1. LandOffer
  2. Blog
  3. All articles
  4. Eraser Review for Engineering Docs: Diagrams, AI Drafts, and Repository Workflows

AI at Work

Eraser Review for Engineering Docs: Diagrams, AI Drafts, and Repository Workflows

Evaluate Eraser with a completed architecture correction and team plan. Separate repository context, Git sync, monitoring, editable source, sharing and metered AI usage before connecting work data.

10 min readLandOffer team

Eraser is worth evaluating when an engineering team needs diagrams and explanatory Markdown in one collaborative workspace, with a deliberate route into repository review. Treat AI output as a draft to compare with source evidence. Choose a plan around member roles, repository limits, editable artifacts and AI usage. A convincing picture is insufficient evidence of a correct architecture. Repository context, Git synchronization and automatic diagram monitoring are different capabilities.

This LandOffer.ai review checks current Eraser documentation and personally inspected public pricing on October 27, 2026. The architecture below is an authored synthetic declaration and deliberately flawed sketch. An actual local comparison checks their listed connections; it does not inspect a real repository or run Eraser. No account, repository authorization, data source, monitor, API call, sharing permission, purchase or Git operation was changed.

Engineers compare a rough track plan with a tangible switching-yard model and an overlooked holding siding.

What the combined document and canvas give you

Eraser's product introduction describes technical diagrams alongside Markdown documents, with AI generation and direct editing. A useful engineering document can therefore retain the reason for a connection beside its visual representation. Record the source, scope and owner beside the arrows so a later reviewer can challenge them.

Start by naming the artifact's job: current-system reference, proposed design, onboarding explanation or review of a change. A proposed queue is different from a deployed queue. A diagram based on a schema is not necessarily a picture of every running service. For a current-system reference, record the environment, source revision, included components and unanswered questions beside the diagram. A reader should know which revision a later change might invalidate.

The codebase-diagram guide describes generating from repository context and focusing a prompt on relevant paths. This can narrow the source material, but it does not prove that configuration is deployed or that every dependency appears in the selected files. Before relying on output, compare meaningful connections and omitted components with the evidence you actually supplied.

Worked correction: a plausible sketch with the wrong connections

The fictional Kestrel team wants a reference for a map-tile preview workflow. The authored source declaration contains seven named nodes and seven directed connections. It stipulates a browser reaching a gateway, then RenderAPI; the API writes metadata and submits work to WorkQueue; TileWorker consumes that work, writes rendered objects and updates metadata. This is an invented reference model, not code extracted from a working service.

The deliberately flawed author-written “AI-style” draft contains five nodes and five connections. It preserves Browser→Gateway, Gateway→RenderAPI and TileWorker→MetadataDB. It also draws Browser→MetadataDB and RenderAPI→TileWorker, while omitting WorkQueue and ObjectStore entirely. Eraser did not generate this draft; its errors are planted so the review can be inspected.

The completed change record is:

Sketch finding Evidence in the authored declaration Completed correction and limit
Browser connects directly to MetadataDB That connection is absent; the declaration routes browser requests through Gateway and RenderAPI. Remove the unsupported arrow. This is not proof that real access controls prevent direct database access.
RenderAPI connects directly to TileWorker The declaration inserts WorkQueue between them. Remove the direct arrow and add RenderAPI→WorkQueue→TileWorker. Queue delivery and retry behavior remain unspecified.
ObjectStore is missing TileWorker→ObjectStore is an explicit declared connection. Add the node and connection; no storage permissions or retention policy is inferred.
RenderAPI metadata connection is absent RenderAPI→MetadataDB is explicit in the source. Add that connection, retaining the worker's separate metadata update. No transaction or consistency guarantee is supplied.

The corrected reference flow, written as a compact text schematic rather than a rendered Eraser diagram, is:

Browser -> Gateway -> RenderAPI -> WorkQueue -> TileWorker -> ObjectStore
RenderAPI -> MetadataDB
TileWorker -> MetadataDB

We actually executed a local Python set comparison of the three authored JSON artifacts. The flawed sketch has two missing nodes, four missing declared connections and two unsupported connections. The corrected reference has exactly the seven declared connections, with no missing or extra pairs under that comparison. The saved source declaration, simulated draft, correction, script and result hashes make those bounded observations reviewable. No external dependency, API or real repository was involved in that local check.

This checks whether listed node pairs match the synthetic declaration. It does not test Eraser generation, source-code inference, diagram rendering, network traffic or runtime behavior. The declaration itself leaves deployment topology, authorization, retries, encryption, retention and performance unknown. Matching its arrows cannot answer those questions. The finished artifact is a corrected reference with visible limits, not an approved production architecture.

Verification is useful only against meaningful inputs

Eraser's Verify documentation describes checking a freeform diagram against attached models, presets, references, generation chats or a named standard. Each check uses one AI credit and reports issues without changing the diagram. If there is nothing to compare against, the product asks for context. That is an important boundary for an evaluation: a check needs an appropriate reference.

A reviewer should examine both the finding and the source's authority. A disagreement with an old architecture image may mean the image is stale. An apparent match with a configuration file may still leave deployment differences unresolved. Naming a security framework does not establish that the application was tested against it or that every relevant control is represented in the diagram.

For Kestrel, Bram owns the diagram record, Liora reviews API-side connections and Noor reviews worker-side connections. Their completed fictional disposition accepts the corrected seven-connection reference against the authored declaration only. Authentication, failure behavior and deployment remain explicitly unverified. Nobody authorizes a release or claims a successful Eraser Verify run.

A reviewer removes an unsupported bypass and checks the corrected route against the physical reference model.

Repository context, Git sync and Eraserbot are separate

The repository-connection guide distinguishes codebase context across GitHub, GitLab and Azure DevOps from monitoring and writeback. It currently supports Eraserbot monitors for GitHub and GitLab, and Git sync for GitHub. Azure DevOps context support does not establish Git sync or monitoring there. Repository authorization and active-versus-synced access also need an explicit decision.

For a GitHub document workflow, the Git sync guide describes pulling Markdown from main, editing in Eraser and pushing to a chosen branch, optionally creating a pull request. It warns that pulling overwrites note-editor content rather than resolving merge conflicts. Preserve unsynchronized work and compare upstream changes before a real pull; canvas content has a different boundary in that documented operation.

Artifact representation matters too. The guide distinguishes .eraserdiagram files, including JSON for freeform diagrams, from diagrams embedded as PNGs in Markdown under /.eraser. An embedded image is not automatically an editable DSL source file. Decide which artifact reviewers need to diff and preserve, then verify the actual saved files in an authorized test. For a team expecting source review, discovering an image-only artifact after adoption would change the decision. “Diagram as code” is not a guarantee that every export or embed carries the same source representation.

Eraserbot is a further workflow. It requires repository access and usage-based billing. The documentation describes monitored-file changes prompting diagram updates around pull requests, with creation attempts and subsequent API-based generation consuming different usage. A monitor cannot establish that reviewers accepted the generated architecture. No monitor, repository app or billing toggle was enabled here.

Current plans and a completed team selection

The personally inspected pricing page shows Free, Business, Business Plus and Enterprise. Business is $45 per member per month billed annually, or $60 monthly. Business Plus is $42 per member monthly equivalent, billed annually with a ten-seat minimum. Enterprise is quoted. These are displayed dollar rates, not a verified currency selection or checkout including tax.

Free currently lists three files, three AI credits and seven-day history. Business lists unlimited files, one Git repository, one data source and one hundred monthly AI credits. Business Plus lists three repositories, three data sources and two hundred fifty monthly credits. The current Plans guide specifies those paid credits per member. Starter is no longer offered to new teams; an older search result for Starter is not a current purchase offer.

Native pricing FAQs distinguish paid members with access to team files from free guests invited to particular files. Guests can edit and comment on those files; they are not necessarily read-only viewers. Private-file access is a separate decision. Unlimited guests does not make a continuing team-wide editor an equivalent free member.

Kestrel's completed selection record uses the same reference-document case:

Requirement Recorded decision Purchase or workflow consequence
Bram, Liora and Noor maintain the team's engineering documents Three ongoing members need collaborative editing. Evaluate three Business seats, not three guests substituted for team-wide access.
Product reviewer Nessa and designer Ivo comment on one approved document Their access is limited to that file. Candidate free guests with file-scoped invitations; no actual access granted or verified.
One GitHub repository holds the reference documentation One connected repository is required if integration is approved. Business's one-repository allowance fits the current requirement; no repository is connected in this review.
Retain editable source and reviewed changes A PNG alone is insufficient for the required source review. Inspect the selected diagram representation, saved source, image and branch diff before adopting writeback.
Private review and controlled links Sensitive draft material must not become broadly editable. Evaluate paid private files and link controls; do not use a reassuring file name as proof of restricted access.
Three repositories might be needed later This is a future expansion, not today's requirement. Re-price Business Plus or another arrangement when needed; the ten-seat minimum cannot be ignored.

At current displayed rates, three Business members imply $1,620 for a year or $180 for one month. Business Plus's ten-seat minimum implies $5,040 annually, not 3 × $42 × 12 = $1,512. These are planning calculations without tax, negotiated terms or usage charges. The finished proposal is a three-member Business evaluation with two file-scoped guest candidates. Nessa and Ivo have not accepted invitations; purchase and repository authorization remain pending.

Budget AI usage separately from member subscriptions

The usage-based guide lists additional AI credits at $60 per hundred, prompt-generation API calls at $80 per hundred and code-render API calls at $20 per hundred. Paid plans can enable this separately with a monthly spending limit; API calls require it. Included credits do not establish unlimited API generation or a fixed cost per finished diagram.

An original billing stress scenario assumes one Business user reaches 107 AI credits and also makes two prompt API calls and three code-render calls, with usage billing enabled. Seven extra credits imply $4.20; the two prompt calls imply $1.60; the three render calls imply $0.60. The total modeled additional charge is $6.40. These are arithmetic inputs, not executed calls, consumed credits or an observed invoice. Other members' usage and contract conditions are outside the scenario.

Set an approved budget before enabling automatic generation. Include retries, verification and monitors according to their actual metering rather than counting only diagrams eventually kept. Leave unused or unapproved integrations disconnected. That protects the evaluation's scope without inventing a promise that every change costs the same amount.

Sharing and export need their own acceptance check

Eraser's sharing documentation distinguishes team files, paid private files and link access. A team file can have broader link access, while a private file is limited to its creator and invited collaborators. Check the actual file type and link setting before sharing engineering context. No real recipient-view test was performed for Kestrel.

The detailed export guide identifies diagram and document outputs separately, including paid-plan restrictions for PDF and several editable export formats. The pricing comparison broadly marks file export included, so do not assume that every listed format is free. For this paid evaluation, inspect the required source, image or document output rather than promising a tested round trip.

Eraser merits evaluation when visual documentation and repository review are real requirements. The next step is a controlled evaluation of the completed correction and team-selection packet: verify the source, representation, intended audience, exact quote and branch behavior before authorizing connections or writeback. LandOffer.ai publishes this review; diagram accuracy, integration reliability and documentation outcomes remain unmeasured.

Sources and Further Reading