Formerly ai-business-skills.
Claude skills for turning messy workplace context into clear facts, asks, decisions, owners, risks, updates, and safe replies.
Most AI tools summarize context. This repo turns context into action while preserving uncertainty.
| Messy context | Clean output |
|---|---|
| A workplace thread with disagreement, legal uncertainty, support readiness risk, a soft deadline, and no named decision owner | Clean ask, decision needed, owner gap, risk/tone note, and a reply draft you can paste |
The promise is small: turn messy workplace context into the next clear move without pretending uncertainty is resolved.
- A Slack thread, email chain, or meeting transcript needs a fast catch-up
- Someone asked a vague question and you need to clarify the ask
- A dense memo, policy doc, vendor writeup, or transcript needs a fact ledger before anyone acts on it
- A conversation has decisions, owners, deadlines, blockers, or open loops
- You need to draft the next context-aware reply
- You need to brief other people on what changed
git clone https://github.com/tmusser/context-to-action-skills.git
cd context-to-action-skills
./install.sh --claude-user --only reduce-to-facts,clear-ask,decision-brief,follow-up-draft
./install.sh --codex-user --only reduce-to-facts,clear-ask,decision-brief,follow-up-draftStart with reduce-to-facts; it is the safest first move for dense or ambiguous context. clear-ask, decision-brief, and follow-up-draft turn the ledger into action, and the full set adds brief-me, meeting-to-actions, and status-update. --include-templates is optional if you also want the supporting templates. Use --force only when you intentionally want to replace an existing destination that differs.
If the source is long or ambiguous, start with reduce-to-facts and ask for a fact ledger before asking for a reply, update, or decision.
Paste messy notes, a transcript, or connected context into Claude/Cowork and ask:
Use reduce-to-facts on this thread and give me a compact fact ledger with source scope, atomic facts, open questions, and action-relevant implications.
reduce-to-facts defaults to source-only mode. Ask for source-expanded mode only when you want outside evidence checked; it adds a Claim Status table and Rhetorical Leaps.
Tiny example:
Messy input:
Alex: "Can we still send the customer note today?"
Sam: "I'd rather not wait; support can catch up after."
Priya: "Support needs the enterprise wording before anything goes out."
Jordan: "Legal has not approved the banner copy yet."
Alex: "The send queue closes at 3pm. Who owns the call?"
Clean output:
Clean ask: decide whether to send today or wait for legal/support readiness.
Decision needed: send today vs. move to tomorrow.
Owner gap: no final decision owner is named.
Suggested reply: "Can we name the decision owner and confirm by 3pm whether legal/support readiness blocks today's send?"
Full version: [examples/messy-thread-to-follow-up.md](examples/messy-thread-to-follow-up.md)
Shareable one-pager: docs/one-pager.md
These are plain-text skills, so the lowest-friction mode is copy/paste: paste the thread, note, email, transcript, or doc excerpt into your AI tool and ask for the skill by name.
The tested path is Claude/Cowork-style skills. The underlying SKILL.md files can also be adapted as reusable prompts for ChatGPT, Gemini, Codex-style agents, or internal AI assistants. No connector is required for pasted-context mode.
Start with reduce-to-facts for dense or ambiguous source, then pass the ledger to clear-ask, decision-brief, status-update, or follow-up-draft.
This is the pasted-context validation boundary for the repo.
The validated path is pasted-context usage with Claude/Cowork-style skills.
The SKILL.md files are plain text and may be adapted for ChatGPT, Gemini, Codex-style agents, or internal assistants, but those paths are compatibility targets rather than fully tested integrations.
Connector and action-mode behavior depends on the host assistant and the tools the user explicitly authorizes. This repo provides guardrails and output contracts; it does not independently grant tool access or mutate systems unless explicitly asked.
For why MCP stays out of scope for now, see integration roadmap.
This repo is a lightweight skill pack for pasted workplace context.
It is not a deterministic workflow engine, ticketing system, MCP server, or project-management backend.
The skills provide reusable instruction contracts and examples. They help an assistant preserve uncertainty, identify asks, surface owners, draft replies, and avoid unsafe action. They do not guarantee structured output unless the host assistant follows the requested format.
For machine-validated workflows, use the optional schemas and examples as a stricter handoff layer.
Slack threads, email chains, doc comments, tickets, and transcripts can all carry the same source state as a live meeting. Treat them as async collaboration: extract facts first, then move to the next clear action.
No coding is required.
Each skill lives in skills/ as a folder with a SKILL.md file.
If your AI tool can read a local skills directory, point it at skills/. If it supports one of the Claude/Codex paths above, use the installer. Otherwise copy the folders into the directory it expects, then paste context and ask for a skill by name. For example:
Use brief-me on this thread and tell me what needs my response.
These skills read and draft by default. They do not send, publish, create events, or update systems unless you explicitly ask.
Start with reduce-to-facts when the source is too long or ambiguous to act on safely, then move to clear-ask, decision-brief, status-update, or follow-up-draft.
- pasted-only mode uses only pasted text
- connected-context mode reads only approved tools
- draft-only mode does not send or publish
- action mode mutates systems only when explicitly asked
- Tone safety: helps avoid accidental escalation, fake certainty, overcommitment, and misreading stakeholder sensitivity
This repo is for the work before the work:
- reduce dense context to safe facts
- understand what people said
- identify what changed
- separate facts from assumptions
- clarify the ask
- preserve open loops
- draft the next context-aware response
This is not a deck generator, project-management framework, or heavy artifact factory.
| Input | State extracted | Output |
|---|---|---|
| Dense context, long thread, memo, policy doc, vendor writeup, or transcript | Facts you can rely on, plus opinions, unsupported claims, contradictions, open questions, and implications | reduce-to-facts fact ledger you can pass downstream |
| Slack thread, email chain, meeting transcript, doc comment, or Jira/Confluence discussion | What changed, clean ask, decision, owner, deadline, blocker, source confidence, and tone risk | Reply draft, follow-up, status update, decision snapshot, or action list |
| Workflow | User question | Skill |
|---|---|---|
| Reduce dense context | What can we safely rely on? | reduce-to-facts |
| Catch up | What did I miss? | brief-me |
| Clarify | What are they actually asking? | clear-ask |
| Extract actions | What are the actions and open loops? | meeting-to-actions |
| Decide | What decision is needed? | decision-brief |
| Respond | What should I say back? | follow-up-draft |
| Broadcast | What should others know? | status-update |
Choose the mode that gives you the right control.
| Mode | You control |
|---|---|
| Pasted-only mode | Use only text you pasted |
| Connected-context mode | Read approved tools like Gmail, Slack, Calendar, Docs, Jira, or Confluence |
| Draft-only mode | Produce messages, updates, or follow-ups without sending or publishing |
| Action mode | Mutate systems only when you explicitly ask |
Use pasted notes, Zoom transcripts, Slack, Gmail, Calendar, Atlassian, and Hex or other data outputs as inputs.
By default, these skills read and draft. They do not send messages, create events, update tickets, or publish outputs unless you explicitly ask.
Shared guidance for mixed-source inputs lives in source-packet.md.
- Reduce to facts vendor risk
- Reduce to facts source expanded
- Messy thread to follow-up
- Gmail buried obligation
- Leadership status update
- Post-meeting follow-up
- Morning brief
- Decision with data
- Async Slack clear ask
See examples/README.md for a quick index of what each example demonstrates.
Every skill is designed to produce something you can send, decide from, or act on within 5 minutes.
Helpful supporting files:
This repo was built using ai-engineering-skills as the control layer.
This repo is one piece of a small set of repos for making AI-assisted work clearer, more bounded, and more verifiable. See Suite map.