From scope, spec and past jobs to a draft schedule

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.

Nothing is uploaded, and nothing reaches the model that it did not ask for. 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, and beside it two things built from it: a count of which words each page uses, so search can rank, and a map of what the folder is made of — its sections, WBS bands, contract milestones, rate rows, the clauses that put a number of days on the work, and the trades they share. Both are built here by code, never by a model. The word counts stay on this machine and are never sent anywhere. The map is given to the assistant, because that is its whole purpose — it asks the map, through query_graph, which section or band concerns a trade before reading anything — but what it carries is counts and names, not the text of your documents: four files, six trades, which sections concern concrete. Forgetting the folder deletes both. Draft → Knowledge folder → What the assistant is told shows you the same map the assistant is given.

On top of both sits the wiki: pages the model writes from the files you tick, one per trade, WBS band, rule, rate table or contract frame. Every line cites the file and section it was read from, and when that file changes the line says so rather than going quietly wrong. You can add what the documents never said — that this client wants a joint survey before the pour — under a heading of your own, so a note is never read back as something a document said. The pages are markdown the assistant reads like any other file in the folder. They are written by the analysis model you chose, so a run costs one request per section; it says how many requests and roughly how many tokens before the first one goes out, and Stop ends it. Save them as a zip; drop the wiki/ folder beside your documents and they are read from there next time, on any machine.

What you can bring

Everything below is optional and stacks. Attach them on the Assistant’s Draft tab.

How a draft runs

A big programme is drafted as a recipe, not as rows. With “Draft as a recipe” on (the default), the model returns the dimensions and the step chains — units, disciplines, the document register and its revisions, packages with lead times, construction chains per discipline, systems — in one reply, and the engine multiplies them into thousands of activities with ids that cannot collide. The recipe appears as tables under the Draft button: change a count or a duration and the activity count follows; Build runs the engine and the gate without a model; Save keeps the recipe as your firm’s template; Learn reads one from the schedule you have open. The library offers a starter per industry and level (process plant, buildings, linear infrastructure at Level 2 and 3) whose durations are typical, not yours. Switch the recipe off for a free-form scope the model should draft row by row.

• 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

• 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.

No key? The built-in model

Settings › Provider › In this browser (no key), or the button under the chat box. A small model is downloaded once into the browser (about 1–2 GB, after you agree) and runs on your machine’s graphics card: nothing you ask it leaves the machine, and it works offline. It is smaller than Claude or GPT, so it is used for prose - explaining a date, the weekly narrative with /weekly, reviewing a returned progress sheet. Drafting and edits are refused with a pointer to Settings: they need a model that can be trusted to fill the schema. Chrome or Edge on a recent machine; it needs WebGPU.

Working as a team, and the weekly loop

• Tools › Sheet. A scoped progress sheet goes out as CSV, Excel, or one HTML file that works on a phone; what comes back reads into proposals you tick, applied as one undo. The assistant can review a sheet before you apply it.
• Reports › Weekly pack. The finish and why it moved, the driving path, a three-week look-ahead, the milestone tracker, and a narrative in which every sentence is computed. /weekly in chat writes it up for a client, numbers kept. Any report tab has Save as HTML: one file that travels.
• File › Team. A folder you already share holds the base programme and a journal per planner; Check in merges them and names what two people both changed. A live session puts two or more planners on one schedule at once over a direct connection; the signalling server carries only the handshake. Behind a strict firewall, a TURN relay (your firm’s own, or Aimavera’s where it is offered) forwards the connection, encrypted end to end, so it cannot read the schedule.
• File › Portfolio, and tabs. More than one schedule open at once; a multi-project XER opens as one tab per project; right-click a tab to compare against it; the open tabs or a whole folder roll up into one table.

4D

Opening a model. View › 4D, then Open model… for an IFC file (2x3, 4 or 4.3) or an .ifczip holding one. Add model… puts another beside it, such as the services next to the structure; each has a tick to show or hide it, and a hidden model keeps its links. A model you have opened before comes back from the geometry cache on this device instead of being read again. Drag to turn, shift-drag or right-drag to pan, scroll to zoom.

Links. Each take-off line, a storey or area by element class, is offered an activity, marked High, Medium or Low with the reasons. Accept all high-confidence takes the sure ones in one step; Next to review walks you through the rest. Type an Activity ID to change a line, or clear it. Links are held by Activity ID, so a P6 update keeps them.

Rules. Under Rules, Add rule links everything that matches a storey, class, system or words in the element name to one activity, and runs again on every open. When your corrections follow a pattern, the rule they imply is offered in a card under Links, with its own Add rule.

Picking in 3D. Click an element to select it, shift-click to add or remove, or use Box select and drag. With the view focused, the arrow keys step through the linked activities in start order and select each; Enter goes to Link to activity. With a selection, Link to activity links it to the Activity ID you type, and Create activity… opens a name, offered from its storey and class, and a number of days (5 to begin with), and makes the activity and links the selection to it. Hover an element for its activity, dates, float and what drives it. Escape clears the selection.

The re-link report. After an update from a newer XER, or a re-issued model, a card lists activities that have gone (with the one of the same name suggested), new elements and elements that left. Apply suggestions, Clear broken or Keep as is, and Undo re-link if it was wrong. Check links runs it at any time. Write to XER fields puts “4D elements” and “4D rule” on each linked activity, so the P6 file carries the summary.

Colour by. Build progress is the default: finished solid, under way amber, critical red, not started a ghost. Total float colours the model in four bands, from no float to more than 20 days, with finished work in grey. Select an element and choose Show driving path to light the chain that drives its activity; Step along path moves the date to each activity in turn. After a risk run, Build at P50 and Build at P80 (every activity at its own P50 or P80 dates) and Chance finished by this date appear in the same list. Crew on site (avg heads/day) colours each zone by its average heads a day over the Monday-to-Sunday week holding the slider date, against the busiest day on the whole schedule. It is offered once a labour resource is assigned to an activity. Stacked trades, two or more trades from different activities in one zone on the same day, are listed below the view as “week commencing” that Monday; click one to select the zone and frame it. Once a clash check has found something, Clashes joins the list and paints the open clashes.

Editing from the model. With a selection, Build after… adds a logic link (with two or more activities selected, the first-selected becomes a predecessor of the others; with one, you type the Activity ID it should build after) and Move to week… asks for the week starting date and sequences the work after the last unfinished activity that finishes before it (work on the selection’s storey first), or, with Use a constraint instead, holds it with a Start On or After constraint. Sequence bottom-up… chains one class of work storey by storey. Each opens a preview first, with the new finish, the DCMA score and the checks that fail after it, with any new failure marked, run on a copy; Apply makes it one undo step, Cancel leaves the schedule as it was.

Forecast and the baseline. The Data date mark on the slider is the data date; click it to jump there. Past it the date reads “forecast”, and work shown as done only because of forecast dates is tinted lighter. Once a baseline is set, Against the baseline colours work the baseline said would be finished, but is not. Compare with baseline shows the current schedule and the active baseline side by side, with one camera and one slider.

Clashes. Find clashes checks Hard, Clearance (a distance in mm, 50 by default, and unticked to begin with), Sequence and Workspace (stacked trades, which needs labour assigned), on this machine; Cancel stops a check. With Only when the work overlaps in time ticked (the default), hard and clearance clashes count only when the two activities’ dates overlap, a clash with one element whose work has no dates is kept, and a clash between two elements with no dates is left out and counted; untick the filter to list them. The list groups clashes by kind; Show frames the elements and ghosts the rest, and each has a status, Open, Accepted or Fixed, kept with the schedule. Run under a Show filter, the check covers what is shown and keeps the clashes found elsewhere. A sequence clash, work built before what holds it up, offers Fix with a link, previewed like any other edit. Export CSV writes every clash; Export BCF writes a BCF 2.1 file (.bcfzip) for other BIM tools, one topic for each clash between two elements, so Workspace clashes are in the CSV only.

Showing less. Show narrows the view to the selected activities, a WBS branch, a storey, a system or a week, and the rest is left out of the picture; This week takes the Monday-to-Sunday week holding the date on the slider, and keeps it when the slider moves.

Getting it out. Export PNG saves the view at the slider date. Record video plays the whole run into a video file (WebM, or MP4 where that is all the browser records; keep the tab in front while the video records; Stop or Esc keeps what is done in either). Weekly frames saves a PNG for each Monday of the run, plus the first and last day, as a zip. Both have the date burned in on an opaque background. File › Export IFC writes IFC 4.3 with the links, each element named by its GlobalId; the XER stays the file of record.

Live sessions. Link and rule edits made in 4D reach everyone in a live session. Each planner opens the IFC on their own machine; only the links travel. Large frames go in parts, so a joiner on the current version gets a schedule of any ordinary size, and the links can hold over 200,000 element links with ordinary Activity IDs; past that the Studio says so and each planner keeps their own. Rules and line links are far smaller.

Size, and simplify first. We have measured a 90,000-element, million-triangle model on an Apple M2. Larger files depend on the machine’s memory, and a multi-GB federation is still workstation work: split it by discipline or storey and open the parts you are sequencing. There is no photoreal rendering.

What stays on the machine. The model, the clash results and the geometry cache stay on this machine; BCF files, video and frames are written in this browser and saved to your device. The IFC is read here and never uploaded, and no AI provider is involved. In a live session the schedule and the 4D links travel between the planners’ browsers; the model and the clash list do not.

Work packages

Setting up. Packages › Packages…, then Set up work packages. It adds three activity codes, Work package, Discipline and System, with a starting discipline list for oil and gas, mining, or heavy industrial and civil work, and changes nothing else. Settings holds the package sizes, the look-ahead and backlog weeks, the lead days for each kind of constraint, and the names of the package types; Light mode for small projects offers only areas, construction packages and installation packages.

Packages. A package is a value of the Work package code, so P6 reads it: work areas (CWA) at the top, construction packages (CWP) in an area, and installation, engineering, procurement and fabrication packages (IWP, EWP, PWP, FWP) in a construction package. New adds a package under the one selected; Assign selected codes the activities selected in the table. An activity sits in one package, and its area and construction package follow from the tree. Show activities filters the table to a package and the packages inside it. What P6 has no field for, such as state, crew and hours, travels in Studio tables that P6 ignores, so the XER round trip stays lossless. The checks warn about size, empty packages, mixed disciplines and activities left on a construction package; they never stop you.

Lists. Import list… reads a CSV or Excel list of components, drawings or material lines, matches its columns for you to confirm, and adds them as items. A tag already in the project is updated, not duplicated, and a package column links items to packages by code. Material lines that are not yet fully received can become constraints.

Rules. Rules… codes activities to packages in bulk by WBS, name or code, such as everything whose name contains PIPE. The first rule that matches an activity wins, a WBS rule includes the bands inside it, and the count updates as you type. Applying is one undo step, and the rules are kept to run again after an update. Propose with AI…, shown once an AI provider is set up, sends the activity list (the first 1,500 activities) to that provider, through the relay described under Where your files go, and offers areas, construction packages and installation packages to tick; nothing is applied until you accept. Mirror to WBS… makes WBS bands for the areas, or the areas and construction packages, and moves their activities in, as one undo step.

Constraints. Packages › Constraints…. A constraint is what must be cleared before an installation package can go to the field. Each is a P6 issue named by its category, tied to its package, with an owner and a need-by date. Without a date of your own, the need-by date is worked out from the package’s start less the lead for its category, and the word beside it says which. A constraint is late when it is open and past its need-by date or forecast after it. Close what is met closes those whose cause is met: every unit on a material line received, or the package it waited on complete. Meeting pack PDF puts the open constraints, the look-ahead and the backlog on paper.

Release. A package cannot be marked constraint-free while any of its constraints is open. Release to field with constraints still open asks why, and the reason is kept in the package’s history.

Look-ahead. Packages › Look-ahead… shows the installation packages starting in the coming weeks, one column per week, 2 to 12 weeks ahead. Ready means no open constraints, at risk means some are open, blocked means one is late. The backlog is the hours of released or ready work with nothing open against it, divided by the hours planned to start each week, against the target in Settings.

Field pack. Packages › Field pack… works on one installation package. Each activity’s steps have a weight, hours, quantity and unit; a hold or witness point earns nothing until it is signed off, when Settings say so, and Sign off records who and the date. Rules of credit… keeps a set of steps for a kind of work. Field pack PDF prints the steps, materials, drawings, constraints, safety items and a sign-off block.

The field form. Field form makes one HTML file for the foreman’s phone, with nothing to install and no network needed. The foreman ticks steps, enters quantities and QC sign-off, and downloads a CSV; what is typed stays on that phone until the CSV is sent back. Read a returned form… shows one line per activity to tick, lists anything it could not match, and applies the ticked changes as one undo step.

Path of Construction. Packages › Path of Construction… puts the areas in the order the site is built; each move is one undo step, and the checks list where the dates or the logic do not hold the order. A plot plan is optional: a PNG, JPEG or WebP picture of the site with a numbered pin for each area, reduced to 1,600 pixels and refused if it is still over about 2 MB. Issue records the order; Changed since issue lists what has moved, been added or been removed. Need dates work engineering, procurement and fabrication back from the construction package each serves, with the slack in calendar days. The ideas register keeps constructability ideas, and only ideas built into the plan count towards the hours saved.

Systems. Packages › Systems… holds the systems and subsystems that are commissioned as one, each with its open ITRs, open punch items and mechanical completion date; a system is ready for turnover when all its activities are complete and both counts are 0. Make turnover activities adds mechanical completion, pre-commissioning and handover for each ticked system, linked in that order. Add test pack holds makes insulation and heat tracing wait for the package that tests the lines, and Add pre-commissioning steps adds a step list for Field pack. For brownfield work, set the shutdown and tick the live areas: the findings list packages overlapping in a live area, packages outside the window of their isolation or permit, and packages needing an isolation outside the shutdown.

Measures. Packages › Measures… works out the backlog, released ahead of need, engineering and procurement on time, RFIs that led to a change, estimates changed after release, the schedule factor, rework and productivity, each with the count behind it. The S-curve and the skyline show the planned hours, and the self-assessment scores five areas, one saved entry per person, with a trend. Measures PDF prints it all, and Report Studio has the same figures as blocks. In 4D, Colour by offers By area, By construction package, By IWP state and By system, and Make an IWP… makes an installation package from a selection.

What never leaves the machine. The schedule, the lists, the plot plan and the packs are read and made in this browser and never uploaded. The lists stay in the saved schedule; the plot plan, the shutdown window and the self-assessment stay in the saved file and are not written to the XER. The one thing that can go out is your choice: Propose with AI sends the activity list, when you press it, to the model provider you chose, through the relay at aimavera.com/api/ai, which passes it on and keeps nothing. A model running on your own machine is called directly from the page, and nothing leaves the device.

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, the word counts that let search rank it, 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.