Opinionated coworker setup for Claude Code · Slack · Notion · Gmail · Granola.
Bootstrap a working AI coworker end-to-end, by interview.
Anson v1 wrote your identity docs. v2 installs the whole coworker: identity + integrations + scheduled automations + custom workflows. You talk to it once, and at the end you have a workspace you can talk to immediately.
- Identity —
AGENTS.md,IDENTITY.md,USER.md,SOUL.md,MEMORY.mdwritten from your answers, not templates - Integrations — Gmail, Google Calendar, Slack, Notion, Granola (macOS), iMessage (macOS) — authenticated, verified, and wired into your agent's tool surface
- Rhythm — morning brief, standup notes, evening wrap-up, weekly digests — whichever you ask for, scheduled and running
- Two-way Slack coworker (optional) — DM the agent from your phone, get replies in thread
- Your own workflows — describe any automation in plain English; Anson generates a skill, dry-runs it on your real data, and installs it if you approve
Opinionated by default. Every choice reflects what's actually worked in production for the founder it was built around. You can override anything later, but anson picks first so you're not making 30 micro-decisions to get to a baseline. Strangers get the same core, customized to their identity and workflow.
Via Claude Code plugin (recommended):
/plugin install superpositionhq/anson-v2
Then in your agent:
Run anson
Or clone manually as a skill:
git clone https://github.com/superpositionhq/anson-v2 ~/.claude/skills/anson-v2
# the skill lives at skills/anson/SKILL.md inside the repo;
# the plugin layout works whether cloned as a skill or installed as a plugin.Then > Run anson.
That's it. Anson handles the rest — step by step, one question at a time, resumable if you get interrupted.
- Claude Code (primary) — Agent SDK works with reduced features
- Strong reasoning model (Claude Sonnet 4.6+ / Opus)
anthropic-skills:skill-creatorpluginuvfor the Slack bridge (optional stage)- macOS or Linux — iMessage + Granola are macOS-only; everything else works on both
The flow is conversational, not a checklist — but at a high level:
| Stage | What happens | What gets written |
|---|---|---|
| A | Anson introduces itself, picks workspace root | <workspace>/ANSON_META.md |
| B | Who are you? (identity interview before any tools) | USER.md, IDENTITY.md, SOUL.md, KEY_PEOPLE.md, MEMORY.md skeleton |
| C | How do you work? (open-ended workflow discovery) | workflow summary in ANSON_META.md |
| D | Anson proposes the install plan based on Stage C, then executes only what your workflow needs | .env, verified integrations, installed skills, scheduled tasks, optional Slack bridge |
| E | Any custom workflows? Describe in plain English; anson generates a skill, dry-runs it, installs it | custom skills (≤5 per session) |
| F | Hand-off | install inventory + next steps |
Total time: ~30–45 minutes, mostly waiting on OAuth.
Identity (Stage B) comes before integrations (Stage D) so installs are tailored to your stated voice and preferences. Integrations are consequences of Stage C — if you don't use Slack, anson never asks for a Slack token.
If you close the terminal mid-install, Anson picks up from the last checkpoint when you say "Run anson" again. Progress is tracked in ANSON_META.md.
Re-run anson any time. It detects the existing install, skips everything done, and jumps you straight to Stage E so you can add more automations.
MIT.
Successor to anson v1 by Superposition. The v1 identity-interview approach is preserved as Stage B; v2 absorbs everything around it.
