| name | repo-business-context |
| description | Repo Business Context (RBC) help agents to manage business context within code repositories. Turn agents from code monkeys into business partners that can code. It stores and updates brand, business, and product context as plain Markdown files in the `/context` folder. Trigger on all interactions and tasks - even if the user doesn't mention RBC. Agent will read relevant files, determine if they are business-related, and retrieve or update context files accordingly. When triggering on a repo without context folder or files, the onboarding happens via 5W2H questions, codebase extraction, or gradual gap-filling. |
| allowed-tools | Read(./context/*) Write(./context/*) Edit(./context/*) Glob(./context/*) |
| license | MIT |
| compatibility | Designed for any skill-aware AI agent, LLM or CLI tools, including amp, antigravity, claude-code, clawdbot, codex, cursor, droid, gemini, gemini-cli, github-copilot, goose, kilo, kiro-cli, opencode, roo, trae, and windsurf. |
| metadata | {"author":"issouza","version":"1.1.0","category":"skill"} |
| author | {"name":"Igor Souza","email":"igor@datasuccess.io","github":"/issouza","linkedin":"/in/igorsansouza","affiliation":{"name":"DataSuccess","url":"https://www.datasuccess.io/"}} |
Repo Business Context (RBC) — Operating System
You are operating in a repository that uses the Repo Business Context (RBC) framework. The /context/ folder holds canonical brand, business, and product knowledge as small focused Markdown files. Your job: consume it selectively, keep it aligned with reality, and help the user populate it with minimal friction.
The codebase is the source of truth for execution. /context/ is the source of truth for intent — pricing, positioning, personas, brand voice, strategy. Things that aren't in the code but shape how the code should behave and how the product should feel.
Core principles
- Load selectively. Never read the entire
/context/ folder. Pick the 1–3 files relevant to the current task. Loading 20 files gives you worse reasoning than loading 2 relevant ones — attention degrades with input size.
- Keep context current. When reality shifts, update the relevant file and announce it. Silent drift is the failure mode.
- Never invent business facts. If the context you need isn't there, ask the user. Do not guess pricing, positioning, audience, or voice from code or assumptions.
- Always announce file changes. Whenever you write or update a context file, say so explicitly in your response. Never silently mutate context.
- Be explicit. In any substantive response, state which
/context/ files you read. This lets the user course-correct quickly.
- For execution tasks, prefer observable behavior. If
/context/ conflicts with what the code actually does, complete the immediate task based on what the code shows, flag the conflict, and update the context file.
- No technical details. The tech stack can be infered from the code itself. Write them down within
/context/business or /context/product if they are business-related only.
- No task management. Project management is assumed to be done as commits, PRs and releases. Also, each AI/agent has its way of managing short-term plans and tasks. You can use
/context/product/backlog.md to write ideas and goals loosely, but never tasks.
Bootstrap — setting up RBC from scratch
Before detecting which situation you're in, check whether RBC scaffolding exists.
Look for /context/README.md and the three subfolders (business/, product/, brand/). If they're absent, scaffold before anything else:
- Create
/context/ and its three subfolders (brand/, business/, and product/).
- Tell the user briefly: "I've scaffolded the RBC folder structure under
/context/. Let's set up your business context."
Then continue to the situation detection below.
Overall instructions
Here are the instructions to follow on before answering any question or doing any task:
- Identify if the task is related to business context.
- If the task is not business-related, neither read or write RBC files, and proceed normaly.
- If the task is business-related, determine if it requires reading, writing or updating context files.
- Choose 2 to 3 files based on the file intent below and the task at hand.
- Infos given by the user are always more important than the context files. If any update is needed, do not change files silently, but ask the user for confirmation.
- Answering the user and completing the tasks is more important than managing RBC files, but keeping RBC files updated to the latest business context can't be neglected.
- If you can't find the right RBC files to work with, just proceed normaly.
- If you can't complete the task without updating RBC files, do it.
First decision: what situation are you in?
Before any substantive task, decide which of these three situations you're in. Glob /context/**/*.md and glance at file sizes — you can tell quickly without reading content.
Internal navigation only — never use the labels "Situation A", "Situation B", or "Situation C" in messages to the user.
Situation A — /context/ is empty or nearly empty
Most of the files under /context/brand/, /context/business/, or /context/product/ are empty or missing. The user hasn't onboarded yet.
Offer the three paths once, briefly, then wait. Do not push — "build gradually" is a valid and recommended default.
- Quick onboarding (~5 min) — 7 questions using the 5W2H method. Drafts the core files in one pass.
- Extract from codebase — scan the repo for business signals and draft inferences for you to confirm. Good for existing codebases with a meaningful README and structure.
- Build gradually (default) — skip the upfront flow. I'll ask 1–2 minimal questions only when I hit a context gap during normal work.
Phrase it conversationally, e.g.: "Before I help with this — your /context/ looks empty, so I don't have your business details yet. Want me to run the 5-minute onboarding, extract what I can from the codebase, or just keep working and fill context as we go?"
If the user picks a path, load the matching reference file:
- Quick →
references/onboarding-quick.md. If the users asks for the recommended approach, this is the answer.
- Codebase →
references/onboarding-codebase.md. Works well for projects already rich in content, or repos with research and other context documents in it.
- Gradually (or no answer, or "just go") →
references/onboarding-gradual.md and proceed with the original task. This is the fallback/default approach.
Situation B — /context/ has content
Normal operating mode. Before acting on a business-flavored task (copy, feature design, pricing, positioning, UX, strategy), open references/routing.md and load the 1–3 files that match the task. State in your response which files you read.
Situation C — reality just changed
If the current conversation or code change shifts business reality (new feature shipped, pricing updated, new provider added, audience narrowed), propose context updates. Open references/maintenance.md for the workflow.
These situations aren't mutually exclusive — a single task can start in B and end in C.
Context files are living documents
/context/ files are plain Markdown. The agent writes them, keeps them current, and announces every change. The user corrects wrong content in conversation; the agent updates immediately.
This is the same trust model as the codebase: you don't gate every AI-written line of code through a review checkpoint. You trust the output, correct what's wrong, and move on.
Writing and updating files
- Write with confidence. Create or update files with the best content available from current signals. Don't hedge — write as if it's true, because you'll keep it true.
- Always announce changes. When you write or update a context file, say so in your response — one line: "Updated
business/revenue-model.md — free tier raised to 50 requests/day." Never silently mutate context.
- Note your sources for inferences. When content is inferred rather than explicitly stated, briefly note the basis: "Set brand voice from README tone — terse and technical." This helps the user catch misreads fast. It's not a disclaimer; it's useful context.
- Correct immediately. When the user says something is wrong, update the file in the same turn. Don't ask for permission to fix what you've been told is wrong.
When to update
Update a context file whenever the work you just did (or discussed) shifts what's true about the product. Use the table in references/maintenance.md for specifics.
The bar: would a thoughtful collaborator update this note before the next session? If yes, update it now.
When files become obsolete
If a file's content no longer reflects reality and there's no clear replacement, delete the stale content and write what's currently true. Don't leave misleading content in place — wrong context is worse than no context.
References — load as needed
Do not read all of these upfront. Open only the one relevant to what you're doing.
references/onboarding-quick.md — The 5W2H 7-question script and how to route answers to multiple files. Open when the user picks quick onboarding.references/onboarding-codebase.md — How to mine an existing repo for business signals and batch-confirm inferences. Open when the user picks codebase extraction.references/onboarding-gradual.md — How to ask 1–2 minimal questions lazily during normal work. Open when the user picks gradual, or by default when context gaps appear.references/routing.md — Task type → which files to load. Open before any substantive business-flavored task (Situation B).references/maintenance.md — When and how to update context files; what triggers an update vs. what doesn't. Open after work that shifts business reality (Situation C).
What this skill does NOT do
- Does not force onboarding. "Build gradually" is valid; respect it.
- Does not block deploys or gate work. Context updates are proposals, not blockers.
- Does not replace human judgment on business decisions. The skill captures decisions the user makes; it doesn't make them.
- Does not deep-dive the codebase on every task. Only during codebase-extraction onboarding, or when the user explicitly asks.
