ASDF
progress_activity

KogniVera

AI Software Development
Framework

ASDF is KogniVera's method for building software with AI agents you can trust. Instead of ad-hoc prompting, every project runs on three governing artifacts — a Constitution that sets inviolable engineering rules, a roster of Agents with crisp scope, and executable Specs that define each unit of work. Agents refuse to write code until the constitution and specs exist; humans review and approve every artifact.

Access is restricted to KogniVera members — sign in to open the builders and your projects.

The Three Phases

01gavel

Inception

The Constitution sets the technology floor, compatibility, security & compliance. Specs decompose the work.

02engineering

Construction

Execute specs — the right agent per spec, run in parallel, each with a clear contract.

03monitoring

Operation

Observe, measure, and evolve. Artifacts and metrics keep humans in control of the verdict.

Harness Agnostic

The framework is independent of any single AI harness. The same phases, specs, and agent discipline run on Gemini CLI, Claude Code, OpenCode, Pi, Copilot, and others — some carry a brand implicitly, others are brand-agnostic and let you point them at any model. It even runs fully offline against local models via Ollama. Only minor naming conventions differ between harnesses, and everything works the same.

terminalGemini CLIsmart_toyClaude CodecodeOpenCoderobot_2PiassistantCopilotdnsLocal models via Ollama

One playbook, any harness

Constitution, specs, and agent roster stay identical regardless of tool.

Minor naming differences

Harnesses label agents, skills, and config slightly differently — the workflow is unchanged.

Swap without rework

Change harness, or mix several at once, without rewriting your specs or process.

What Can Go Wrong — and How ASDF Prevents It

AI agents fail in predictable ways. ASDF exists because each failure mode has a working countermeasure.

warningWatch out

layers_clear

Ignoring your tech stack

Models default to their favorite stack — they will quietly swap your chosen framework, database, or library for what they saw most in training unless the choice is pinned down in writing.

psychology_alt

Hallucinations

Confidently invented APIs, packages that don't exist, config options that were never real. Plausible-looking code is not the same as correct code.

call_split

Unrequested scenarios

Agents love to be helpful — extra features, edge cases, and flows nobody asked for creep into the code, growing scope and review surface silently.

visibility_off

Silent assumptions

When a requirement is ambiguous the model picks an interpretation and moves on without telling you. The gap surfaces weeks later as a defect.

shieldGuardrails

gavel

Always refer to the constitution

The constitution is the supreme engineering authority. Every agent reads it before writing a line of code — stack, standards, and forbidden choices are pinned there, not in prompts.

smart_toy

Use the right agent for the work

Match each spec to a role-scoped agent — a senior dev engineer for features, a code reviewer for diffs, a security specialist for auth. One generalist agent doing everything is how scope drifts.

difference

Review your git repo constantly

Diff every change the agent makes. Check the repo after each task and make sure the changes are exactly what was requested — nothing more, nothing missing.

fact_check

Specs with acceptance criteria

Every requirement carries Given/When/Then criteria, and unanswered questions are listed as Open Questions — never invented. If it's not in the spec, it doesn't get built.

Guiding Principles

Specs over prompts

The spec is the reviewable unit of work, not the prompt.

No silent decisions

Every default and assumption is visible, editable, and recorded.

Humans own the verdict

AI drafts; people review, edit, and approve before anything ships.

Foundations

ASDF isn't invented from scratch — these are the ideas it builds on.