Skip to content

Design intent

This page states the direction the framework is being built towards. Nothing on it is built yet unless a section says so. Each section links the epic that builds it. A section moves to Principles in the change that makes it true, so this page shrinks as the framework catches up with it.

What kac is for

kac keeps an organisation's engineering knowledge as validated, identified and versioned Markdown, and generates what each agent needs at each moment of the work.

The first half is true today. The second half is the direction. Today kac packages a corpus as one Claude Code plugin, and its skills look a record up when an agent decides to. The sections below say what replaces that.

The moments knowledge serves

Knowledge is used at distinct moments in the work, by different readers. Each moment needs a different slice of the corpus, delivered a different way.

Moment Who What they need
Design an agent or engineer choosing an approach the decisions already made, the policies that bind, the targets to meet
Build an agent writing code every rule that applies to the code being changed, wherever the rule is filed
Review a reviewer, human or AI the rules to check the change against, each with an id to cite
Operate whoever is on call the runbook, and whether anybody has hit this failure before
Contribute an agent or engineer writing back the level of detail a record is written at, and the route for a finding
Assure the corpus owner, or an auditor which policies have no standard, which rules no control proves, what is stale
Learn a person new to the team the domain and the house style, read as a wiki

The use cases for each moment, and for each type, are written as features under trials/use-cases/. Each one names the trial case that measures it. Tracked in epic #940.

What an agent does with each type

Each type has an intent: what an agent should do with a record of it. The intent sets the tone of the skill that delivers the type, and of anything generated from it.

Intent What the agent does Types
Follow follows it closely, as a person would runbooks, and fixes once the symptom matches and somebody has verified the fix
Flow follows the shape of the work and judges each step processes
Nudge treats each MUST as the baseline, and each SHOULD as a default it may depart from, saying why standards, policies
Respect does not go against it silently: honours it, or raises it ADRs, NFRs once agreed, tools (only an approved tool for new work), deviations inside their scope and period
Inform uses it as context where it is relevant glossary, services, data, integrations, offerings, explanations, postmortems, controls, reports, draft fixes

A Nudge type grades each clause by its keyword, so one record states obligations and latitude side by side. A Respect type is one obligation as a whole: a decision, a target, a choice of tool, or an accepted departure.

Each row is a hypothesis until a trial confirms it. The per-type index lists the use cases that test each row.

Cadence and delivery

A type whose records change once a year can be built into what ships. A type whose records change with every incident is better read live. Story #946 decides how the fast types reach an agent, using the cadence each type page states. Tracked in epic #940.

The example corpora as exemplars

The corpora under examples/ show what a good record of each type looks like, and adopters model their own records on them. Each type's worked examples are compared with well-regarded public ones. Where a record differs from good public practice, the corpus says why. Almost every rule in the example standards is still a MUST, and they are rewritten first, in the shape Principles describes. Tracked in epic #930.

Delivery at the moment of use

The plugin gives an agent the smallest set of relevant knowledge at the moment the agent is likely to use it. Text loaded early and left unused draws on the agent's attention for the rest of the session. So the plugin pushes a short pointer at the moment of use, and the agent reads the full text the pointer names. This follows Anthropic's guidance on context engineering and on progressive disclosure in skills.

Moment When it arrives What reaches the agent Where the full text is
Build the session's first tool call an index of every standard rule, by title the standard, read through a lookup
Review somebody asks for a review the review view, in full, in a reviewer of its own the review view
Design, operate, contribute, assure the agent decides to look something up a lookup skill's answer the record
Learn a person opens the corpus the published corpus, read as a wiki the record

A moment whose lookups keep missing in a trial moves to a pointer the plugin pushes. Tracked in epic #940.

Views generated for each moment

kac generates a view (an export shaped for one use) when it bundles the plugin. The corpus declares each view in its plugin manifest. Each entry in a view states its record's status, so a draft rule reads as a draft.

  • The build index lists every standard rule by title, under its standard's id. A hook the plugin ships pushes it once to each agent in a session, and again after a compaction.
  • The review view lists each rule's obligations: each MUST to check, and each SHOULD to note a departure from. It lists the controls that check each standard, with each control's mechanism and status. A control can name a whole standard, so the reviewer reads the control's record to learn which obligations it checks.

The per-type export stays the format another corpus imports. Tracked in epic #940.

Formats other agents read

Claude Code comes first, and nothing generated is exclusive to it. Skills use the open SKILL.md format, and each view is Markdown.

The plugin ships its own reviewer, modelled on the Standards review in code-review. A repository can also copy the review view to where its own review tools read standards, such as CODING_STANDARDS.md.

No record states which files a rule applies to, so nothing loads by path. Issue #993 keeps that option open until the build trial shows whether the index is enough. Tracked in epic #940.

Where to go next

Principles says what the framework does today, and why.