One short document that gives every agent and every skill in your library a shared frame of reference — built from evidence in your existing work, not from a workshop full of opinions.

What it is

A constitution is a short, always-loaded document that captures your company's foundational vocabulary, operating models, and principles in one place. Every AI agent on your platform reads it on every turn, and every other skill in your library cites it rather than restating it.

In practical terms: it makes your agents reason the way your company reasons — automatically, consistently, and without anyone having to re-explain how the business works for every new skill someone writes.

The constitution is the grammar of how your team operates. Leaf skills are the sentences. You don't want one document that's both the dictionary and the novel.

The problem it solves

If your team is building a skills library, you're already running into — or about to run into — four expensive problems:

01 Inconsistency

The same question, routed through two different skills, produces meaningfully different answers because each skill reasons from slightly different implicit assumptions about how the business works.

02 Slow skill creation

Every new skill starts from scratch on context. Authors either re-explain your operating model or accept that the agent might miss it. Both options bloat the library.

03 Silent drift

As skills get edited over months, your shared understanding subtly diverges. The definition of “activation” in one skill no longer matches another. By the time anyone notices, refactoring is a project nobody wants.

04 Experience loss

The actual model of how the company works lives in three senior people's heads. When they leave, it walks out with them. Procedures without their underlying principles are brittle.

How it works

The builder treats your existing skills library as the primary evidence source. Foundations live below conscious awareness — workshops produce aspirations, but the library produces facts. A foundation that already appears in fourteen skills is more real than a foundation someone declared in a meeting.

The team's role shifts from primary author to editor and gap-filler — which is the role they're genuinely better at.

Pass One - Scan

Read every skill file in the library. Extract candidate vocabulary, frameworks, principles, and contradictions, with each candidate cited back to the skills where it appears. The output is an evidence brief — no synthesis yet, just evidence.

Pass Two - Propose

Score each candidate on frequency, consistency, load-bearing weight, and stability. Sort into strong candidates (include), weak candidates (flag for validation), and rejected (only in one skill, or unstable).

Pass Three - Validate

A focused 60–90 minute session with two to four team members. The agent presents the evidence one section at a time; the team confirms, refines, rejects, resolves contradictions, and surfaces gaps — foundations the team uses but never wrote into any skill.

Pass Four - Draft & audit

Assemble the validated foundations into the constitution (under 300 lines). Run four audit tests: the traceback test, the newcomer test, the procedure test, and the grammar test. Ship only when all four pass.

Why scan-first beats interview-first. Memory is biased. Aspirations creep in. Workshops produce opinions; the library produces facts. Scanning first ensures every foundation in the constitution is traceable to skills already using it — nothing aspirational, nothing on faith.

What it produces

A single skill file, typically 80–200 lines, structured like this:

---
name: <team-or-org>-constitution
description: Foundational models, principles, and vocabulary used by every
  other skill in the library. Always-loaded; cited by leaf skills.
implicit: true
---

# <Team or Org Name> Operating Constitution

## Vocabulary
<Validated terms, one sentence each.>

## Models & Frameworks
<Validated frameworks, with stages explained briefly.>

## Principles
<Validated imperatives, under three lines each.>

## How Leaf Skills Should Cite This
<Two-line note on citation conventions.>

Hard length budget: under 300 lines. Above that, procedures have leaked in.

What changes when it's in place

Before

• Skills duplicate foundational context, slowly diverging
• New skill authors waste time re-establishing shared ground
• Agent answers feel “AI generic” rather than “your company”
• The operating model lives in three people's heads
• A model change requires updating dozens of skills
• Foundations are debated with no evidence

After

• One source of truth for foundations, cited by all skills
• Authors focus on the specific procedure, nothing else
• Agent answers carry your house style automatically
• The model lives in a document the whole team owns
• A foundational change is one edit, applied everywhere
• Foundations are surfaced from real library evidence

How to put it in place

Four phases. Read all four before starting — the mistakes that matter are the ones made in phase 1 because phase 4 wasn't yet visible.

Phase One - Install the builder skill

Drop the constitution-builder skill into your platform's skills folder. Confirm the agent running it has read access to your full skills library — scan-first only works if the agent can see what you have. Limit invocation to people who will actually facilitate the validation session.

Phase Two -Run the analysis

Hygiene-check your library (remove deprecated skills, resolve duplicates). Invoke the builder — it produces an evidence brief and a ranked proposal in 20–45 minutes. Review the proposal alone for 15–30 minutes before bringing the team in. Run a focused 60–90 minute validation session with two to four people. Draft and audit.

Phase Three - Install the constitution

Save it to your skills folder. Mark it always-loaded (frontmatter flag, system prompt include, or explicit reference from each leaf skill — in that order of preference). Add a one-line citation note to each existing leaf skill. Run a validation comparison — pick one skill, run three requests with and without the constitution loaded, confirm the outputs improve.

Phase Four - Long-term operation

Establish a custodianship model on day one — who proposes, who approves, how often you review. Run a quarterly re-scan of the library to surface new candidates and remove items that no longer have library evidence. When adding new leaf skills, check them against the constitution before merging.

Signals to watch

The constitution is working when:

• Leaf skills get shorter over time, not longer
• New team members ramp on skill-authoring faster
• Agent outputs feel on-brand without anyone tuning
• Disagreements surface earlier and resolve cleaner
• Each quarterly re-scan finds fewer surprises

It's decaying when:

• It keeps growing past 300 lines
• Leaf skills start ignoring it
• You can't remember when it was last reviewed
• It contains items no skill ever cites
• It contradicts how the team actually works

Bad signals are not crises — they're prompts to schedule the next review sooner.

Who benefits

• Your agents
  • Get a consistent frame of reference. They reason the way your company reasons, on every turn, without anyone re-prompting.
• Your team
  • Gets a shared written model grounded in what they have actually written, not in what they remembered in a workshop. Disagreements surface earlier; skill-writing gets cleaner; reviews get sharper.
• Your library
  • Stays maintainable as it grows. New skills don't add weight to old ones. The library scales from twenty skills to two hundred without falling apart.
• Your clients
  • Get coherent, on-brand work regardless of which skill or which team member surfaced it. The work feels like it came from one mind because, in a real sense, it did.
• Your leadership
  • Gets institutional knowledge that doesn't depend on any single person. The thinking is in the file, not in someone's head.

What it is not

Setting expectations clearly matters — misuse will kill the value before it starts. The constitution is not:

• A playbook. Specific tactics and how-tos belong in leaf skills.
• A mission statement. It's operational, not aspirational.
• A replacement for human judgment. Humans still review outputs and handle edge cases.
• A static document. It's reviewed quarterly, re-scanned against the library each review.
• A long document. Past 300 lines, procedures have leaked in.
• The product of one workshop. It's built from evidence, not memory.

It is a small piece of operational infrastructure. The whole point is that it's boring — durable enough to lean on, plain enough that nobody argues about formatting.

The underlying idea

Most companies have a clear model of how they operate — but it lives implicitly, in conversations, in senior judgment, in pattern-matching that took years to develop. Writing it down was always a good idea. With agents now doing real work alongside the team, writing it down is no longer optional. The agents need it as much as new hires do.

But you cannot write a useful constitution from scratch. The foundations that matter are already showing up in your library — duplicated across skills, restated slightly differently each time, sometimes contradicting each other. The work isn't to invent a new model. It's to surface the model that's already implicit in what your team has written, validate it, and make it explicit in one place.

The constitution is the thing you wish you had written down years ago, finally written down — and grounded in the work you have actually done.

Short. Plain. Durable. Cited by everything that touches your business.