Planning: deliverables, work, dependencies, schedule and resources · Agent Skill

Every project has work to choose and records to keep, so read at least §Choosing methods by need and §Work records in the project’s tool on every project. Read the rest when there are dependencies, dates, shared resources, risks, or changes to agreed scope. These are established project-management methods (ISO 21502 and the APM body of knowledge describe them); what follows is enough to apply them without either. They are chosen by need, not all at once, and none of them is a separate document for its own sake.

Choosing methods by need

Mechanism	What it supplies	Use it when
Deliverable list; product and work breakdown	What is produced, and the work each needs	Always a list; a hierarchy when there are several deliverables, teams or substantial scope
Work packages and tasks	Assignable units with acceptance criteria	Always; detailed enough to choose and verify the next action, no further
Dependency network	What is blocked by which prerequisite	Whenever one piece of work needs another first, or an outside party must hand something over
Milestones and a Gantt schedule	Dates, durations, parallel work, handoffs	When deadlines, lead times, availability or coordination materially affect delivery
Critical path and resource analysis	Which delays move the finish; whether the plan is feasible	For connected schedules; include calendars, unavailable equipment and shared people
Task board; Kanban when appropriate	Work in progress, waiting and done	For active flow; real Kanban also has an explicit workflow and work-in-progress limits
Responsibility matrix (RACI)	Who performs, decides, is consulted or informed	When roles or handoffs become ambiguous; otherwise a named owner and acceptor suffice
Risks, assumptions, issues, dependencies (RAID)	Uncertainty, obstacles and responses	Consider material items on every project; one consolidated record, not empty logs
Change control	Effects on scope, resources, dates, quality, acceptance	Whenever something agreed would change
Rolling-wave planning	Near-term detail, honest distant estimates	When uncertainty makes a detailed long-term schedule unreliable

These are different relationships: a breakdown says what work belongs to a deliverable; a dependency says what must happen before other work can proceed; a Gantt chart shows time; a board shows workflow state; an architecture diagram describes the product. None substitutes for another.

Deliverables and work breakdown

List the deliverables (the product breakdown): every output the project produces or hands over, including the ones that are not the main product: documentation, training, certification files, packaging, a handover pack, an archive.
For each deliverable, list the work it needs (the work breakdown): design, build or make, procure, inspect, test, install, rehearse, review, accept.
The breakdown covers everything and nothing twice: all the work, at every level, is the sum of its children; work that belongs to no deliverable is a question (it serves nothing, or a deliverable is missing).
Break down only as far as needed: a leaf is a work package when one owner can do it, its acceptance is checkable, and it fits within one reporting period. Do not atomise into keystrokes.

Work packages and acceptance criteria

Each work package records, in the project’s tool (§Work records in the project’s tool):

the deliverable it belongs to;
the responsible actor, and who accepts it;
acceptance criteria: concrete, checkable conditions on the actual result, agreed before work starts;
the procedure that will produce and verify it (../SKILL.md §4. Procedures and evidence);
its prerequisites and the resources it needs;
an estimate, and dates where the schedule needs them;
its state: not ready, ready, in progress, waiting (on what), done (with a link to evidence).

Milestones as usable slices

Vertical slices, not layers. “1: backend. 2: frontend. 3: hardware. 4: integration.” leaves nothing anyone can use until the last layer, so the first setback leaves nothing. Each milestone leaves something a person can use, and names the measures it moves.
A walking skeleton before any breadth: the thinnest version of the first journey, end to end through every stage (built from clean, a test seen to fail, released or deployed or installed, observed in use, and rolled back once). For a physical product, one real reading or action on the bench, confirmed by an independent measurement (physical.md §Commands and observed outcomes). Create only the folders, services, parts and dependencies the skeleton uses; an empty folder for a service that may exist one day is a decision taken without evidence.
The first release to real people is an early milestone, with its route to them ready.
Each milestone ends with a review before the next starts (quality.md §Reviewing work; for anything already running, operations.md §Periodic review).

Dependencies and the schedule

Across projects, link only real prerequisites and shared-resource constraints in their existing records. Each project retains its own objective, authority and acceptance; being in one workspace does not make all adjacent or parked work commissioned (../SKILL.md §1. Classify the project).

Record each real prerequisite as a link between work packages (“B cannot start until A is done”; occasionally “cannot finish until”, or a lag such as “concrete cures for 7 days”). Record outside handoffs (a supplier delivery, an approval, a client’s content) as work packages or milestones owned by that party, with the date they committed to.
Estimate each duration in working time, and note the calendar it runs on (a person’s working days, a supplier’s lead time in calendar days, a lab’s opening hours).
Draw the network or the Gantt chart when it makes coordination clearer; the records hold the links and dates, the chart is a view of them.
Keep the schedule honest: a date is either committed by the party who controls it, or an estimate marked as such. What may stand as that source, and what may not, is §Time limits.

Critical path and resources

For a connected schedule, work out which delays move the finish:

Forward pass: each package’s earliest start is the latest earliest-finish of its prerequisites; earliest finish = earliest start + duration.
Backward pass: from the finish date, each package’s latest finish is the earliest latest-start of the work that depends on it; latest start = latest finish − duration.
Float = latest start − earliest start. Packages with zero float form the critical path: any delay to them delays the finish. Near-zero float is nearly critical.
Check resources, not just arrows. For each person, machine, room or budget, add up the demand per period against availability (holidays, other projects, a machine booked elsewhere). Where demand exceeds it, move work within its float first; if that is not enough, the finish moves, and the plan says so.
When anything changes (a delivery slips, a person is unavailable, an estimate was wrong), recompute, update the forecast, and tell the owner which milestones and acceptances move. Pick the next ready work from what the new schedule allows; do not mark delayed work done or rebuild unrelated work to look busy.

A timeline view in a tool is not a computed critical path or a resource-feasible schedule. If the project needs these and its tool cannot provide them, choose an existing scheduling tool on that requirement (decisions.md); do not build one.

Flow: boards and Kanban

A board is a view of the work records’ state (not ready, ready, in progress, waiting, in review, done). It is not a second plan.
Real Kanban adds explicit rules: the columns and what moves an item between them, a limit on work in progress per column, and attention to items that wait. Adopt them when flow matters (continuous work, support, operations), not for a one-off build.
Limit work in progress even without Kanban: finishing one thing beats starting three.

Responsibilities

Every deliverable and work package has one responsible actor and one acceptor. For a small project that is enough.
Use a responsibility matrix when roles or handoffs become ambiguous: for each deliverable or decision, who is Responsible (does it), Accountable (one person who decides and accepts), Consulted (asked before), Informed (told after). One Accountable per row.
The operating model records, for each procedure, who performs it today (a person, an agent or automation), who approves it, and the fallback (../SKILL.md §4. Procedures and evidence). A procedure only one person or one agent can perform is one absence from stalling.

Risks, assumptions, issues and dependencies

Keep one consolidated record of the material items, not four empty logs:
- a risk is something that may happen, with its likelihood, its impact, the response (avoid, reduce, transfer, accept) and an owner;
- an assumption is something taken as true without evidence, with how and when it will be checked;
- an issue is something happening now that needs action, with an owner and a date;
- a dependency on something outside the project, with who controls it.
Consider them on every project; revisit them at each review and whenever the evidence changes. A risk that happened becomes an issue; an assumption that failed changes the plan.

Change control

Anything that changes agreed scope, resources, dates, quality or acceptance is a change request, however small it looks and whoever asks for it.
Assess it: what it affects (deliverables, schedule, cost, risks, acceptance), and the options, including not making it.
Decide it within the delegated authority set in the briefing, and record the decision and its effects; outside that authority, take it to the person who can decide, with the assessment, and do not act on it until decided.
Update the records (work packages, schedule, acceptance criteria, risks) so the plan and the work agree.

Rolling-wave planning and exploration

Plan the near term in detail and the distant work honestly: the next milestone as work packages with estimates; later ones as deliverables with ranges. Detail the next wave as the current one nears its end.
Where nothing is known (an invention, a system nobody has built), write the unknowns as questions; for each, the cheapest experiment, simulation or prototype that would answer it; run the ones that decide the most first; record every result, failures included, as evidence. Search prior work first, so the experiment starts where others stopped.

Resources and their sources

Choose every resource’s source; do not take the first. Accounts, keys, models and the agent’s token budget, hosting, domains, parts and test devices, suppliers, venues, people and hours, money. For each, compare at least two sources, including what the owner already has and building or making it, on cost now and at the target size, lock-in, lead time, and who has to act (decisions.md).
Record, for each resource, the decision that chose its source, or that the owner already has it.
Send the owner one message with the whole resource table and a recommendation; never nag toward a vendor. A resource the owner cannot provide is a constraint to route around: a free tier, a grant, a cheaper part, a borrowed tool, something built.
Set a budget for agent and model use, and record spend against it.
At each review compare spend with budget for every line (hosting, services, parts, contractors, agent and model use), note free tiers that ended and prices that moved, work out the runway, and route around what changed: another source, a cheaper design, a funding route. Money is a constraint to engineer around, never the reason a project stops.

Estimates and arithmetic

Show every calculation with its units and inputs, so another person can redo it.
Recompute every number you rely on, including a supplier’s or vendor’s, which is a claim until a second source or a measurement confirms it.
Where a calculation would rest on a guess, simulate or prototype.
Name the assumptions that would change the answer, and by how much they would have to move.
Estimate durations as ranges where uncertain (an optimistic, a likely and a pessimistic value; a common point estimate is (optimistic + 4 × likely + pessimistic) ÷ 6), and say which estimates are guesses.

Failures this prevents: a battery that lasts two days where the idea promised a season; a free tier that becomes a four-figure bill at the target size; a bill of materials above the price.

Time limits

A time limit is a check, in the same way a budget is (quality.md §Budgets as checks). A date, a quarter or a duration in the schedule is kept only when its source is named. Otherwise the check fails and the run stops. This section says what the conductor can time, how that limit is reached, and the line an unsourced date may not cross.

What can be timed

The next wave, as a distribution from work the doer has already finished. For software, the method is Joel Spolsky, Evidence-Based Scheduling, 26 October 2007 (https://www.joelonsoftware.com/2007/10/26/evidence-based-scheduling/). Break a task the person has done before into at most 16 hours. A larger package means the steps were not listed. Only the actor who will do the work estimates it. Another actor’s number is not a source. Do not press an estimate shorter to protect a date. Velocity is that actor’s estimate divided by the time the task actually took. His practice drops velocities older than about six months. A new estimator keeps a stand-in history, deliberately wide, until about half a dozen real tasks exist. Do not add the estimates into one ship date. Draw many futures. His essay draws 100, and treats each as 1 percent. In each future, divide each estimate by a velocity drawn at random from that history, then place the hours on that person’s calendar. The result is a distribution, not a day. In each round, the team is done when the last person finishes. A person shared with other work is not free for the whole of that round (§Critical path and resources).
A committed date. Only when the party who controls the date has committed, and the record names the artefact (a signed date, a written confirmation, a contract clause). An estimate is not that artefact.
A reference class of the same kind of finished work. Bent Flyvbjerg, “From Nobel Prize to Project Management: Getting Risks Right”, Project Management Journal, vol. 37, no. 3, August 2006, pp. 5-15 (arXiv:1302.3642), gives three steps. Identify a class of past similar projects, broad enough to be statistically meaningful and narrow enough to be comparable. Establish the probability distribution from credible empirical data for a sufficient number of projects in that class. Place this project on that distribution. The outside view uses the outcomes of similar completed actions. It does not forecast the events inside this one. The paper’s curriculum case is one project of that kind. The team’s own estimates ran from 18 to 30 months. About 40 percent of the comparable efforts were abandoned. The rest took seven to ten years. The work finished eight years later. Those figures time curriculum work of that kind. They do not time a software change. A different kind of work is not a source. Record that the class was rejected, and do not copy its figures.
The same method on a different class is still not this project’s numbers. Flyvbjerg, Hon and Fok, arXiv:1710.09419, submitted 3 October 2017, report a 2012 Hong Kong Development Bureau study. It covered 25 roadwork projects. Forecast costs and durations were compared with actual outcomes. The projects were benchmarked against 863 similar projects. The abstract states those counts. It does not state a percentage to lift. The roadworks distribution does not time software.

When nothing above exists

No history of finished tasks, and no opened class of the same kind, means the next work is to measure or to open a class. Until then the item is unscheduled. It carries no calendar date, no quarter and no duration. Spolsky’s half-dozen tasks are his stated practice for a software estimator. They are not a law for every trade. Say whose practice a threshold is.

What this procedure will not treat as a limit

The weighted average in §Estimates and arithmetic is an inside view of this task. It is not a reference class and it is not a measured velocity. The curriculum case is why: the inside estimates and the outcomes of similar finished work were different distributions.
One ship date as the sum of the estimates. Spolsky’s point is that the sum sounds right and is the wrong result.
A shorter estimate written to keep a date. His picture is a box of wood blocks. Use a bigger box, or fewer blocks. Do not shrink the blocks. If the date must move, cut scope or move the date. On Excel 5, the feature list would not fit the schedule, so the team cut it and called the cuts a deferral to the next version. When that next list was reviewed, not one deferred feature was worth doing. Cutting scope was the schedule. Shrinking the estimates would have hidden the choice.
An old date kept by adding people. His essay says new people will probably work at 50 percent for several months, and will slow the people who teach them. That 50 percent is his illustration, not a constant to paste into a plan. Adding people starts a new estimate.
A multiplier in place of a breakdown. He writes that thinking of the code without listing the steps makes the work seem to take n, when listing the steps makes it more like 4n. That is what an unlistable package hides. It is not a measured ratio to multiply into a schedule. The remedy is the 16-hour breakdown, not a multiplier.
A space-flight confidence level copied onto a small task. NASA’s planning page states a joint confidence level as the probability that cost is at or under the target and the schedule finishes at or under the target date. It describes CADRe as the historical record of cost, schedule and technical attributes for analogous projects, completed at each milestone and stored in ONCE (https://www.nasa.gov/ocfo/ppc-corner/ppc-guidance-documents/). The Schedule Management Handbook and Cost Estimating Handbook Appendix J are listed there. This procedure does not take a percentage from a handbook it has not opened, and it does not apply one agency’s threshold to a different kind of work.

How a limit is written, so a check can see it

In the schedule record, under a heading ## Schedule, each item uses one of these markers. The text after the colon is not empty.

measured: the doer’s own completed tasks, which velocities were kept, and the distribution that followed. For software, name Evidence-Based Scheduling when that is the method used.
committed: the artefact and the party who controls the date.
reference class: the opened source, the class, and the figure that source states. A class of a different kind of work is recorded as rejected, not used.
unscheduled: why there is no history and no opened class. No date, quarter or duration on that item.

scripts/check_time_limits.py reads that section. A date, a quarter or a duration with none of the first three markers fails, and the message names the claim. A marker with nothing after the colon fails. An unscheduled item that still states a date, a quarter or a duration fails. A file with no ## Schedule section passes, because it claimed no schedule. The check has to be run. A rule that is not run does not block the work.

When the date slips

Recompute. If the ship date moves later by more than one day per day, work is being added faster than it is finished. That is Spolsky’s stated test, not a new ratio. Cut scope or move the date. Do not edit the velocities to recover the old day.

Work records in the project’s tool

Use the fields and links the tool already has. For example, in an issue tracker with projects: a deliverable as a parent item with its work packages as sub-items; prerequisites as “blocked by” links; owner as the assignee; dates, estimates and state as project fields; milestones as the tool’s milestones; the board and the roadmap as views. In a spreadsheet: one row per work package with the same columns. Offline or with nothing else: one Markdown file with a table, kept in the repository.
No invented ID scheme or file format. Refer to items by the tool’s own identifiers (an issue number, a spreadsheet row). A plain file has none: number the rows of each table (1, 2, 3) and refer to them in words, such as “work item 3” or “the second measure”. Do not create families of prefixed codes (G1, M2, J3, W4, Q5) or rules for cross-referencing them between files; that is the private format this guidance exists to avoid.
Mark every export with its source and date, and do not edit it as a separate plan.
Use an existing linked view for shared state and takeover. Show objective, authority, accepted core, active work, duty modes, evidence and exceptions from the authoritative records. If no suitable view exists, ../templates/state.md is a small optional Markdown fallback, not another plan, store or source of authority. Verify the next operator’s actual access, authority and capability before declaring a takeover complete.