Taste is the judgment you can recognize but can't easily put into words. Cisely writes down its roots — the beliefs your company holds, and the choices they justify — so your people and your AI can apply it without you in the room.
AI has turned intelligence into a utility — limitless, and the same for everyone. When anyone can generate anything, generation stops being the scarce half; what's scarce is the judgment to recognize which output is right — what the industry has started calling taste. Taste is easy to recognize and hard to state, and at the level of a company it is rooted in shared belief: what a business holds true is what decides what "good" even looks like.
But taste you can't express is taste your team and your AI can't apply — it stays locked in a founder's head. Cisely exists to make it concrete. It doesn't try to bottle the ineffable judgment itself; it writes down its roots — your beliefs, and the choices they justify (whom to serve, which expectations to meet, which to deliberately decline) — as a shared, machine-legible context graph. An AI dropped into that context stops being a tool waiting for instructions and becomes a teammate that has inherited what the company is trying to do, for whom, and why — able to apply your judgment without you in the loop. Cisely keeps a living, continuously-revised approximation of that judgment; it amplifies the vision you already hold — it doesn't author it for you.
This package is how you build that context with your coding assistant. It teaches the Cisely agency model and drives the Cisely MCP to author it with you: the response side (mission → strategy → initiatives → metrics) and the demand side (personas → goals; contexts → stakeholders → expectations), joined by the beliefs both sides root in — the business's shared "mind," written down once and navigable by every human and agent that works in it.
Works with Claude Code, OpenAI Codex, Google Antigravity, Cursor, OpenCode, Aider, and other
AGENTS.md-aware tools — one payload, thin per-harness adapters. Seedocs/install.md.
Cisely models a business as two sides that meet where value actually changes hands.
- Demand — what your stakeholders want. Anything that acts is an agent. Its invariant identity is a persona, which holds latent goals (emotional deliverables — to feel secure, in control, proud). Viewed through an operating context (onboarding, security, billing…), a persona becomes a stakeholder carrying concrete expectations — the only directly measurable proxy for the goal underneath.
- Response — what the business does about it. A belief (its mission) roots a strategy, which chooses whom to serve and which expectations to meet — and which to deliberately decline. Strategy spawns initiatives that build touchpoints; metrics check whether the expectations were actually met.
The two sides are one equation — joined at the touchpoint, audited by the metric:
RESPONSE belief → strategy → initiative → touchpoint
↓ MEETS
DEMAND belief → goal (latent) → expectation
AUDIT metric measures the meeting → evidence reshapes the strategy
A business is strongest when the belief rooting its strategy and the belief rooting its customers' goals are the same belief — then serving them stops being a guess about strangers and becomes an expression of shared conviction. (Author agency once, and Jobs-to-be-Done and a Value Proposition Canvas fall out as views of the same graph — derived, never maintained by hand.)
An AI agent is itself an agent in this model — a delegated one, acting for a principal on inherited goals. Alignment is precisely the gap between the goals it inherited and the actions it takes — and agents drift because that context was never written down. They were handed a task, not the business.
Lend the agent this context graph and it discovers the business by traversal — from a touchpoint to the expectation it meets, to the stakeholder that holds it, to the persona and its goal, to the belief, to the strategy that answers. It stops guessing what "good" means and inherits your company's actual judgment: alignment implanted by construction, not hoped for.
That is what turns a capable model into a 10× teammate — not more raw intelligence, but the substrate it stands on. A 10× engineer is usually an ordinary engineer on a 10× platform; move them off it and the multiplier evaporates. The Cisely context graph is that platform for your AI — the difference between an assistant that executes prompts and a teammate that shares the mission.
- Skill
cisely-agency-model— the worldview, the modeling discipline, and the playbook for wielding the Cisely MCP tools in the right order. Claude uses it automatically when you talk about modeling your company, personas, strategy, goals, or value proposition. - Skill
cisely-sub-strategy— how to define your own strategy kinds (a technology strategy, a GTM strategy, a people-&-culture strategy): author a company-wide schema, then fill in records. - Commands:
/cisely:model— start a guided modeling session (interview, or ingest from your own docs)./cisely:review— traverse the graph and report coverage, gaps, and belief-alignment./cisely:canvas— synthesize Jobs-to-be-Done and a Value Proposition Canvas from the graph.
Sign up at app.cisely.dev and create (or join) your company's workspace. To author model content you need tenant-admin rights; any member can read. This is also where you'll view everything you build — the Agency Board (demand side) and the Purpose Board (response side).
This package works across coding assistants. In Claude Code:
/plugin marketplace add CiselyAI/agency-skills
/plugin install cisely@cisely-marketplace
(/plugin marketplace add also accepts a full Git URL if you're not using GitHub.)
Using Codex, Antigravity, Cursor, OpenCode, Aider, or another AGENTS.md-aware
tool? See docs/install.md for the one-per-tool setup. Every
harness reads the same skills; each just needs the Cisely MCP wired up (step 3).
The plugin registers the remote MCP server at https://app.cisely.dev/mcp. Run /mcp, select
cisely, and complete the OAuth sign-in in your browser using the account from step 1.
Prefer to set it up manually?
claude mcp add --transport http cisely https://app.cisely.dev/mcp
/cisely:model
Then /cisely:review to check coverage and belief-alignment, and /cisely:canvas to synthesize
Jobs-to-be-Done and a Value Proposition Canvas for any persona.
New versions ship as pushes to this repo. How you pull them depends on your harness — there are two models. (The Cisely MCP is a hosted server, so it updates on its own; you only ever refresh the skills and adapters.)
Plugin-managed harnesses — the tool's plugin manager fetches the new version:
- Claude Code:
This refreshes the catalog; Claude Code picks up the new release. To force it, reinstall with
/plugin marketplace update cisely-marketplace/plugin uninstall ciselythen/plugin install cisely@cisely-marketplace. Use/reload-pluginsto apply without restarting. - OpenAI Codex: reinstall cisely from the Plugins list (app) or
/plugins(CLI). - Google Antigravity: re-run
agy plugin install https://github.com/CiselyAI/agency-skills.
File-based harnesses — the tool reads the skills straight from your checkout, so a pull is the update:
git -C /path/to/agency-skills pullApplies to Cursor (.cursor/ rules), OpenCode (cloned repo), Aider
(.aider.conf.yml), and any other AGENTS.md-aware tool. Restart the tool (or start a new session)
so it re-reads the files. See docs/install.md for per-tool specifics.
Everything you create via the MCP is visible in the Cisely web app at https://app.cisely.dev
(the Agency Board for the demand side, the Purpose Board for the response side). After each
change, the plugin gives you a direct link to view it.
The plugin is high-rigor by design: it coaches the model's discipline (every node goes
Draft → Revise → Activate), treats a declined expectation or persona as a deliberate decision,
and pushes toward the model's core payoff — a business whose strategy roots in the same beliefs
its stakeholders hold.
The agency model is summarized in
skills/cisely-agency-model/references/agency-model.md.
Licensed under the Apache License 2.0. Copyright 2026 Bhumika AI Pte Ltd.