Target Workflow
Daily AI Poem — Selected as the highest-AIC workflow not optimized in the last 14 days (all higher-spend workflows were optimized 2026-06-17 or later). It posts a daily developer-themed poem as a GitHub Discussion.
Analysis Period & Runs
Period: 2026-06-18 – 2026-06-19 | Runs analyzed: 3 (all successful)
Spend Profile
| Metric |
Run 1 |
Run 2 |
Run 3 |
Avg |
| AIC |
17.686 |
17.390 |
16.653 |
17.24 |
| Input tokens |
49,326 |
49,190 |
48,760 |
49,092 |
| Cache-read tokens |
21,720 |
21,724 |
21,707 |
21,717 |
| Output tokens |
1,491 |
1,321 |
916 |
1,243 |
| Turns |
— |
— |
2 |
2 |
| Conclusion |
success |
success |
success |
100% |
Cache efficiency: 30.7% (cache-read / total input+cache). Model: claude-sonnet-4.6 (9× AIC multiplier). Tool calls per run: 1 (create_discussion only).
Ranked Recommendations
#1 — Downgrade Model from claude-sonnet-4.6 to haiku
Estimated savings: ~15 AIC/run (~87% reduction)
The workflow's entire task is: compute (day_of_month mod 7), write a 3–4 stanza poem, post it as a discussion. This is pure creative text generation with no code, no reasoning chains, and no tool orchestration. It completes in 2 turns in every run.
claude-sonnet-4.6 carries a 9× AIC multiplier. Switching to haiku (claude-haiku-4-5, 0.33× multiplier) would yield:
- New estimated AIC/run ≈ 17.24 × (0.33 / 9) ≈ 0.63 AIC/run
- Savings: ~16.6 AIC/run (conservative estimate: ~15 AIC/run accounting for possible token variation)
Smaller models are well-suited for: generating short-form creative writing, applying a simple formula (date mod 7), formatting output to a template. Haiku-class models routinely produce high-quality poetry with good rhythm and wit.
Action: Add model: haiku to the daily-ai-poem.md frontmatter:
---
description: Daily poem about AI, agents, and developer life — posted as a GitHub Discussion
on:
schedule:
- cron: "daily around 08:00"
workflow_dispatch:
permissions:
contents: read
model: haiku
tracker-id: daily-ai-poem
safe-outputs:
create-discussion:
title-prefix: "[poem] "
expires: 8d
max: 1
category: general
---Evidence:
- All 3 runs used
claude-sonnet-4.6; none required advanced reasoning, code analysis, or multi-step tool use
- Output tokens are small (avg 1,243), confirming the task is a short single-pass generation
References: §27740464271 · §27751122636 · §27818573144
#2 — Remove Unused GitHub MCP Tools from Agent Context
Estimated savings: ~1–2 AIC/run (after model downgrade applied)
The GitHub MCP server (24 tools) is loaded in every run but zero GitHub tool calls were observed across all 3 runs. The only tool used is safeoutputs.create_discussion. These 24 tool definitions inflate the ambient context and are injected into every request.
Similar workflows that require no GitHub reads (e.g., if the poem needed no repository context) can explicitly restrict their tool surface. The daily-xkcd.md workflow demonstrates the pattern with an explicit tools: allowlist in frontmatter.
Action: Add an explicit empty tools restriction to daily-ai-poem.md to prevent the GitHub MCP server tools from being injected into context:
This eliminates the 24-tool GitHub MCP schema from the model's tool list, reducing ambient context overhead.
Evidence: MCP gateway logs confirm 2 servers loaded (github, safeoutputs); only 1 safeoutputs.create_discussion call was made. The github server contributed 24 tool definitions to context, all unused.
Caveats
Generated by Agentic Workflow AIC Usage Optimizer · 415.1 AIC · ⊞ 21.6K · ◷
Target Workflow
Daily AI Poem — Selected as the highest-AIC workflow not optimized in the last 14 days (all higher-spend workflows were optimized 2026-06-17 or later). It posts a daily developer-themed poem as a GitHub Discussion.
Analysis Period & Runs
Period: 2026-06-18 – 2026-06-19 | Runs analyzed: 3 (all successful)
Spend Profile
Cache efficiency: 30.7% (cache-read / total input+cache). Model:
claude-sonnet-4.6(9× AIC multiplier). Tool calls per run: 1 (create_discussiononly).Ranked Recommendations
#1 — Downgrade Model from
claude-sonnet-4.6tohaikuEstimated savings: ~15 AIC/run (~87% reduction)
The workflow's entire task is: compute
(day_of_month mod 7), write a 3–4 stanza poem, post it as a discussion. This is pure creative text generation with no code, no reasoning chains, and no tool orchestration. It completes in 2 turns in every run.claude-sonnet-4.6carries a 9× AIC multiplier. Switching tohaiku(claude-haiku-4-5, 0.33× multiplier) would yield:Smaller models are well-suited for: generating short-form creative writing, applying a simple formula (date mod 7), formatting output to a template. Haiku-class models routinely produce high-quality poetry with good rhythm and wit.
Action: Add
model: haikuto thedaily-ai-poem.mdfrontmatter:Evidence:
claude-sonnet-4.6; none required advanced reasoning, code analysis, or multi-step tool useReferences: §27740464271 · §27751122636 · §27818573144
#2 — Remove Unused GitHub MCP Tools from Agent Context
Estimated savings: ~1–2 AIC/run (after model downgrade applied)
The GitHub MCP server (24 tools) is loaded in every run but zero GitHub tool calls were observed across all 3 runs. The only tool used is
safeoutputs.create_discussion. These 24 tool definitions inflate the ambient context and are injected into every request.Similar workflows that require no GitHub reads (e.g., if the poem needed no repository context) can explicitly restrict their tool surface. The
daily-xkcd.mdworkflow demonstrates the pattern with an explicittools:allowlist in frontmatter.Action: Add an explicit empty tools restriction to
daily-ai-poem.mdto prevent the GitHub MCP server tools from being injected into context:This eliminates the 24-tool GitHub MCP schema from the model's tool list, reducing ambient context overhead.
Evidence: MCP gateway logs confirm 2 servers loaded (
github,safeoutputs); only 1safeoutputs.create_discussioncall was made. Thegithubserver contributed 24 tool definitions to context, all unused.Caveats
gpt-5-mini(0.33× multiplier) as an alternative small-model option before reverting to sonnet.