agents

Plan Set Comparison now opens the revised drawings with changes marked directly on the PDF. Revision clouds and numbered callouts put each change in context, with a searchable change list alongside the plans. Select a change to jump to its location and compare close-ups of the original and revised detail.

Use Download compiled PDF to take the marked-up set out of Specset. The export includes your annotation edits and the changes you choose to include, with an optional change-log cover that links to the revised sheets. Export the full comparison or select specific sheets, ready to share with your team.

Comparisons are also more accurate when drawings have been reorganized. On redrawn sheets, the agent compares individual details, schedules, and notes to separate content changes from layout shifts. It recognizes renumbered sheets and follows details that move to another sheet, reporting the move and any edits without counting the same detail as both removed and newly added.

See Plan Comparison for the review and export workflow.

Skip the keyboard when speaking is faster. The new waveform button lets you dictate directly into a Specset composer.

Click Dictate, speak while the timer runs, then stop the recording. Specset transcribes your words and adds them to the text already in the composer. You can review, edit, and add context before anything is sent or filed.

Dictation is available in every project conversation and in the Daily Log quick entry where that tool is enabled. Your microphone is only active while the red recording timer is visible.

See Ask the project for the complete workflow and microphone-permission guidance.

Submittal reviews check a package against hundreds of spec requirements, and a review that casts a net that wide will inevitably flag some things that turn out to be fine. Until now, sorting those out was your job. Every issue a review files is now re-checked by a second, independent AI before it enters your issue list.

The verifier goes back to the package and actively tries to disprove the finding. It searches the attachments for content the review called missing, checks the submittal register for a separate package that covers the deliverable, looks for an RFI or addendum that supersedes the requirement, and reads the governing spec section for or-equal language that already permits the submitted product.

Findings that do not survive are dismissed automatically with the reason recorded, so they never cost you a click. Open a dismissed issue and hover its AI rejected badge to read exactly what disproved the finding, with citations that jump straight to the submittal page or drawing schedule the verifier used. Nothing is hidden: rejected issues stay in the Dismissed list, and one click re-opens any of them if you disagree. Issues that survive carry an AI confirmed verdict on the issue itself, with the verification evidence a hover away.

Read more in Submittal review.

Specset can now take you straight into the right project chat when you open the installed app on site. Instead of finding the project and starting from scratch, you arrive with that project's conversation context already loaded.

Project managers set the site's coordinates and matching radius in Project Settings > General. Each user chooses whether to enable the feature in the installed app. When one project matches, its chat opens automatically. When sites overlap, Specset asks which nearby project you want.

The live device location is used only to perform the match locally. It is not stored or sent to Specset, and nearby-project detection can be turned off at any time from the account menu.

See Install the app to configure a project site and enable nearby-project detection.

Design quality reviews cast a wide net on purpose, and until now that meant you did the sifting. Every issue a review files is now checked by a second, independent AI before it enters your list. The verifier goes back to the plans and specs and actively tries to disprove the finding: it re-reads the cited sheets, checks drafting conventions, and looks for the answer elsewhere in the documents.

Issues that hold up show an AI confirmed badge. Hover it to read what was checked and why the issue stands, with a link to the full verification session. Findings that do not survive are archived automatically with the reasoning saved, so they never cost you a click. The verifier also gives severity a second opinion and rewrites vague titles, so what you see is worth your time.

Real-but-minor items, like typos, leftover boilerplate, and details a later revision will fill in, are routed to a separate Cleanup list instead of your review queue. The queue keeps only findings that need someone on the design team to make a call.

Read more in Issues.

Review tables, issue lists, extractions, drafted documents: everything the agent produces is an output, and every output now lives in the Library. It ships to every workspace today, at both the project and organization level.

• One place for everything the agent makes. The Library lists every output with its type, status, and the session that produced it. Search and filter to find a deliverable weeks later, or start a new output right from the page.
• Open outputs in place. Click a row and the output opens in a side panel, so you can drill into individual checks and issues without leaving the list, or jump back into the chat that produced it.
• Chat history knows what was produced. A new "Has output" filter and per-conversation output counts make it easy to find the conversation behind a deliverable instead of scrolling by title alone.
• Checklist skills can file issues. Turn on issue creation for a checklist skill and every flagged check becomes a draft project issue, cited and ready to triage.

The Library takes over from the Agent Sessions page as the home for agent outputs. Workspaces still using Agent Sessions keep it while they transition.

Read more in Skills.

Specset is now available as a native plugin for both Claude and Codex. Install one bundle to give either agent the project context and construction workflows it needs.

• Nine shared skills. Search specs and drawings, set up projects, manage submittals and RFIs, work with closeout records, ask project agents, and administer your organization.
• Native MCP tools. Search and read structured Specset content directly, with the same tool registry on both agents.
• Secure sign-in. Local plugins use browser or device authorization through the Specset CLI. Hosted ChatGPT connections use organization-scoped OAuth.
• File transfers. Claude and Codex can stream local uploads and downloads through the CLI. Hosted clients can pass attachments and receive short-lived download links through MCP.

The plugin defaults to read-only work and asks before changing project data. Credentials stay in the local CLI or the hosted OAuth connection rather than in the conversation.

Follow the setup guides for Claude and Codex.

Tell the agent something once and it sticks. Memory carries your preferences, decisions, and working conventions across conversations, so every new chat starts already knowing how your team works.

• Saved as you chat. State a preference ("always quote fire-rating requirements verbatim") and the agent saves it as a memory, marked with a chip in the conversation so nothing is stored silently.
• Applied everywhere. Memories load into every future conversation, org-wide or scoped to a single project.
• Yours to manage. Review, edit, and delete memories under Org Settings or Project Settings, including everything the agent saved on its own.

Read more in Ask the project.

Every team has review checklists, QA workflows, and extraction jobs that are theirs alone. Skills let you teach them to the agent once, then run them anywhere with a /.

• Create by describing. Tell the Skill Creator what you want in plain language, from the skills settings page or straight from the skill picker in chat, and it authors the skill with you. Prefer full control? Build it manually in the editor.
• Checklists and review tables. Define your own categories, checks, and custom columns, and the agent fills the table with cited findings each time the skill runs.
• Project-wide extractions. Extraction skills sweep every document, spec section, drawing sheet, or RFI in scope and build a results table row by row, with a citation on every value.
• Run it anywhere. Type / in any conversation to launch a skill, org-wide or scoped to a project.

Read more in Skills.

Your deliverables already have a format. The agent now drafts into it instead of starting from a blank page.

• Word templates. Attach a .docx and the agent maps its structure, then fills each section with content grounded in your project record. Every fill is backed by citations, and sections that need information only you have are flagged for input rather than invented.
• Excel templates. Attach a .xlsx form or log and the agent fills the fields and table rows while your formulas, styles, dropdowns, and images stay byte-for-byte intact. The result opens in Excel like you filled it in yourself.
• Review before it ships. Drafts open in the paginated document editor for a final pass, then export to DOCX or spreadsheet.

Templates can be attached ad hoc in a chat or registered on a skill so the whole team drafts the same way. Read more in Skills.

AI review findings are only worth your time if you can trust them. Issues now get a built-in second opinion: an independent verification agent that re-researches each finding from scratch, with no stake in the original conclusion.

• Confirm or reject, with receipts. The verifier re-reads the cited documents and either confirms the issue, rejects it, or refines it. Its full reasoning is available as a dedicated verification thread on the issue.
• Severity gets a sanity check. Overstated or understated severities are corrected as part of verification, so the high-severity list stays trustworthy.
• Automatic on new reviews. Projects can turn on auto-verification so design quality, deep design, and drawing review issues are verified as they're created.
• Filter by verdict. Verification status appears on every issue, and the list and Overview dashboard filter by it, so you can start with confirmed issues and work outward.

Read more in Issues.

Design Quality Review works spec-first: it walks your specification sections and checks the design against them. The new Drawing Review agent works from the other direction, leading with the sheets themselves.

• Every sheet gets its own reviewer. The agent fans out across the drawing set in parallel, one review per sheet.
• Cross-discipline checks. Each sheet is compared against the related sheets it coordinates with, across disciplines: structural against architectural, mechanical against plumbing, plans against their schedules and details.
• Coordination defects surface as issues. Clashes, conflicting values, and broken callouts or cross-references land in your Issues list with citations to the exact sheet regions involved.

Run it alongside Design Quality Review rather than instead of it. The two catch different classes of problems. Read more in Design review.