Purpose
The source announcement describes projects that move "from folder to conversation": you describe a goal
and Claude manages the work across multiple sessions instead of you splitting tasks, managing handoffs,
and merging results. This skill turns that model into a setup-and-steering procedure.
When to use
- Work spans several repos or many independent sub-tasks that can run in parallel.
- The effort runs for days or weeks and decisions must persist (shared memory).
- You want to delegate at the level of outcomes, not individual prompts.
Instructions
- Write the goal as an outcome. Pick one measurable goal and the repos or documents it needs.
Claude then suggests work it can start immediately. Use
templates/coordinator-brief.md.
- Configure the project. Cloud environment, connectors, plugins, project instructions, and the model
and effort level for both the coordinator and the worker threads.
- Brief the coordinator like a chief of staff. State priorities, constraints, who to consult, and what
"done" means. The coordinator routes tasks to new or existing threads.
- Let threads do the work. Each thread is a Claude Code cloud session on its own branch and copy of the
repo. With repos connected, threads open PRs and run tests; with documents, they read and draft.
Threads may split work further with subagents, loops, and workflows.
- Steer at the right level. Follow progress in the main project chat; open a single thread only to
steer details. You can steer from your phone, and work continues after you step away.
- Treat overlap as normal PR work. Overlapping changes surface as merge conflicts, as with any PR.
Ask the coordinator which PRs must merge first.
- Use shared memory instead of heavy prompt engineering. Every thread adds to and draws from it (e.g.
a release moved to Friday, why an export was dropped, who to consult before changing billing). Put
reusable files in the library; artifacts Claude produces land there too.
- Tune the working style. Adjust how often Claude checks in, how often it starts new threads, and how
detailed its updates are.
- Watch usage. Several full sessions run at once, so limits can be reached faster. Check
project-specific usage and lower model/effort for worker threads when appropriate.
Availability and limits as of the announcement are in
references/availability-and-limits.md.
Examples
Latency goal
Goal: reduce checkout p75 latency. Repo: checkout service.
Claude profiles each endpoint, tests optimizations, and opens PRs in parallel threads.
Endpoint retirement across repos
Connect the API, web, and mobile repos. Goal: retire the deprecated v1 endpoint.
The coordinator creates one thread per repo to migrate callers, run tests, and open PRs, then reports
which PRs must merge first.
1---2name: goal-driven-multi-thread-projects3description: Set up and steer a redesigned Claude Code project (beta, September 2026) where you state a goal and a coordinator routes work to parallel threads - each thread a Claude Code cloud session on its own branch - backed by shared memory and a library. Covers choosing the goal and repos/context, configuring environment, connectors, plugins, instructions, and model/effort for coordinator and workers, briefing the coordinator like a chief of staff, steering threads (including from a phone), handling merge order and conflicts, tuning check-in frequency and update detail, and watching project usage. Use when long-running, multi-part work (latency reduction, multi-repo migrations, endpoint retirement) would otherwise require you to split tasks, hand off context, and merge results by hand.4metadata:5 source: https://skillmd.com/skills/uygnoey/goal-driven-multi-thread-projects6---78## Purpose9The source announcement describes projects that move "from folder to conversation": you describe a goal10and Claude manages the work across multiple sessions instead of you splitting tasks, managing handoffs,11and merging results. This skill turns that model into a setup-and-steering procedure.1213## When to use14- Work spans several repos or many independent sub-tasks that can run in parallel.15- The effort runs for days or weeks and decisions must persist (shared memory).16- You want to delegate at the level of outcomes, not individual prompts.1718## Instructions19201. **Write the goal as an outcome.** Pick one measurable goal and the repos or documents it needs.21 Claude then suggests work it can start immediately. Use22 [templates/coordinator-brief.md](templates/coordinator-brief.md).232. **Configure the project.** Cloud environment, connectors, plugins, project instructions, and the model24 and effort level for both the coordinator and the worker threads.253. **Brief the coordinator like a chief of staff.** State priorities, constraints, who to consult, and what26 "done" means. The coordinator routes tasks to new or existing threads.274. **Let threads do the work.** Each thread is a Claude Code cloud session on its own branch and copy of the28 repo. With repos connected, threads open PRs and run tests; with documents, they read and draft.29 Threads may split work further with subagents, loops, and workflows.305. **Steer at the right level.** Follow progress in the main project chat; open a single thread only to31 steer details. You can steer from your phone, and work continues after you step away.326. **Treat overlap as normal PR work.** Overlapping changes surface as merge conflicts, as with any PR.33 Ask the coordinator which PRs must merge first.347. **Use shared memory instead of heavy prompt engineering.** Every thread adds to and draws from it (e.g.35 a release moved to Friday, why an export was dropped, who to consult before changing billing). Put36 reusable files in the library; artifacts Claude produces land there too.378. **Tune the working style.** Adjust how often Claude checks in, how often it starts new threads, and how38 detailed its updates are.399. **Watch usage.** Several full sessions run at once, so limits can be reached faster. Check40 project-specific usage and lower model/effort for worker threads when appropriate.4142Availability and limits as of the announcement are in43[references/availability-and-limits.md](references/availability-and-limits.md).4445## Examples4647**Latency goal**48> Goal: reduce checkout p75 latency. Repo: checkout service.49> Claude profiles each endpoint, tests optimizations, and opens PRs in parallel threads.5051**Endpoint retirement across repos**52> Connect the API, web, and mobile repos. Goal: retire the deprecated v1 endpoint.53> The coordinator creates one thread per repo to migrate callers, run tests, and open PRs, then reports54> which PRs must merge first.