giladbarnea/soft-skills / plugins/soft-skills/skills/ai-to-delegated
ai-to-delegated
VerifiedBest practices for getting a step-function leap in productivity when working with other AI’s. Load `ai-to-delegated` before spawning a sub-agent, creating a team, talking to a teammate agent, or writing a doc likely read by other agents. Trigger words include “dispatch an agent”, “spawn a team”, and “fork the session to ...”.
| File | Kind | Size |
|---|---|---|
| SKILL.md | skill md | 2.9 KB |
| briefing/examples.md | markdown | 12.7 KB |
| briefing/forked-context.md | markdown | 613 B |
| briefing/fresh-context.md | markdown | 1.6 KB |
| configurations/colloquium.md | markdown | 2.2 KB |
| configurations/external-advisor.md | markdown | 1.1 KB |
| configurations/hillclimber-watcher/protocol.md | markdown | 4.0 KB |
| configurations/hillclimber-watcher/watcher.md | markdown | 1.5 KB |
| configurations/pair-programming.md | markdown | 5.1 KB |
| configurations/research-fanout.md | markdown | 2.8 KB |
| coordination/budget-balancing.md | markdown | 1.0 KB |
| coordination/delegates.md | markdown | 4.9 KB |
| coordination/leading-leaders.md | markdown | 4.2 KB |
| coordination/many-hours-supervision.md | markdown | 3.6 KB |
| coordination/peers.md | markdown | 2.4 KB |
| planning/choose-configuration.md | markdown | 7.9 KB |
| planning/dispatch-outside-the-harness.md | markdown | 1.3 KB |
| planning/when-to-delegate.md | markdown | 8.7 KB |
planning/choose-configuration.md
View on GitHub---
name: choose-configuration
description: Choose the delegation shape, concurrency, context, model, and thinking level for the work.
---
# Choose a Delegation Configuration
Read this when choosing or changing a configuration. Preserve choices already made by your leader.
If you are still deciding whether delegation helps, read [when-to-delegate.md](planning/when-to-delegate.md).
## Delegation parameters
Decision matrix:
1. **Delegation shape:** subagent or team?
2. **Concurrency:** subagents: parallel subagents or a single subagent? Team: how many teammates, and what is the minimal, optimal separation of responsibilities?
3. **Context:** inherit the session’s context window or start fresh? See [Fork or fresh](#fork-or-fresh).
4. **Model:** which model?
5. **Thinking:** which thinking level? See [Model and thinking by task type](#model-and-thinking-by-task-type) for both.
## Fork or fresh
Fork by default. A fork needs no brief beyond the assignment, which saves your own context window, and it is guaranteed to understand the problem space and the intended solution. Start fresh in two cases:
1. You want zero bias: an independent review, an external advisor.
2. The session you would fork is bloated. Two tests, which usually agree:
1. **Runway.** Context windows are finite. They typically end at 1M. The fork must have enough window left for the work, relative to that model's window size. Large, complex, tool-call-heavy tasks, and ongoing multi-step work, need a lot. Its token headroom is simply the fork's model's context window size minus your context window usage at the time of forking. A safe runway is 500k free tokens for the fork. If the task is large, it may need arbitrarily more. Use your judgment. Certainly, 200k is too little. 400k is enough for _some_ tasks, usually bounded ones.
2. **Performance.** Past a size, models lose some performance and focus. Rough numbers as of October 2026: big and flagship models become a little less sharp at 600k. Medium models: 300k to 400k. Small models: 200k to 300k. Note: “a little less sharp” does not mean the model is malfunctioning, and does not mean it needs replacement. It means exactly and only this — a weight to balance in the context management debate.
When a fork would fail either test, apply [briefing a fresh-context](briefing/fresh-context.md) and give it to a fresh agent. Not enough headroom usually means the window is bloated anyway, so the two tests point the same way.
## Model and thinking by task type
Three tiers. Confirm the pick with your leader. Model names are examples as of October 2026. If your harness cannot set the combination you chose, read [dispatch outside the harness](planning/dispatch-outside-the-harness.md).
| Tier | Task | Model | Thinking |
| --- | --- | --- | --- |
| Judgment | Making decisions, judging what matters more, deep understanding and expertise, large knowledge space, nuance, connection-mapping and synthesis (the bigger picture), or managing agents | Big or flagship | ≥ high |
| Implementation | Decisions already made, judgment and vision clear; the work only needs implementing | Medium, or big for harder cases. Not flagship | Medium: ≥ high. Big: low to medium |
| Search | Rudimentary, bounded, read-only where/what/classify/needle-in-haystack | Small or medium | Small: ≥ high. Medium: low |
**Big vs. flagship, in the judgment tier.** Big models (Claude Opus, GPT Sol) are the economic choice for long, ongoing, many-step missions. They are very capable workhorses even if not the absolute cutting edge. Flagship models (Claude Fable, GPT Astra) fit bounded tasks where the more intelligence the better: an unbiased review, an external advisor, root cause analysis of a complex issue, navigating multi-faceted tradeoffs.
**Medium vs. big, in the implementation tier.** Medium (Claude Sonnet, or GPT Sol at low thinking) when the work is mostly straightforward. Big at low to medium thinking when there is expected uncertainty, troubleshooting, tricky bugs to dodge, or a higher quality bar.
**Small vs. medium, in the search tier.** Small (Claude Haiku, GPT Luna) at ≥ high thinking, or a medium model at low thinking. Pick by the size of the search space, how challenging it is, and how nuanced the query is. GPT Luna at ≥ high thinking, even xhigh, counts as a good medium-sized searcher.
**After dispatch.** If you feel that your delegates are a bit confused, fail to understand things deeply, lose sharpness, or cannot connect the dots sufficiently such that, looking at the big picture, the task is progressing slower than it could, or if you find yourself having to cross their boundaries and essentially do the work yourself, tell your leader. The fix is probably to upgrade the configuration. Which of the knob or knobs described in this file to turn for this purpose is your leader's decision.
## Subagents and Teams are two different things
**The difference:**
- Sub-agents are isolated from each other and report only to you;
- Teammates talk amongst themselves in real time, between tool calls, without routing through you.
Each has its own use cases and advantages.
**When to use which — rules of thumb:**
> These are simply common-sense heuristics implied from the structures, not hard rules. Use your judgment.
- Spawn a single *sub-agent* when you are the main context owner, and you could offload a bounded task to keep your own context window focused and devoid of D-tours.
- Spawn *multiple parallel sub-agents* when you are the main context owner, and when a wide task fans out horizontally into independent threads and you expect to do the synthesis yourself — i.e. when there is no special reason for the sub-agents to exchange findings and opinions before reporting back to you. The offloading argument in the single sub-agent case applies here as well, just in a distributed form.
- Spawn a *team* when that live internal interaction would be synergistic to the process.
**A team is a superset of the parallel sub-agent structure:**
A team = concurrent sub-agents + live communication. This unlocks a deeper value, because unlike sub-agents, the team can perform the work *you* would have done otherwise, before reporting back to you: the team can synthesize their own findings, adverserially review each other’s work and converge on a consensus, brainstorm ideas and come back with a lean plan, share issues and unblock each other, and so on.
Therefore, use teams to lift you up to a decision-making level, rather than a task-execution level. This has its tradeoffs, but it is a powerful tool when used judiciously.
## Configuration examples
These are examples to generalize from, not an exhaustive list or required structures. Read only the configuration relevant to the work.
| Situation | Configuration |
| --- | --- |
| Independent research tasks, followed by synthesis | [Research fanout](configurations/research-fanout.md) |
| Adjacent research efforts benefit from live exchange | [Colloquium](configurations/colloquium.md) |
| Two peers agree on a plan, then implement and review once | [Pair programming](configurations/pair-programming.md) |
| An optimizer needs outside judgment while pursuing a measurable target | [Hillclimber–watcher](configurations/hillclimber-watcher/protocol.md) |
| Your delegates will lead their own agents | Also read [leading-leaders.md](coordination/leading-leaders.md) |
## Before dispatch
Read [delegate coordination](coordination/delegates.md) before dispatching.
If the recipient does not share your context, read [fresh-context briefing](briefing/fresh-context.md).
Since teammates talk to each other, tell each of them to load [peer coordination](coordination/peers.md), alongside the project context they need.
If you are spawning an adversary among them, tell it to load the [peer-review skill](peer-review/SKILL.md) too.
Agents and teams can take a long time to run. Give a generous timeout.
For an illustration of choosing a team, see the [implementer–reviewer example](briefing/examples.md#choosing-an-implementerreviewer-team).
Related skills
| Skill | Stars | Updated | Works with | Status | Install |
|---|---|---|---|---|---|
| reverse-engineer-anythingmorluto/reaReverse engineer native, managed, Electron/JavaScript, packaged, firmware, and browser targets with REA, and build agent environments (AgentEnv bundles) of web applications from a live instance or a capture, and of Linux desktop apps with a task from a request. Use shipped-artifact or requested runtime evidence to explain features, compare versions, decompile code, or guide a reconstruction. Skip REA for ordinary source-repository architecture analysis. | Stars: 472k | Updated | Claude CodeCodex | Verified | Install |
| diagnosing-superpowersobra/superpowersUse when a superpowers session went wrong and your human partner wants to know why — repeated work, ignored plans, stumbles, poor results, a skill that didn't fire, "it took too long", "why is it so expensive", "what is it doing" — or wants to build a bug report for the superpowers maintainers, for the current session or a past one identified by id or path, on any harness. | Stars: 323k | Updated | Claude CodeCodex | Verified | Install |
| dispatching-parallel-agentsobra/superpowersUse when facing 2+ independent tasks that can be worked on without shared state or sequential dependencies | Stars: 323k | Updated | Claude CodeCodex | Verified | Install |
| executing-plansobra/superpowersUse when executing an implementation plan in the current session as the implementer yourself — your human partner chose inline execution, or no subagent tool is available | Stars: 323k | Updated | Claude CodeCodex | Verified | Install |