NOKTA — NAIM's Orchestrated Knowledge-To-Artifact

The spec layer of the NAIM loop. NAIM builds. NOKTA thinks first.

1 file. 0 human review. CI decides.

Nokta is part of the NAIM ecosystem — Naim Agentic Iterative Mobile. NAIM iterates on apps; NOKTA defines what to build before the first line of code.

A single idea spark — a dot — enters the system. Guided LLM questioning expands it into a structured product spec. Dot. Line. Paragraph. Page.

What?

program.md is the Karpathy-style spec file for the Nokta mobile app. This file serves as a complete specification from which AI agents (Claude Code / Codex) can build the application from scratch.

Humans write the spec, machines write the code.

How It Works?

There are three contribution paths:

Path A — Develop a section of program.md (SPEC EDITING)

Edit existing sections in program.md to improve specification quality.

Pick a section (see Section List below)
    ↓
Open a branch: section/04-data-contracts
    ↓
Edit the relevant section in program.md
    ↓
Open a PR to main
    ↓
CI scores automatically (0-100)
    ↓
Score ≥ current score on main → ✅ Auto-merge
Score < current score on main → ❌ Auto-reject
    ↓
Fix, push, try again

🤖 Fully automated. No human review. The metric decides. Score never drops.

Branch naming: section/<number>-<description> (e.g., section/04-data-contracts)

Path B — Propose a new feature spec (SPEC CREATION)

Create new feature specifications as standalone documents.

Copy specs/TEMPLATE.md → specs/your-feature.md
    ↓
Open a branch: spec/your-feature
    ↓
Fill in all 5 sections, delete all > TODO lines
    ↓
Open a PR to main
    ↓
CI scores against spec_generic.yml (0-100)
    ↓
First PR for this file → ✅ Establishes baseline (score ≥ 0)
Subsequent PRs → must match or beat the baseline
    ↓
Fix, push, try again

🤖 Fully automated. No human review. First merge sets the bar. Score never drops.

Branch naming: spec/<feature-name> (e.g., spec/user-profile)

Path C — Implement code from specs (IMPLEMENTATION)

Write actual code (TypeScript, React Native) based on approved specs.

Read program.md or specs/*.md (approved spec)
    ↓
Open a branch: implement/feature-name
    ↓
Write code that implements the spec exactly
    ↓
Run hard gates locally:
  • npx tsc --noEmit (TypeScript)
  • npx eslint . --ext .ts,.tsx (ESLint)
  • npm test (Jest tests)
  • npx expo export --dump-sourcemap (Bundle size < 2MB)
    ↓
Open a PR to submissions branch
    ↓
CI runs evaluation workflow:
  • Validates hard gates
  • Runs golden flow tests (if available)
  • Scores quality/effort
  • Posts scorecard to PR
  • Updates LEADERBOARD.md
    ↓
All gates pass → ✅ Eligible for human review
Any gate fails → ❌ Fix and push again
    ↓
👤 HUMAN REVIEW REQUIRED (even if CI passes)
  • Maintainer reviews code quality
  • Security check (no API keys, XSS, injection)
  • Architecture compliance
  • Spec compliance
    ↓
Approved → 🎉 Merged to main

👤 Human approval required. CI validates, but human reviews for security, architecture, and design patterns.

Branch naming: implement/<feature-name>, fix/<bug>, test/<test-name>

Evidence required: Screenshots, screen recording, or Expo Go QR code/APK

Summary: Which Path Should I Use?

Path A/B = Spec work → Auto-merge
Path C = Implementation → Human review required

Quick Start

# 1. Fork or clone the repo
git clone https://github.com/seyyah/nokta.git
cd nokta

# 2. Create a branch (choose your section number)
git checkout -b section/04-data-contracts

# 3. Open program.md, find your section, write it
# Each section has <!-- COMMENT --> instructions inside

# 4. Check your score (optional, local test)
pip install pyyaml
python scripts/section_score.py --section 4

# 5. Commit + push
git add program.md
git commit -m "feat(section-04): add data contracts"
git push origin section/04-data-contracts

# 6. Open a PR on GitHub → CI runs automatically → wait for result

Section List

Scoring

Each section has a checklist (checklists/section_XX.yml). CI reads this checklist and evaluates the section:

• Structural checks: Are required headings present? Is the minimum word count met?
• Content checks: Are specific concepts/patterns present? (e.g. TypeScript code block, JSON schema)
• Each check carries a weight. Scored out of 100 points.

Scoring rule: Your score cannot drop below the current score on the main branch. Equal or higher means you merge.

Test Locally

# Single section (Path A)
python scripts/section_score.py --section 4

# All sections (Path A)
python scripts/section_score.py --all

# CI-formatted output (Path A)
python scripts/section_score.py --all --ci-comment

# Score a spec file (Path B)
python scripts/section_score.py --spec-file specs/your-feature.md

Spec Contributions (Path B)

Use Path B to propose a new Nokta mobile app feature as a standalone spec.

# 1. Copy the template
cp specs/TEMPLATE.md specs/your-feature-name.md

# 2. Create a branch
git checkout -b spec/your-feature-name

# 3. Fill in all 5 sections, then delete every > TODO line

# 4. Check your score locally
python scripts/section_score.py --spec-file specs/your-feature-name.md

# 5. Commit + push
git add specs/your-feature-name.md
git commit -m "spec(your-feature): describe feature"
git push origin spec/your-feature-name

# 6. Open a PR → spec-ratchet CI runs automatically

Spec Checklist (checklists/spec_generic.yml)

Ratchet rule: First PR for a new spec file — passes with any score > 0. Subsequent PRs for the same spec file — must match or beat the score on main.

Rules

1. Only edit program.md (Path A) or add a new specs/*.md file (Path B). Do not edit checklist YAMLs, CI workflows, or scripts — these are IMMUTABLE INFRA.
2. Section 1 (IDENTITY) is already complete. Use it as a reference and follow its format.
3. Multiple people can write the same section. The highest score gets merged.
4. Delete TODO placeholders. As long as > TODO: lines remain, the section score stays low.
5. AI tools are welcome — vibe-writing is fine. But you still need to pass the checklist.
6. English or Turkish — both are supported by the checklist.

FAQ

Q: What if someone else is writing the same section?
First to merge wins. The next PR must beat that score. Ratchet.

Q: My PR was rejected — what should I do?
Read the CI comment — it will show which checklist items failed. Fix them, push, and CI runs again.

Q: Section 1 is already complete — can I improve it?
Yes — but you need to beat its current score (100/100). Delete content and the score drops; you'll be rejected.

Q: Can I contribute to multiple sections?
Yes — each in its own branch + its own PR.