents: `<1-5>` — `<why this is the smallest effective count>`
32	- Validation owner: `<agent-id\|claude>` — `<aggregate checks it runs once>`
33
34	\| Agent \| Wave \| Depends on \| Scope \| Implementation brief \| Completion evidence \|
35	\| ----- \| ---- \| ---------- \| ------------------ \| ------------------------------------------------------ \| ----------------------------------- \|
36	\| `A1` \| `1` \| `none` \| `<files/behavior>` \| `<outcome, edits, constraints, and stopping criteria>` \| `<commands and observable results>` \|
37
38	- Code polish: `<required\|not required>` — `<reason>`
39	```
40
41	Choose the execution shape from repository evidence and the approved work:
42
43	- Use sequential agents when one agent depends on another, their write scopes overlap, or a later agent owns integration
44	or aggregate validation.
45	- Use parallel agents only for independent work with explicitly disjoint write scopes. Agents may inspect shared
46	context, but must not write outside their assigned scope.
47	- Use hybrid execution for dependency-ordered waves: run independent agents within a wave in parallel, reconcile the
48	entire wave, then start its dependents.
49
50	A wave finishes with its slowest agent. Keep every agent's scope focused and move deferrable validation to the
51	validation owner.
52
53	The five-agent limit applies to the entire handoff, not each wave. Assign every agent a stable ID, exact dependencies,
54	an implementation scope, and its own stopping criteria. If parallel work does not collectively prove the overall plan,
55	reserve a later sequential agent for integration and aggregate validation.
56
57	Assign aggregate validation to exactly one owner per handoff: package-wide or repo-wide checks (full test suites,
58	whole-package typecheck or lint, catalog-wide checks) run once — by the integration agent when one exists, otherwise by
59	Claude during post-wave reconciliation. Every other agent's completion evidence must be the narrowest checks that prove
60	its own edits: file-scoped lint, format, or typecheck plus targeted tests for the files it touched. Duplicate aggregate
61	runs across a wave's agents are wasted wall-clock time, not extra assurance.
62
63	Every agent runs on the latest Sonnet model through the `general-purpose` subagent type; the Agent tool exposes no
64	per-agent effort or timeout controls, so balance a wave by decomposing scope, not by tuning configuration.
65
66	Require `$code-polish` for nonlocal invariants, concurrency or state machines, migrations or parsing, auth or security,
67	retry or error semantics, and public API or data-contract changes. File count alone is not a trigger.
68
69	Do not spawn subagents until the user approves the plan and Claude leaves Plan mode.
70
71	## Execution Phase
72
73	### Launch
74
75	Launch each agent with the Agent tool: `subagent_type: "general-purpose"`, `model: "sonnet"`, and a description like
76	`A1/3: <scope>`. Start every agent in a parallel wave in the same message as parallel tool calls; start sequential
77	agents only after reconciling their dependencies. Claude Code renders subagent progress natively — do not build bespoke
78	progress dashboards, polling loops, or status tables around the calls. After launc

People who liked this also viewed

6 picks

Type

find-skillsHelps users discover and install agent skills when they ask questions like "how do I do X", "find a skill for X", "is there a skill that can...", or express interest in extending capabilities. This skSkillsJul 20262.6M27k
grill-meA relentless interview to sharpen a plan or design.SkillsJul 2026633k182k
vercel-react-best-practicesVercel's official collection of agent skillsSkillsJul 2026573k29k
improve-codebase-architectureScan a codebase for deepening opportunities, present them as a visual HTML report, then grill through whichever one you pick.SkillsJul 2026522k182k
setup-matt-pocock-skillsConfigure this repo for the engineering skills — set up its issue tracker, triage label vocabulary, and domain doc layout. Run once before first use of the other engineering skills.SkillsJul 2026454k182k
triageMove issues and external PRs through a state machine of triage roles — categorise, verify, grill if needed, and write agent-ready briefs.SkillsJul 2026424k182k

claude-handoff

68stars3forks

View on GitHub

How to install claude-handoff?

paulrberg/agent-skills/claude-handoff

$npx -y skills add PaulRBerg/agent-skills --skill claude-handoff

Installs into the current project.

Summary

Orchestrate one to five Sonnet subagents to implement an approved Claude Code plan.

Files · 1

View on GitHub

SKILL.md

1	# Claude Handoff
2
3	Plan in Claude Code with the session's planning model, then hand the approved implementation to one to five Sonnet
4	subagents in the sequence the task requires. Subagents run in-session through the Agent tool with the host session's
5	permissions, so they can write anywhere the session can. Use this skill to reserve the expensive planning model for
6	thinking, orchestration, and verification while cheaper Sonnet agents do the implementation.
7
8	## Contract
9
10	- Run only after the user explicitly invokes this skill in Plan mode. If Plan mode is not active, ask the user to switch
11	and stop.
12	- Claude owns discovery, decisions, the implementation plan, and agent orchestration. Do not consult subagents while
13	planning.
14	- Each subagent implements its assigned part of the approved plan. It may inspect, edit, and validate, but must not
15	redesign the solution or return another plan.
16	- Use the smallest effective team. One agent remains valid; use additional agents only when decomposition materially
17	improves latency, correctness, or verification. Never exceed five agents total across the handoff.
18	- Keep Claude's implementation work to orchestration, integrity checks, failure handling, and the conditional polish
19	pass.
20
21	Use `$ARGUMENTS` as the task when present; otherwise use the active user request.
22
23	## Plan Phase
24
25	Produce a decision-complete plan with this section:
26
27	```markdown
28	## Claude Handoff
29
30	- Strategy: `<sequential\|parallel\|hybrid>`
31	- Agents: `<1-5>` — `<why this is the smallest effective count>`
32	- Validation owner: `<agent-id\|claude>` — `<aggregate checks it runs once>`
33
34	\| Agent \| Wave \| Depends on \| Scope \| Implementation brief \| Completion evidence \|
35	\| ----- \| ---- \| ---------- \| ------------------ \| ------------------------------------------------------ \| ----------------------------------- \|
36	\| `A1` \| `1` \| `none` \| `<files/behavior>` \| `<outcome, edits, constraints, and stopping criteria>` \| `<commands and observable results>` \|
37
38	- Code polish: `<required\|not required>` — `<reason>`
39	```
40
41	Choose the execution shape from repository evidence and the approved work:
42
43	- Use sequential agents when one agent depends on another, their write scopes overlap, or a later agent owns integration
44	or aggregate validation.
45	- Use parallel agents only for independent work with explicitly disjoint write scopes. Agents may inspect shared
46	context, but must not write outside their assigned scope.
47	- Use hybrid execution for dependency-ordered waves: run independent agents within a wave in parallel, reconcile the
48	entire wave, then start its dependents.
49
50	A wave finishes with its slowest agent. Keep every agent's scope focused and move deferrable validation to the
51	validation owner.
52
53	The five-agent limit applies to the entire handoff, not each wave. Assign every agent a stable ID, exact dependencies,
54	an implementation scope, and its own stopping criteria. If parallel work does not collectively prove the overall plan,
55	reserve a later sequential agent for integration and aggregate validation.
56
57	Assign aggregate validation to exactly one owner per handoff: package-wide or repo-wide checks (full test suites,
58	whole-package typecheck or lint, catalog-wide checks) run once — by the integration agent when one exists, otherwise by
59	Claude during post-wave reconciliation. Every other agent's completion evidence must be the narrowest checks that prove
60	its own edits: file-scoped lint, format, or typecheck plus targeted tests for the files it touched. Duplicate aggregate
61	runs across a wave's agents are wasted wall-clock time, not extra assurance.
62
63	Every agent runs on the latest Sonnet model through the `general-purpose` subagent type; the Agent tool exposes no
64	per-agent effort or timeout controls, so balance a wave by decomposing scope, not by tuning configuration.
65
66	Require `$code-polish` for nonlocal invariants, concurrency or state machines, migrations or parsing, auth or security,
67	retry or error semantics, and public API or data-contract changes. File count alone is not a trigger.
68
69	Do not spawn subagents until the user approves the plan and Claude leaves Plan mode.
70
71	## Execution Phase
72
73	### Launch
74
75	Launch each agent with the Agent tool: `subagent_type: "general-purpose"`, `model: "sonnet"`, and a description like
76	`A1/3: <scope>`. Start every agent in a parallel wave in the same message as parallel tool calls; start sequential
77	agents only after reconciling their dependencies. Claude Code renders subagent progress natively — do not build bespoke
78	progress dashboards, polling loops, or status tables around the calls. After launc

People who liked this also viewed

6 picks

Type