Session Handoff · Decision Log · Idea Developer · Agent Creator · Install · Add a skill · License

A small, hand-maintained collection of open agent skills I keep improving. Each one lives under a category folder in skills/ and stays small enough to read in full. Drop the ones you want into your own skills directory.

Right now there are four, under skills/continuity/, skills/creativity/, and skills/opencode/. That list will grow.

Sessions are ephemeral; the model and the human both forget context between them. Session Handoff keeps a tiny set of project docs so work survives. When a session starts it reads them in order, orients on where things stand, and proposes the highest-value next step. When it ends it writes current state and next actions back down, so the next session picks up cleanly instead of re-discovering everything.

The docs it maintains:

• HANDOFF.md holds where things stand and what's next.
• SESSION_LOG.md is a running record of each session.
• RUNBOOK.md lists the commands worth not rediscovering.

Durable decisions are recorded by Decision Log, not here. It ships with references/templates.md (exact formats for each doc) and references/read-order.md.

The "why" of a choice is the part that doesn't survive on its own. The what lives in code and config; the reasoning evaporates. Decision Log writes it down at the moment the call is made, so three months later nobody has to re-litigate an architecture debate because the tradeoffs were never captured.

It records the context, the options considered, the accepted tradeoffs, the rejected alternatives, and the condition that would flip the call, into docs/DECISIONS.md. When a decision is revisited it reads the old entry first, so the new debate starts from the recorded reasoning instead of from scratch. It ships with references/templates.md for the entry format.

Session Handoff hands decision capture over to this skill rather than owning it itself.

A rough idea is not a plan. Idea Developer takes the half-formed tech or product thought and interviews it into a one-page plan. It asks one question at a time, adapts to whatever the user just said, and probes the weak spots: an audience of "everyone", a differentiator of "it's just better", a scope that quietly covers three products. After five to ten questions it hands back a filled template with a one-liner, problem, audience, solution, differentiators, MVP scope, risks, and next actions. The template fills as the conversation runs, so the plan is already written by the time the interview ends.

It ships with references/questions.md, the question bank organized by theme, each section carrying a probe that names the weak spot to push on.

Agents are just files with a strict shape: a YAML frontmatter from a fixed field list, and a body that becomes the prompt. The user thinks in plain language, the skill thinks in schema. Agent Creator turns a request like "I need an agent that reviews my pull requests" into a working .opencode/agent/<name>.md. It interviews briefly for purpose, mode, model, and permissions, writes a valid definition to the right scope, and reminds you to restart opencode, which loads config at startup and ignores new agents until then.

It ships with references/agent-format.md, the exact file format: locations, the allowed frontmatter fields, how permission rules behave, and worked examples.

Install every skill in this repo:

npx skills add pkhamre/skills

Or add a single one:

npx skills add pkhamre/skills --skill session-handoff
npx skills add pkhamre/skills --skill decision-log
npx skills add pkhamre/skills --skill idea-developer
npx skills add pkhamre/skills --skill agent-creator

Adding a skill means creating skills/<category>/<name>/SKILL.md with a name and description in the frontmatter, plus any bundled references/, scripts/, or assets/ it needs. The description is what makes it trigger, so it should say both what the skill does and when to reach for it. Keep the body under a few hundred lines and push detail into reference files. See opencode's skill docs for the layout.

Installed upstream skills and the opencode install manifest aren't tracked here. Authored skills are the contents of skills/.

MIT. See LICENSE.