cognitive-skills

Capture how a person thinks in their role, then make any AI tool think the same way for the same work.

A toolkit for encoding human judgment as portable AI artifacts. Run an interview, get a cognitive clone, paste it into Claude, ChatGPT, Cursor, Copilot, Gemini, or any tool that accepts a system prompt. The outputs are plain Markdown + YAML — no vendor lock-in, no SDK, no API key.

What ships in this repo:

• 15 role-specific Cognitive Extraction Engines (CEETs) — interviews + synthesis templates for backend, frontend, devops/SRE, data, product, UX, UI, copy, marketing, sales, customer success, finance, legal, people-ops, and founder/CEO.
• A production-grade Jira requirements pipeline — six commands that turn ambiguous tickets into validated functional contracts with traceability, an LLM-augmented refinement step, and 85% test coverage.
• A deterministic eval harness — three benchmark tasks, no judge model, release-gate ready.
• Three meta-skills — auto-routing (autodiscover), sub-agent orchestration with validated handoff contracts, and the impersonator for drafting CEET packs from public evidence.

60-second example

git clone https://github.com/CMolG/cognitive-skills.git
cd cognitive-skills

Open examples/ready-to-use/backend-netflix-tech-blog/cognitive-profile.md, paste it into your AI tool's system prompt, and ask:

"Review this pull request for a database migration that renames users.email to users.primary_email, backfills data, and adds a unique index."

You will get a review framed in expand-and-contract phases, dual-write windows, lock-build strategy, and rollback toggles — instead of a generic "add tests, check rollback" checklist. See the before/after table below for the exact difference.

For the Jira pipeline specifically, watch the 90-second walkthrough:

bash examples/ready-to-use/demo-jira-pipeline.sh

It runs against a synthetic ticket fixture (no Jira credentials needed). Record cleanly with asciinema rec — see the script header for the exact command.

What's new

• 1.2.0 (in progress) — Jira pipeline hardening: rule-based core fixed, LLM-augmented mode added, 76 tests at 82% coverage. Deterministic eval harness with three benchmark tasks. Meta-skills now have validated contracts (OrchestrationPlan, SubAgentResult) and a provenance enforcer for the impersonator. See CHANGELOG.md.
• 1.1.x — English documentation pass; autodiscover and sub-agent orchestration skills.
• 1.0.0 — Initial public release.

Visual usage guide for SKILLS

Each skill has its own subsection with a distinct Material icon, description, target audience, and usage examples.

autodiscover

What it does: Discovers and automatically routes flows/skills according to the task.
Focused on: Users who want to reduce manual skill selection.

• Analyze this objective and tell me which skill should run first.
• Route this request to the correct flow without me choosing a role.

ceet-backend-engineer

What it does: Cognitive extraction for backend engineering (data, APIs, invariants).
Focused on: Backend engineers and technical reviewers.

• Generate a backend cognitive profile from an interview.
• Activate an AI environment to review migrations and API contracts.

ceet-copywriter

What it does: Captures copy criteria (voice, structure, conversion).
Focused on: Copywriters and content/brand teams.

• Extract my tone rules for B2B landing pages.
• Create prompts for rewrites with brand voice.

ceet-customer-success

What it does: Models decision-making for onboarding, retention, and expansion.
Focused on: CSMs, support leads, and post-sales teams.

• Synthesize my at-risk account playbook.
• Generates health score prioritization rules.

ceet-data-analytics

What it does: Extracts analysis, experimentation, and metrics frameworks.
Focused on: Analysts, data practitioners, and growth teams.

• Converts my analysis method into a cognitive clone.
• Creates commands to review hypotheses and biases in dashboards.

ceet-devops-sre

What it does: Captures SRE/DevOps criteria for operations, incidents, and reliability.
Focused on: SREs, platform engineers, and on-call leads.

• Models how I decide rollback vs forward-fix.
• Generate rules for postmortems and high-risk changes.

ceet-financial

What it does: Structures finance heuristics (models, forecast, controls).
Focused on: Finance teams and founders focused on unit economics.

• Extracts my logic for quarterly forecasting.
• Create prompts to validate pricing and margin assumptions.

ceet-founder-ceo

What it does: Synthesizes strategy, narrative, and organizational design criteria.
Focused on: Founders, CEOs, and strategic staff.

• Documents my process for deciding strategic bets.
• Generate an AI environment to prepare capital decisions.

ceet-frontend-engineer

What it does: Extracts decision patterns in UI state, rendering, and accessibility.
Focused on: Frontend engineers and web product teams.

• Creates a clone for frontend performance review.
• Defines accessibility and interaction quality rules.

ceet-legal-compliance

What it does: Captures legal risk, policy, and compliance criteria.
Focused on: Legal ops, compliance officers, and risk teams.

• Extracts my contractual review checklist.
• Generates directives for regulatory risk classification.

ceet-marketing

What it does: Models reasoning for positioning, channels, and funnels.
Focused on: Performance marketers and brand/growth leads.

• Synthesizes my multichannel acquisition strategy.
• Creates prompts for funnel and messaging audits.

ceet-people-ops

What it does: Extracts hiring, performance, and culture criteria.
Focused on: HR, People Ops, and talent managers.

• Converts my evaluation framework into operational rules.
• Generate artifacts for onboarding and professional development.

ceet-product-manager

What it does: Captures prioritization, discovery, and roadmap frameworks.
Focused on: Product managers and product leads.

• Extracts how I prioritize between technical debt and features.
• Create commands to prepare RFCs and scope decisions.

ceet-sales

What it does: Structures discovery, objection handling, and closing playbooks.
Focused on: SDR/AE, consultative sales, and revenue teams.

• Model my process for qualifying enterprise opportunities.
• Generates objection response guides by segment.

ceet-sub-agent-orchestration

What it does: Defines subagent coordination and distribution of cognitive tasks.
Focused on: Teams that design multi-agent systems.

• Designs agent orchestration for technical auditing.
• Set handoff rules between specialist agents.

ceet-ui-designer

What it does: Captures visual criteria for design systems, components, and motion.
Focused on: UI designers and design systems teams.

• Extract my principles for cross-product visual consistency.
• Generates prompts for hierarchy and contrast reviews.

ceet-ux-researcher

What it does: Models behavioral research thinking and findings synthesis.
Focused on: UX researchers and product discovery squads.

• Converts my interviews into product decision rules.
• Generates a template for synthesizing behavior patterns.

impersonator

What it does: Initializes simulated CEET packs from public evidence or repository history.
Focused on: Users who need a quick draft without a live interview.

• Creates an initial pack for a known technical author.
• Generate a cognitive profile draft from a repository.

jira-agentic-requirements-pipeline

What it does: Structures an agentic requirements pipeline based on Jira.
Focused on: Product/engineering teams with Jira-centered operations.

• Define a flow from intake to refined ticket.
• Generates quality policies for Jira user stories.

The CLI is six subcommands. The top-level help lists them:

$ python3 jira-agentic-requirements-pipeline/scripts/jira_pipeline_cli.py --help
usage: jira_pipeline_cli.py [-h]
                            {fetch-issue,discovery,generate-questions,collect-input,resolve-contract,base-branch-plan}
                            ...

Agentic requirements pipeline for Jira

positional arguments:
  {fetch-issue,discovery,generate-questions,collect-input,resolve-contract,base-branch-plan}
    fetch-issue         Fetch issue from Jira
    discovery           Analyze missing business requirements
    generate-questions  Generate prioritized business questions
    collect-input       Capture business answers with resumable state
    resolve-contract    Resolve a functional contract from answers
    base-branch-plan    Generate base branch implementation plan

Each subcommand has its own --help — for example, generate-questions --help documents the --baseline-budget and --signal-budget flags. See the skill's SKILL.md, EXAMPLES.md, and TROUBLESHOOTING.md for the full quick start.

Three ways to use this repo

1. Use a ready-made pack — open examples/ready-to-use/backend-netflix-tech-blog/ (or any other pack) and copy cognitive-profile.md into your AI tool's system prompt. Provenance is in each pack's evidence-map.md.
2. Draft a pack from public evidence — point any skill-aware AI at impersonator/SKILL.md. It generates a simulated pack for a public figure or repo author and runs through impersonator/scripts/validate_pack.py before shipping.
3. Run a real CEET interview — open the role's SKILL.md and follow the interview flow. Output is a first-person cognitive clone.

Before and after: base model vs CEET pack

Same prompt, two arms. Verifiable with the harness in evals/.

Prompt:

"Review this pull request for a database migration that renames users.email to users.primary_email, backfills data, and adds a unique index."

Base model (generic)	CEET pack (backend-netflix-tech-blog)
Recommends adding tests and checking migration rollback.	Breaks migration into expand/contract phases and explicitly asks for dual-write windows before rename cutover.
Mentions performance and downtime in general terms.	Calls out index build strategy, lock behavior, query plan verification, and rollback toggles under active traffic.
Suggests validating data after migration.	Requests invariant checks (null, duplicate, stale writer paths), replay safety, and observability signals for each phase.
Gives a broad checklist.	Prioritizes blast radius controls: canary rollout, feature flags, and explicit fail-fast criteria tied to SLO/error budget impact.

Harness numbers (seed run, synthetic outputs)

The eval harness scores three benchmark tasks with deterministic metrics: required-phrase coverage, required-section coverage, and TF-cosine voice alignment vs a per-task corpus. Composite is the mean of the three.

Task	baseline	ceet	generic	Δ ceet−baseline
engineering-pr-review	0.59	0.85	0.44	+0.27
product-prd-draft	0.78	0.81	0.73	+0.03
copy-headlines	0.54	0.77	0.41	+0.23

The seed outputs under evals/results/seed-2026-05-07/ are synthetic — written to self-test the harness, not as a benchmark. Replace them with outputs from your own model and re-run python3 evals/scripts/run.py --run-id <your-id> to produce real numbers. The release gate (--gate --min-delta 0.05) requires ceet to beat baseline on at least 2/3 tasks.

The 15 roles

Folder	Role	Focus
ceet-backend-engineer	Backend Engineer	Systems, data, invariants, failure modes, APIs
ceet-frontend-engineer	Frontend Engineer	UI state, rendering, accessibility, client perf
ceet-devops-sre	DevOps / SRE	Infra, CI/CD, observability, incident response
ceet-data-analytics	Data / Analytics	Metrics, hypotheses, experimentation, SQL/Python
ceet-product-manager	Product Manager	Prioritization, discovery, roadmap, stakeholders
ceet-ux-researcher	UX Researcher (conductual)	Behavior, interviews, synthesis, jobs-to-be-done
ceet-ui-designer	UI Designer	Visual system, components, motion, craft
ceet-copywriter	Copywriter	Voice, structure, conversion, brand tone
ceet-marketing	Marketing	Positioning, channels, funnels, growth loops
ceet-sales	Sales	Discovery, objections, pipeline, closing
ceet-customer-success	Customer Success	Onboarding, retention, expansion, health signals
ceet-financial	Finance	Models, unit economics, forecasting, controls
ceet-legal-compliance	Legal / Compliance	Contracts, risk, regulation, policy
ceet-people-ops	People Ops / HR	Hiring, performance, culture, policy design
ceet-founder-ceo	Founder / CEO	Strategy, capital, narrative, org design

Independent skills

Folder	Purpose
impersonator	Initialize any CEET role with a simulated draft pack inferred from public-figure evidence or repository-author commit/code history (no interview).

The shared loop (every CEET follows this)

┌──────────────┐   ┌──────────────┐   ┌────────────────┐   ┌──────────────────┐
│  INTERVIEW   │ → │   EXTRACT    │ → │   SYNTHESIZE    │ → │    ACTIVATE      │
│ deep, Q+A    │   │ patterns &   │   │ cognitive clone │   │ portable AI env  │
│ role-scoped  │   │ heuristics   │   │ + decision map  │   │ for any AI tool  │
└──────────────┘   └──────────────┘   └────────────────┘   └──────────────────┘

See METHODOLOGY.md for the full methodology and each ceet-<role>/ folder for the role-specific interview script, synthesis logic, templates, and examples.

Using a CEET in any AI tool

The outputs of every CEET are two files:

• cognitive-clone.md — a portable, vendor-neutral description of how the person thinks in their role.
• ai-environment.md — a system-prompt-ready configuration you can paste into:
  • Claude Projects / custom instructions
  • ChatGPT custom GPTs / instructions
  • Cursor / Copilot rules files (.cursorrules, .github/copilot-instructions.md)
  • Gemini Gems / system prompts
  • Any other AI tool that accepts a system prompt or context file

How to run a CEET

Pick the folder that matches the role, open its SKILL.md, and follow it. If you're using this inside an AI tool that supports Anthropic-style skills, the skill will trigger automatically when you ask for a cognitive extraction for that role.

See docs/how-to-use.md for the full step-by-step guide.

Each role folder structure

Every ceet-<role>/ folder is fully self-contained:

ceet-<role>/
├── README.md                     # Role overview, interview flow, output artifacts
├── SKILL.md                      # AI-triggerable skill definition
├── interview/
│   └── questions.md              # Role-specific interview questions
├── templates/                    # Role-specific artifact templates
│   ├── agents/                   # 5 agent templates (e.g., code-reviewer, debugger)
│   ├── skills/                   # 5–6 skill templates (e.g., style-enforcer, test-writer)
│   ├── commands/                 # 8–10 command templates (e.g., /review, /debug)
│   ├── rules/                    # 5–6 global rule templates (e.g., coding-standards)
│   ├── hooks/                    # 3 hook templates (engineering roles only)
│   └── cognitive-profile.md      # Full cognitive profile template
└── examples/
    └── README.md                 # How to generate and use example outputs

All templates use {directives.domain.field} placeholders that are injected from the cognitive profile during generation — zero conditional logic.

Documentation

Document	What it covers
METHODOLOGY.md	The four stages: interview → extract → synthesize → activate
docs/how-to-use.md	Step-by-step guide for running a CEET extraction
docs/synthesis-rules.md	12 strict rules for converting interview responses to cognitive profiles
docs/tool-integration.md	How to load outputs into Claude, ChatGPT, Cursor, Copilot, Gemini, and more
docs/extending.md	How to add a new role pack to the toolkit
impersonator/README.md	Independent non-interview skill for simulated CEET initialization

Project status

This toolkit is under active construction. Each CEET folder is self-contained and will become a standalone skill package.