Estimate

estimate

Produce an honest, validated estimate: confirmed scope split from assumptions, a technique that fits the evidence, AI-assisted staffing scenarios, and an interactive what-if page served on localhost.

Hard rules

1. Every estimate row carries confidence + assumptions. A row nobody estimated renders not estimated — never 0.
2. Every scope item is labeled stated or proposed — no unlabeled scope.
3. Agent judges, script computes: every number in a deliverable comes from scripts/compute.mjs. Never total, average, or price by hand.
4. Never apply one blanket AI multiplier to a whole project — per-task category only (references/ai-multipliers.md).
5. node scripts/validate.mjs must exit 0 before the page renders.
6. Agentic estimates: baselines and confidence come from scripts reading measurements.jsonl — the agent never writes a duration, confidence, or evidence row.

Flow

1. Detect evidence: requirements/RFP? ARCHITECTURE.md? codebase? none? State findings; the user can override. (references/interview.md §1)
2. Depth and delivery mode: ask QUICK / STANDARD / DEEP, then TRADITIONAL / AGENTIC (references/interview.md §2b). Delivery mode decides step 4.
3. Interview: follow references/interview.md — pre-fill from evidence, ask only holes, run the clear-vs-assumed gate before sizing. Before proposing milestones, read references/slicing.md — slices are judged there, not computed.
4. Technique: TRADITIONAL-only — recommend from references/techniques.md, state why, cite the method's sources (its §Sources — attribution + link, never quoted text), confirm. AGENTIC reads references/agentic-estimation.md and references/task-shapes.md instead: no technique to recommend, shape + scope + seed minutes per task.
5. Size: write judgments to estimation-inputs.json (references/writing.md — the booking fixture is the canonical shape).
6. Compute: node scripts/compute.mjs --inputs estimation-inputs.json --out estimation.json
7. Write: estimation.md per references/writing.md.
8. Validate: node scripts/validate.mjs --md estimation.md --json estimation.json — fix findings, re-run until clean.
9. Render + serve: node scripts/render.mjs --json estimation.json --md estimation.md --out <dir> (add --client-only for a client-safe file) — render re-runs the validation checks itself and refuses on findings, so an unvalidated page cannot ship. deliveryMode: "agentic" routes rendering to the agentic HTML template automatically; nothing to choose here. Then serve with the analyze-requirements skill's serve.mjs; report the URL.

Companion mode

When ARCHITECTURE.md exists: §6 components seed the WBS, §15 risks seed the risk register, seed the components roster from §6 rows (use the LikeC4 model's container and component ids where one exists — dotted ids like atlas.goldStore) and tag every feature with the component it implements, flip the electedDocs estimation entry to elected: true, place estimation.md beside ARCHITECTURE.md so the analyze-requirements viewer picks it up as a companion page; link the interactive page from it. Render estimate.html into <dir>/dist/ — the same folder the viewer renders into. When a rendered viewer exists, add --viewer index.html for the back-link (internal view only; the client render strips it) — index.html and estimate.html then ship as one self-contained folder, and the viewer's estimation tab links the copy inside it.

Dependency

Node ≥ 20. No npm install needed — the scripts are dependency-free.