Agent skill · v1.0.0 · July 21, 2026

Research that survives the chat.

The Internet Research Steward converts web discovery into governed, versioned, retrievable knowledge—with provenance, freshness controls, explicit human approval, and evaluation built in.

Download skill See the system

Dependency-free SHA-256 manifest Human approval boundary

DiscoverBounded search and source capture INPUT

NormalizeTyped claims, decisions and evidence OBJECTS

ChallengeConflicts, alternatives and uncertainty REVIEW

ApproveAuthorized promotion boundary HUMAN

PublishVersioned views plus release manifest OUTPUT

release-manifest.json
algorithm: sha256
files: 26
status: validated

26versioned package files

0runtime dependencies

8mandatory governance gates

9routing and boundary fixtures

The problem

The internet teaches the model. The chat forgets the lesson.

Search can find current information. Conversation can synthesize it. Neither alone creates durable, governed organizational knowledge.

01 / MEMORY DRIFT

Important details decay into vague preferences.

Chat memory improves continuity, but it is not a complete, inspectable knowledge ledger. Evidence, boundaries and superseded guidance can disappear from active context.

02 / RESEARCH SPRAWL

Reports accumulate faster than understanding.

Multiple research runs create overlapping conclusions, duplicate sources and conflicting recommendations. Future work must reconstruct which report is current.

03 / SILENT MUTATION

“Improve it” can overwrite accepted knowledge.

Without stable IDs, explicit deltas and approval boundaries, an AI can silently upgrade confidence, repair citations, or replace a decision without preserving history.

The architecture

Separate the learning system from the presentation.

The skill is intentionally thin. It routes the work, retrieves only relevant knowledge, applies governance, and generates bounded views from a versioned source of truth.

Evidence inputs

Public webCurrent external evidence

Connected sourcesApproved internal context

Prior decisionsExisting approved state

Regression fixturesKnown failures to protect

Governed core

Research StewardRouting and workflow

Knowledge objectsStable IDs and provenance

Promotion lifecycleProposal before approval

ValidationSchema, links and manifest

Derived outputs

Current-state guideNewest approved understanding

Decision briefImplications and next action

HTML and JSON-LDCommunication projections

Audit viewHistory, approval and hashes

Potential solutions

Choose the governance depth that matches the consequence.

Not every topic requires an enterprise research system. The skill defines a progression from lightweight continuity to durable institutional knowledge.

LEVEL 01 · LIGHTWEIGHT

Project context only

Use a dedicated ChatGPT Project, concise instructions, curated files and a current-state note.

Fast setup and strong conversational continuity.
Useful for low-risk personal research.
Weak version history, approval control and portability.

Best fit: personal exploration, low-risk topics, short-lived initiatives.

LEVEL 02 · CONTROLLED

Project plus canonical source

Keep the Project as the working set while a versioned document or repository remains authoritative.

Durable current-state guidance and explicit change history.
Supports human review and basic provenance.
Manual promotion and retrieval can still drift.

Best fit: teams, recurring research, design systems and architecture standards.

LEVEL 03 · GOVERNED

Steward skill plus knowledge bundle

Use the reusable skill to research, normalize, challenge, propose, validate and publish from canonical objects.

Stable IDs, explicit deltas, supersession and freshness.
Human approval boundary and deterministic release checks.
Evaluation proves triggering, retrieval and application.

Best fit: enterprise standards, AI platforms, regulated knowledge and reusable handbooks.

Operational workflow

A research flywheel with an explicit stop line.

The model can discover, synthesize and propose. It cannot silently approve its own conclusions as institutional truth.

Orient

Define purpose, audience, scope, freshness, source policy and completion criteria.

Retrieve

Load the smallest sufficient set of current claims, decisions, regressions and open challenges.

Research

Capture authoritative evidence with dates, exact locators, independence and limitations.

Normalize

Represent claims, findings, decisions and regressions as typed objects with stable IDs.

Compare

Produce NEW, CHANGED, CONFIRMED, DEPRECATED, CONFLICTING and UNCERTAIN deltas.

Challenge

Test alternatives, applicability boundaries and evidence quality before promotion.

Propose

Submit explicit knowledge changes. Stop before approval without authorized review.

Validate

Check object integrity, references, freshness, regression protection and communication quality.

Publish

Generate current-state, decision, evidence and audit views from the approved bundle.

Trust controls

“Looks good” is not a release gate.

Eight mandatory checks turn retention and governance into observable system behavior.

TriggeringActivates for durable research and stays out of unrelated work.

TraceabilityEvery material promoted claim retains provenance.

Epistemic clarityFact, inference, recommendation and decision stay distinct.

Explicit deltaChanges are visible instead of silently overwriting history.

Approval boundaryThe agent cannot self-approve proposed knowledge.

FreshnessReview dates and event triggers identify aging evidence.

Bounded retrievalOnly the smallest sufficient knowledge set enters context.

Regression protectionKnown protected behavior is checked before release.

Inside the package

Small enough to inspect. Structured enough to scale.

The skill uses progressive disclosure: a concise router at the top, focused reference modules behind it, typed templates, deterministic scripts and evaluation fixtures.

internet-research-steward / release

# validate knowledge objects
$ npm run validate
Validated 1 knowledge object(s).

# generate content-addressed manifest
$ npm run manifest
Wrote release-manifest.json with 26 file(s).

# package contract
runtime: Node.js 24+
dependencies: 0
algorithm: SHA-256
status: validated

Core routing

SKILL.md
README.md
package.json

Progressive references

01-research-contract.md
02-source-quality-policy.md
03-knowledge-model.md
04-promotion-workflow.md
05-freshness-policy.md
06-evaluation-contract.md
07-output-contract.md
08-current-state-template.md
09-project-setup.md

Typed knowledge templates

research-run.json
source.json
claim.json
decision.json
regression.json
change-set.json
handoff.json

Validation and evaluation

scripts/validate.mjs
scripts/build-manifest.mjs
evals/prompts.csv
evals/rubric.json
release-manifest.json

Release proof

Portable, inspectable and tamper-evident.

The package ships with a file-level manifest and an independently verifiable archive digest.

Release manifest

Every package file records its relative path, byte count and SHA-256 digest. A changed file produces a changed release manifest.

26 files · 22,690 bytes · release-manifest.json

Archive digest

Use the digest below to verify the downloadable ZIP after transport.

5f2f94c02f7bc8e7ef4db7d3f6d05c4f30021c04f0944b318b608388e81f4b81

Install the operating model

Stop hoping the research sticks.

Give the model a governed path from discovery to durable knowledge. Preserve evidence. Make change explicit. Prove what was retrieved and applied.

Download v1.0.0 Read SKILL.md