Product blog
Three authoring methods — measured time and ROI
Same business goal, three workflows — with field measurements, a cost model in CNY, and calculators you can tune.
Scenario: after login, scan a list, open entries that match a rule, run a fixed sub-flow on each. Below we compare how teams actually build that automation in Automation Skill Builder. Step-by-step loops, branches, and MCP recording (especially Method 3) are in the companion post From linear recording to loops and branches.
The three methods
| Method | Summary |
|---|---|
| Method 1 — Pure AI code | Describe the flow; AI writes Playwright/Python from scratch. No ASB required for authoring. |
| Method 2 — Inner record + exe, outer run_code | Record inner steps once, parameterize and package as exe/skill; record outer loop and call inner via run_code. Best reuse for nested flows. |
| Method 3 — Full record + AI refactor | One MCP-recorded pass, paste script, ask AI to add loops/conditions. Fastest for first delivery. |
Measured baseline (single loop)
Empirical data from the same product team (GLM-class model, familiar with AI assistants; single-loop list task):
| Method | Build time (measured / est.) | AI debug rounds | ASB subscription |
|---|---|---|---|
| Method 1 — Pure AI code | ~60 min | ~10–15 | Not required |
| Method 2 — Inner record + exe, outer run_code | ~105 min | ~5–8 | Yes (~$29/mo) |
| Method 3 — Full record + AI refactor | ~25 min | ~3 | Yes (~$29/mo) |
Cost model: labor, API, and ASB
Per skill, total cost ≈ labor + API + ASB allocation (Methods 2–3). Default labor ¥360/hr (~$50/hr at FX 7.2).
- Labor dominates at single-loop complexity — API matters more only at high token volume.
- Measured Method 3 authoring used ~3.1M tokens in one long chat (~82% input) — agentic context replay, not a 5k-token chat. Use a fresh chat for comparable benchmarks.
- Method 1 token multiplier is modeled higher (~4× Method 3 rounds); replace with your measured runs when available.
GLM-5 on OpenRouter vs SiliconFlow
Calculators use GLM-5 list pricing: OpenRouter about $0.60/M input and $1.92/M output; SiliconFlow about ¥4/M input and ¥16/M output (roughly 35–40% lower API at the same token volume). Switch platforms in the tool — labor and ASB lines are unchanged.
Runtime: zero model tokens
Packaged skills run locally through ASB — measured runtime LLM tokens = 0. A $29/mo plan with ~20,000 runs/month spreads broker cost across executions; building skills is where API spend happens.
When nesting changes the winner
Flat loops (1L–4L): Method 3 stays ahead. One nested loop (1N): Method 2’s inner module + outer shell starts to pay back. Deeper nesting (2N+): reuse multiplier on inner exes compounds — adjust “reuse count” in the calculator to see crossover.
Method 1 has no ASB line — at very low monthly skill volume it can look cheaper on paper than Methods 2–3 that carry subscription; crossover shifts as soon as build time or volume grows.
Interactive ROI calculators
Use the calculator below (sliders for labor ¥/hr, skills/week, runs/month, reuse, FX). It follows the site language and recalculates live.
Serve product-site over HTTP for Chart.js and iframe embed. Open in a new tab if the embed is blank.
Practical recommendation
Default: Method 3 to prove the flow (~25 min in our sample). Promote to Method 2 when the same inner block ships in 3+ flows or nesting makes monolithic refactor fragile.
Method 1 fits developers and pure-code paths without ASB; split “run once via MCP” and “refactor in a new chat” to avoid context pollution. Enterprise gaps (cross-app files, hooks, audit via machine id + service key) are orthogonal — see ecosystem and positioning posts.