agentsclimarketplace

Cost model

Skill avelikiy/great_cto/skills/cost-model

Don't buy software. Get the work done. GreatCTO ships AI autopilots that run a whole business function — medical coding, legal docs, procurement, accounting, IT, tax — from intake to outcome. A qualified human signs only the judgment calls. Live connectors, built-in compliance.

Install
npx -y skills add avelikiy/great_cto --skill cost-model

Assembled from the repository path, not quoted from the project. Check it against their README if it does not work.

What its author says it does

Copied from the file, not written here

Standardized cost-estimation framework for great_cto plans. Forces explicit LLM cost, infra cost, human-supervision time, and the (defensible) human-equivalent comparison. Output format is parsable by the board's /api/cost path — must follow exactly.

SKILL.md

5.0 KB, as published. Nobody here has run it

Cost model — make cost claims defensible

great_cto reports cost numbers on the board. Those numbers MUST be auditable, because a wrong "7,638×" claim killed credibility (see docs/blog/cost-dashboard-rebuild.md). This skill defines the format.

The 4-line cost section

Every PLAN-.md and ARCH-.md cost section follows this exact template:

## Cost estimate

**LLM**: $<low>–<high> (<N> calls × $<per-call avg>)
**Human equiv**: $<low>–<high> (<hours> × $<rate>/h)
**Infra delta**: $<low>–<high>/month
**Time to ship**: <hours> agent-time, <hours> wall-clock

> Methodology: <one-sentence rationale for each range>

Why this exact format?

The board's getCostHistory() parser anchors on line-start "LLM" and "Human" labels. Mid-line references are ignored to prevent the $240-trap regression. Stick to the template.

How to estimate each line

LLM cost

For each agent in the pipeline, estimate:

  • Prompt tokens = (system prompt size) + (context the agent receives)
  • Completion tokens = (typical output for that agent type)

Quick reference for Sonnet 4 ($3/M in, $15/M out):

AgentTypical promptTypical outputPer-call cost
architect14k1.5k~$0.06
pm6k0.6k~$0.03
senior-dev8k0.8k~$0.04
qa-engineer11k0.5k~$0.04
reviewer (avg)8-12k0.6k~$0.04
security-officer12k1k~$0.05
devops9k0.8k~$0.04

For Haiku ($0.80/M / $4/M), divide by ~4. For Opus 4 ($15/M / $75/M), multiply by ~5.

Sum across the pipeline stages that actually fire (use gatesFor() and reviewersFor() from archetypes.ts to know the count).

Human equiv

The human cost to do the SAME work without agents. This is the "if I hired a senior engineer, how long would this task take, at what rate?"

  • Senior engineer: $120-180/hour (mid-market US/EU)
  • Staff engineer / specialist: $200-300/hour
  • Domain expert (security, compliance): $250-400/hour

Estimate hours conservatively. A "small feature" the LLM does in 15 minutes might take a human 2-4 hours (it's never just the typing).

Infra delta

Only count what's NEW. If the feature adds a Redis instance, count Redis. If it adds 10MB/month of S3 storage, that's noise — don't list.

Time to ship

Two numbers — both useful:

  • Agent-time: wall-clock of LLM calls (typically 5-30 min)
  • Wall-clock: actual elapsed including human gates (typically hours to days)

Sanity check before writing

Before committing the section to the plan, verify:

ratio = human_equiv / llm_cost

If ratio > 1000, something is wrong. Common bugs:

BugHow to detectFix
Wrong unit ($ vs ¢)LLM cost ends in /M tokens not $Convert: tokens / 1M × price
Counting savings not spend"Human time saved" not "Human cost"Use cost of doing it, not value of skipping
Mid-line label pollutionPlan has "$X LLM$Y human" on one line
Forecast vs actual mixedLLM forecast counts toward total_llmSeparate forecast section if needed

Cost gates

For AI archetypes (mlops, ai-system, agent-product), the pipeline opens gate:cost after architect's forecast. CTO must approve the projected monthly burn before senior-dev starts.

Use the GATE template:

## Gate:cost forecast

| Production volume | Monthly LLM cost |
|---|---|
| 1K req/day | $X |
| 10K req/day | $Y |
| 100K req/day | $Z |

Recommended monthly cap: $<cap>
Triggers above cap: <what alerts fire, who gets paged>

Anti-patterns

Round-number theatre. "$0.50 LLM | $7,500 human" — looks suspicious. Use realistic ranges: "$0.50–1.20 | $225–360".

Single point estimates. Always provide a range. Single numbers hide uncertainty.

No methodology line. Just numbers without rationale is unverifiable.

Hand-waved infra. "Some hosting cost" is not a number. Either give $, or say "infra: no change."

Example — good

## Cost estimate

**LLM**: $0.75–1.85 (3 tasks × $0.25–0.62 per Sonnet call)
**Human equiv**: $225–300 (1.5–2h × $150/h, mid-market senior)
**Infra delta**: $0/month (uses existing Express + Postgres)
**Time to ship**: ~15min agent-time, ~3h wall-clock (1 human gate)

> Methodology: tasks sized by line-count estimate; per-call cost from
> historical Sonnet 4 averages on this archetype's plans.

Ratio = 300/1.85 = 162×. Plausible. Defensible.

Keep looking

Skills are one crate of 328,083. Ordering is by how many stacks a row turns up in, so the top of any crate is what has actually been picked rather than what has the most stars.