Make the idea earn the build.
From a rough request to a researched decision and the files needed to act on it.
Demo | Workflow | Flow diagram | Outputs | Compatibility | Install | Examples | Evaluation
ConceptOps is an open Agent Skill for product discovery, feature planning, technology evaluation, market research, feasibility checks, unit economics, validation, prototyping, documentation, presentations, websites, brand systems, and visual concepts.
Give it a short, messy request. It finds the relevant context, separates facts from assumptions, compares realistic options, tests the key unknowns, makes a decision, and creates only the outputs needed for the next step.
When the task is already a clear implementation, it steps aside: the host executes directly, without questions or .ideas/ artifacts.
Install the same portable skill in OpenAI Codex, Claude Code, or GitHub Copilot. There is no hosted service to trust, account to create, or runtime dependency to maintain.
offline search. private. 20k markdown notes. no cloud.
what should we use? implementation can wait.
Mode: technology decision
Goal: choose a private local search approach
Constraints: one laptop, no cloud, no always-running service
Compare: current project -> native features -> installed tools
-> mature libraries -> minimal custom work
Decision: Go / Conditional / Freeze
Proof: a measurable retrieval-quality experiment
Outputs: task brief + solution blueprint + sources + usage summary
No invented requirements. No five-layer architecture. No polished landing page before the idea has earned one.
flowchart TD
A["Request"] --> B{Needs discovery,<br/>evidence, or a decision?}
B -->|No| X["Direct execution<br/>No ConceptOps run files"]
B -->|Yes| C["Normalize intent,<br/>context, and constraints"]
C --> D["Select the strategy<br/>and output set"]
D --> E{Would independent parallel<br/>or specialist work help?}
E -->|Yes| F["Run bounded specialist<br/>subagents in parallel"]
E -->|No| G["Main agent handles<br/>the run directly"]
F --> H["Main agent consolidates evidence,<br/>resolves conflicts, and decides"]
G --> H
H --> I["Write structured outputs<br/>to .ideas/<slug>/"]
- Gate: bypass ConceptOps when the request already needs only direct implementation.
- Normalize: preserve the request while making its objective, context, constraints, and unknowns explicit.
- Select: choose the smallest useful output set, evidence plan, and execution strategy.
- Delegate: use bounded specialist subagents only when independent work can run in parallel or improve the decision.
- Consolidate: keep the main agent responsible for conflicts, evidence, verdict, and final artifacts.
- Persist: write the manifest, brief, sources, selected outputs, and usage summary under
.ideas/<slug>/.
- Messy input is expected. A fragment is enough to begin.
- Clear work stays direct. Well-scoped implementation bypasses the ConceptOps workflow.
- Existing projects come first. ConceptOps reads the real codebase, instructions, dependencies, patterns, and affected flow before recommending a change.
- Reuse beats invention. It checks current capabilities, native platform features, and installed dependencies before adding anything.
- Current claims need current sources. Prices, regulations, competitors, libraries, and benchmarks are researched instead of guessed.
- Facts stay separate from assumptions. Calculations use formulas; unsupported precision is rejected.
- Effort follows risk. Stronger models, deeper reasoning, and specialist subagents are used only where they can change the decision.
- External actions stay gated. Production deployment, purchases, and public release require explicit authorization.
Choose one component or combine several with AND/OR semantics. ConceptOps recommends the smallest useful set after it understands the task.
| Component | What it leaves behind |
|---|---|
| Brief report | Evidence, risks, verdict, and next action |
| Presentation | Editable decision narrative |
| Website | Landing page source and local preview |
| Documentation | Product, technical, API, user, or operations documents |
| Brand book | Positioning, voice, visual direction, and reusable tokens |
| Unit economics | Editable assumptions, scenarios, formulas, and sensitivity |
| Research pack | Sourced market, customer, competitor, or technology evidence |
| Validation kit | Hypotheses, experiments, thresholds, and kill criteria |
| Prototype | Wireframe, faithful mockup, or interactive concept |
| Go-to-market | Segment, positioning, channels, experiments, and metrics |
| Solution blueprint | Project-grounded implementation options and recommendation |
Every package includes summary.md: what was created, why it exists, how to use it, what remains uncertain, and what to do next.
ConceptOps works as a native Agent Skill in:
| Host | Status | Surfaces | Version |
|---|---|---|---|
| OpenAI Codex | First-class | App, CLI, and IDE extension | Current stable |
| Claude Code | First-class | CLI and IDE integrations | 2.0.20 or newer |
| GitHub Copilot | First-class | Cloud agent, code review, CLI, app, VS Code, and JetBrains | Current stable |
ConceptOps uses the portable core of the open Agent Skills specification: a SKILL.md file, YAML metadata, relative references, and bundled assets. It does not depend on host-specific hooks, subagent fields, or shell permissions. The optional agents/openai.yaml file adds presentation metadata in Codex and is ignored by other hosts.
No single minimum version is published across every Codex and GitHub Copilot surface, so current stable releases are recommended. Claude Code added Skills support in 2.0.20.
Install ConceptOps globally for Codex, Claude Code, and GitHub Copilot:
npx skills add uginy/conceptops -g -a codex -a claude-code -a github-copilotThe command is provided by the open skills CLI. Use npx skills add uginy/conceptops --list to inspect the skill before installing it.
npx skills add uginy/conceptops -g -a codexInvoke with $conceptops or select it from /skills.
npx skills add uginy/conceptops -g -a claude-codeInvoke with /conceptops or let Claude load it when the request matches.
npx skills add uginy/conceptops -g -a github-copilotInvoke with /conceptops in Copilot CLI or let Copilot load it when relevant.
Copy the conceptops/ directory into the host's personal or project skills folder:
| Host | Personal | Project |
|---|---|---|
| Codex | ~/.agents/skills/conceptops |
.agents/skills/conceptops |
| Claude Code | ~/.claude/skills/conceptops |
.claude/skills/conceptops |
| GitHub Copilot | ~/.copilot/skills/conceptops |
.github/skills/conceptops |
Restart the host only if the newly installed skill does not appear.
Use $conceptops. English. Maybe a tiny service that helps independent cafés
predict tomorrow's pastry waste? I don't know if owners would use another app.
Give me the smallest useful package.
Use $conceptops. Work in English. The filtering screen in this inventory app is
slow and confusing. Decide whether we should replace the current table library.
I need a solution blueprint AND a low-fidelity prototype, but no code changes.
Use $conceptops. English. Need private semantic search across roughly 20,000
Markdown notes on one laptop. No cloud and no always-running service. Help me
choose an approach; implementation can wait.
See expected direction and pass conditions in evals/scenarios.md:
- Rough product idea
- Existing-project change
- Technology decision
- Language contract and AND/OR
- Direct execution scope gate
ConceptOps chooses the visual mode by what the artifact needs to prove:
- A real screenshot with a tutorial overlay for exact current behavior.
- A project-faithful mockup for a realistic product extension.
- A structural wireframe for flow or layout.
- A creative or AAA concept when the task calls for open visual exploration.
It can match an existing product closely without forcing every creative task into the current UI style.
evals/scenarios.md holds the synthetic demos and behavioral contracts in one place: prompt, expected direction, and pass conditions for each scenario.
- rough product idea and output restraint;
- existing-project reuse and scope boundaries;
- technology decision without premature implementation;
- language consistency and AND/OR component selection.
- direct execution without ConceptOps ceremony or run files.
These are behavioral contracts, not claims of universal model performance. The skill structure is also checked with the validator bundled with OpenAI's Skill Creator.
conceptops/
├── assets/
├── conceptops/
│ ├── agents/openai.yaml
│ ├── assets/
│ ├── references/
│ └── SKILL.md
├── FLOW_DIAGRAM.md
├── evals/
├── LICENSE
├── README.md
└── SECURITY.md
ConceptOps is instruction-only and has no backend or telemetry. The skill still treats external writes, public release, production access, private data, money, medical or legal decisions, minors, and physical actions as explicit trust boundaries.
Report vulnerabilities through GitHub's private vulnerability reporting. See SECURITY.md.
MIT. See LICENSE.