Harvest

Skill ajaygunalan/flow-friction/skills/harvest

Claude Code skills for researchers, by a researcher. Humans + agents exploring what hasn't been built yet — without losing alignment along the way.

Install

npx -y skills add ajaygunalan/flow-friction --skill harvest

Assembled from the repository path, not quoted from the project. Check it against their README if it does not work.

2 things to look at

no licenseNo license file was found in the repository. Code published without one is not open source by default, so using it at work is a question for whoever answers licensing questions where you are.
2 stars2 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.

What its author says it does

Copied from the file, not written here

End-of-session distillation — scans the conversation and existing issues, filters signal from noise through adaptive questioning, then creates, updates, or closes issues. Use when the user says "harvest", wants to wrap up a session, capture decisions before closing, or bridge work to a future conversation.

SKILL.md

2.0 KB, as published. Nobody here has run it

Phase 1: SCAN

Read the conversation from beginning to end. Read open issues in docs/issues/ (skip status: done). If $ARGUMENTS is given, use it to focus.

Build a mental map: what was done, what changed, what broke, what was decided, what's still open.

Phase 2: FUNNEL

Present a short summary of what you extracted — threads, decisions, open questions. Then ask the user what matters using AskUserQuestion.

This is adaptive, not fixed. Keep filtering until the user says that's enough. One round may suffice for a focused session. Five rounds may be needed for a sprawling brainstorm. Draw from these angles as needed — they are prompts, not a mandatory sequence:

What are we solving?
What are we building?
How do we know it's done?
What would a fresh agent get wrong?

Stop when the signal is clear.

Phase 3: DECIDE

Based on the filtered signal + existing issues, propose a set of actions. Present them as a list for user approval:

Create <slug> — new issue
Edit <slug> — any change (update, rewrite, status change)
Delete <slug> — no longer relevant
No issue — nothing to file

Multiple actions in one harvest is normal.

Do NOT write anything yet. Get explicit approval on the action list first.

Phase 4: WRITE

Execute the approved actions. Edit existing files or create new ones. Let the content find its natural shape — narrative, bullets, decision log, spec fragment, whatever makes the signal clearest.

Follow the issue conventions defined in global CLAUDE.md (location, header format, statuses, types).