Who does what
Two parties build a draft. The model (Claude, GPT, Gemini, or a local runtime you choose) reads and proposes. The engine (the CPM scheduler running in your browser) never guesses: it parses your files, judges every proposed edit, and reschedules.
| The model proposes | The engine judges |
|---|---|
| Reads the scope, the ticked spec sections, the worked example, the folder | Parses PDF, Word, PowerPoint, Excel, XER, XML and IFC in the browser |
| Decides which folder files to open — up to 12 reads per draft | Checks every edit: shape, both ends exist, no cycles, no actuals |
| Proposes activities, WBS bands, durations and links as edits | Runs CPM, DCMA 14-point and the milestone gate on a copy first |
| Judges which logic findings matter, and ranks them | Finds open ends, leads, constraints and cycles itself, for free |
Nothing is “ingested”. There is no index of your folder sitting anywhere. The model asks for a file, a section or a page; the engine reads it here; only that passage goes to the model. What was read is kept on this machine, so the next session does not read it again.
What you can bring
Everything below is optional and stacks. Attach them in the Assistant’s Guided tasks view.
| Input | Formats | What reaches the model |
|---|---|---|
| Scope | typed text | As written. A tender description, a list of what has to be built, or one line: “steel frame, sequence it like Substation-42”. |
| Specification | .pdf .docx .md .txt | Two requests. First the CSI section index only — the model says which sections a planner needs; you tick or untick. Then the ticked sections: PART 1 and PART 3 whole, PART 2 products cut to a head, up to 150,000 characters. Save as .md to reopen instantly. |
| Previous schedule | .xer .xml | A worked example: the WBS tree, activities as code | name | wbs | days, logic as A1020 -> A1030 FS lag. Dates, float, status and ids are stripped — habits, not content. Whole up to 1,500 activities; above that, tick the WBS bands the new scope shares. |
| Model | .ifc .ifczip .csv .xlsx | A takeoff: one scope line per breakdown and class with its quantity. No ids, no geometry. Read in byte ranges by workers, never held whole. |
| ITT | .pdf .docx .md .txt | Dated contract milestones — the frame the draft must land inside — and the client’s mandated top-level WBS. Both editable before sending. |
| Knowledge folder | a folder | Read on request through five tools: list_files, outline_file, read_file, search_files, view_page. Put standards, method statements, rate tables, past programmes and sequence diagrams here. Chrome and Edge remember the folder; other browsers ask each session. |
How a draft runs
- You attach and tick. ITT first, then the past job (tick its bands if it is big), then the spec sections, then the folder. Type the scope last, naming the files you want followed.
- Draft activities. One request goes out with the rules, a summary of the schedule you already have, and everything ticked. With a folder on, the model may read first: the panel lists each read as it happens, with a Stop link.
- The model proposes edits — activities, WBS bands, durations, links — as the same operations the undo stack uses, plus a summary that names the file each duration or sequence came from.
- The engine validates every edit. Wrong shape, a link to an activity that does not exist, a cycle, an attempt to set actuals: refused, and the model is told why once. You see “38 edits, 2 refused”.
- The gate runs on a copy. CPM, DCMA 14-point and the contract milestones, before anything touches your schedule. A draft that misses the completion milestone says so here.
- Apply or discard. Applied edits land as one change, one undo away. “Built from: spec · folder · template” stays on the result, so an estimate is never mistaken for a grounded one.
- Review the logic. The engine finds open ends, leads, negative lags, hard constraints and cycles itself; the model only ranks which matter. Then “Change it in words”: “push procurement out two weeks”.
How each file is read
PDF. The text layer, page by page, in the browser. Specifications are indexed by CSI MasterFormat section, so a 2,000-page book is read by section rather than by page count. A flow chart in a PDF comes through as its words only — the arrows are not in the text layer. For that, the model can call view_page, which draws the page as a picture for a vision model to read. A scanned PDF has no text layer at all; view_page is the only route.
PowerPoint. Slides are read as text with their logic reconstructed: each box’s text, its swimlane (from the vertical labels down the left edge), arrows as flow: Excavate -> Blind -> Pour whether snapped to the boxes or drawn loose over them, and tables row by row. A diagram pasted in as a picture is a picture: only view_page sees it.
Previous schedules (.xer, P6 .xml, MS Project .xml). Three readings of one file. In chat, a six-line summary for questions. As a template, the worked example above. In the folder, outline_file lists the WBS bands with activity and link counts, and read_file section=STEEL returns one band whole, with links into and out of it marked (outside band). A band bigger than one read names its sub-bands rather than clipping.
Word, Excel, CSV, text. Paragraphs, sheets as CSV, plain text. Rate tables in Excel read well; put units in the header row.
Images. Drawings and photos go to a vision model as pictures, downscaled to 1,600 px on the long edge. In the folder they are reached with view_page.
Limits
| Where | Ceiling | Past it |
|---|---|---|
| One-shot template | 1,500 activities · 3,000 links | The rest is cut top-down. The band list turns red; tick bands instead. |
| Spec sections sent | 150,000 characters | Later sections are left out; the counter says so before you send. |
| Folder reads per draft | 12 | The draft stops and says to narrow the scope or pause the folder. |
| One folder read | 12,000 characters | The read says where to continue, or which sub-bands to read. |
| Folder size | 2,000 files | The walker stops listing; use subfolders. |
| Search | first 50 files | The first search reads every file it can — minutes on a big folder of PDFs, once. |
| Page as picture | 1,600 px · about 1–2k tokens each | Only the newest picture stays in the conversation; older ones become captions. |
- The model transfers logic; nobody copies it. “Sequence like Substation-42” is pattern transfer by the model: it recognises the same trades and mirrors the links. There is no deterministic name-matching. A frontier model does this well; a 3B local model does not.
- Local models and tools. Small local runtimes are poor at tool use and cannot see pictures. With them, tick bands and sections yourself rather than relying on the folder, and expect drafts without
view_page. - Templates carry habits, not calendars or resources. Naming, sizing, WBS shape and logic go; calendars, resources and costs do not.
- One template at a time. Two past jobs: put both in the folder and name them in the scope.
- PowerPoint heuristics. Lanes are the vertical labels at the left edge; a loose arrow is matched to the boxes its ends touch; slide order is file order. A diagram that spans slides is two diagrams.
- The folder is a Chrome and Edge feature. Safari and Firefox can attach files one at a time in chat instead. What was read is cached per browser, per machine.
Getting a good draft
- Frame first. Attach the ITT before anything else. The milestones bound the draft and the client WBS becomes the top level, so the model is not inventing a breakdown you will have to rename.
- Pick the band, not the job. For a past programme over a few hundred activities, tick only the WBS bands the new scope shares. Four relevant bands beat the first 1,500 activities of a 9,000-activity job every time.
- Let the spec index choose, then prune. Accept the model’s section picks, then untick product-data sections. What carries work and sequence is PART 3 EXECUTION.
- Name the files in the scope. “Follow Substation-42’s STEEL band. Strip and cure per sequencing-rules.pdf section 4. Steel lead time from rates.xlsx.” A named file is read; an unnamed one may not be.
- Draft in bands. For an EPC job, draft one WBS band at a time. Each later draft extends the schedule you have and is told not to duplicate it. A 300-activity draft is reviewable; a 3,000-activity one is not.
- Use the right model for the job. Set a frontier model as the analysis model in Settings: drafts and the spec index use it, chat can stay on something cheaper or local. Drafting is where model quality shows.
- Read the gate, not the summary. The gate verdict is the engine’s; the summary is the model’s. If DCMA shows 40% missing logic or the finish milestone is missed, discard and narrow rather than apply and fix.
- Then review the logic. Open ends and leads the engine found are facts; the model’s ranking is opinion. Fix the facts first.
- Keep what you read. Save the converted spec as .md; keep the folder connected in Chrome or Edge. Second sessions cost seconds, not minutes.
A good scope line: “Substation building: civil substructure, steel frame and envelope. Client WBS from the ITT. Sequence civil like Substation-42 band SUB, steel like band STEEL. Formwork strip and cure per standards/sequencing-rules.pdf §4. Procurement lead times from rates.xlsx. Level 3, metric.”
What produces poor drafts: a whole 9,000-activity XER as the template; every spec section ticked; a folder of a hundred files and a scope that names none of them; a local 3B model asked to use tools; applying a draft that failed the milestone gate and fixing it by hand.
Where your files go
Files are parsed in the browser. Nothing is uploaded to Aimavera. What goes to the model you chose is only the text you ticked or the passages it asked for, through a relay that stores and logs nothing — or, with a local runtime, directly from this page and nowhere else. Extracted text and the folder handle live in this browser’s storage on this machine; forgetting the folder clears them. The folder itself is opened read-only; nothing is ever written into it.