← Shiv Rajappa
01 / 10
Use buttons, arrow keys, or swipe
An interactive product thesis

It's only words and words are all I have

Why grammar is reshaping product development.
When software can interpret language and take action, product teams must design the vocabulary, rules, and evidence behind every outcome.

NounKernel
VerbBrief
AdjectiveSignificant
RuleCite every claim
Brief me every morning on what significantly changed, citing every claim.
The case

Compountier helps product and engineering leaders be at the frontier of what matters in new product development using AI

Compountier is a research product that follows 165 sources, each source is extracted into kernels of knowledge

165 sources

Videos
Articles
Release notes
Papers
broken into

Kernels

Own the context. Rent the model.talk
Verification is the durable moat.paper
Model prices fell again this quarter.release note
compounded into

One morning brief

What changed todayLive
Written fresh every morning. Read today’s →
The main objective:-

Brief meVerb every morningAdverb on what significantly changedAdjective citing every claimRule

Read today’s briefcompountier.com
The problem

When an agent gets it wrong, we blame the model or the associated harness.

The first version of Compountier ran that sentence with three of its four parts undefined. Every tool update landed in one bucket called Tooling. About half of the extracted facts carried no concept labels, because nothing stated what a concept label meant.

Brief me Verb How far does briefing go? The agent could rank the items, cut them, or summarize everything. Nothing said where the cutting stops.
every morning Adverb This one was written down. The window is fixed at the last 24 hours. The cutoff falls at the same time every day.
what significantly changed Adjective What makes a change significant? A rebrand is loud, and a price cut is quiet. The early version had no test, so it treated both as the same kind of news.
citing every claim Rule What counts as a citation? A link, a name, and a quote set three different standards. No code checked for any of them.
Six parts · this failure came from three Nouns Verbs Adjectives Adverbs Sentences Rules

The failures typically result from a lack of definition and bounded constraints that can be observed, evaluated and tested against.

The investigation

Observability & traceability uncover the gaps

LangSmith records the steps between a request and an answer as a trace, and the trace surfaces the gaps.

The challenge ✗ The rebrand led

Brief meno bounds every morningdefined ✓ on what significantly changedno test citing every claimnot enforced

The trace of that morning

1
Pull the last 24 hours: 41 items
Ok
2
Sort significant from minor

No definition existed. The model ranked the items by how loud they sounded.

Guessed
3
Draft the brief
Done
4
Cite each claim
Skipped
5
Publish the digest
Done

The brief that morning produced

A morning from the early version
  1. 1A major AI lab unveils a new name and logo.
  2. 2A new benchmark ranks the popular coding agents.
  3. ✗A major model provider cuts its API prices by 80%.Cut for length.
The brief reads cleanly, and no claim carries a citation. The one item that changes what you can afford to build sits in the cut material.
What LangSmith does with itLangSmith observability →
Step 1: The trace

The trace above recorded the run. Step two had no test to pass, so the model guessed, and only the trace shows the guess.

Step 2: Saved as a case
CaseA rebrand and an 80% price cut land on the same day.
ExpectedThe price cut leads the brief.
GotThe rebrand led. Fail

Every morning that goes wrong gets saved like this. Twelve are saved so far.

Step 3: All twelve, re-run
4 of 12mornings produce the right brief

Each dot is one saved morning. The early version gets eight of them wrong.

The fix

Clarifying and articulating the bounded constraints and evaluation criteria make it cleare

LangSmith proved the fix by re-running the evaluation set.

The same sentence, rerun ✓ The price cut led

Brief mebounded ✓ every morningdefined ✓ on what significantly changedtested ✓ citing every claimenforced ✓

"Significant", as it now ships in the codebase

CountsCapability release, model release, breaking change, deprecation, pricing change, substantive tooling.
Does notBug fix, chore, docs, a rebrand.
Concepts tooAll twelve concept labels now carry written definitions, with examples that qualify and examples that do not.
EnforcedCode files every entry before a model writes a word.

The brief the same morning produces now

The same morning, current version
  1. 1A major model provider cuts its API prices by 80%.k1
  2. 2A new benchmark ranks the popular coding agents.k2
  3. 3A major AI lab unveils a new name and logo.k3
Every claim carries a numbered citation. The save step checks that each citation points to a real kernel before the brief can publish.
LangSmith re-runs the same twelve mornings
Step 1: The trace

The same run, recorded again after the definition landed. No step is left with a guess to make.

Step 2: The same case
CaseA rebrand and an 80% price cut land on the same day.
ExpectedThe price cut leads the brief.
NowThe price cut leads. ✓ Pass
Step 3: All twelve, re-run
12 of 12mornings produce the right brief

If a change ever breaks one of these mornings, the release stops.

The model never got smarter. The sentence did.  Open the live brief that this sentence produces →

What changed

Responsibility sits across five areas.

Four actors shape the morning brief, while the trace observes the whole interaction.

Select an actor for its role
Shared objectBrief me every morning on what significantly changed, citing every claim

User

Intent and authority

Model

Interpretation

Harness

Tools and boundaries

Code

Validation and execution

Trace observes every transition

Request · evidence · decision · tool · outcome

UserSets the stakes

The user asks for the morning brief and owns what significant must mean: consequence to their stack, not noise volume.

The bigger picture

Graphical software gives people paths. Agentic software constructs them.

This was never just a Compountier story. Every agentic product moves composition from screens and clicks into runtime interpretation.

Graphical software

The interface owns the route
1 User
2 Open the feed
3 Skim today
4 Judge it

The person learns the product's language and follows its sequence.

Agentic software

Intent creates the route
User intent
Discover
Triage
Extract
Cite

The model translates human language into the product's executable vocabulary.

The checklist

Six parts to define product interactions with grammar.

Tap a word type to change the visual and explanation
Compountier briefs the significant developments every morning, citing every claim.
Objects

Nouns define what exists.

    Your turn

    A product sentence becomes an executable capability.

    Compose one of your own. Choose an action, a condition, and an object; the system must make the resulting sentence safe and real.

    Brief me every morning on the significant developments, citing every claim.
    The model composes the sentence. Code enforces the citation rule.
    The product discipline

    Design the language of action.

    A capable model needs a clear vocabulary, enforceable rules, and evidence from every sentence it constructs.

    1
    NounsDefine what exists.
    2
    VerbsInventory every action.
    3
    AdjectivesSpecify consequential conditions.
    4
    ResponsibilitySeparate judgment and enforcement.
    5
    TracesInspect intent to consequence.
    6
    EvaluationImprove grammar from evidence.
    The interface tells us what a person can do. The trace tells us what an agent did. Grammar design connects the two.

    Compountier's sentences are about AI intelligence. Yours are about invoices, tickets, campaigns. The grammar is the same.  See the worked case → · Read a trace with LangSmith →