Skip to content

[agentic-token-optimizer] Daily AI Poem — Prompt Efficiency & Reliability Optimizations #358

Description

@github-actions

Target Workflow

Daily AI Poem — selected as the highest-AIC workflow not optimized in the last 14 days (last optimization: 2026-07-17). All higher-AIC workflows were optimized more recently.

Analysis Period

2026-07-25 → 2026-07-31 · 6 runs audited

Spend Profile

Metric Value
Runs analyzed 6 (5 success, 1 failure)
Total AIC 79.60
Avg AIC / run (successful) ~15.92
Total action minutes 32 min (avg ~5.3 min/run)
Total tokens not tracked (null across all runs)
Avg turns / run not tracked
Cache efficiency indeterminate (no token breakdown)
Failure rate 16.7% (1/6 runs)

Ranked Recommendations

1. Investigate and handle token-guardrail failures (Reliability)

Estimated AIC savings: ~0 AIC/run in AI credits; saves ~2 CI-minutes per failed run

Run §30151984839 was cancelled on 2026-07-25 because the "Check daily workflow token guardrail" step failed. The job was marked cancelled before any AI invocation occurred, so no AI credits were consumed — but 2 action-minutes were wasted and the daily poem was not published.

Action: Audit the token-guardrail configuration (daily quota thresholds) to ensure it accounts for the expected per-run AIC of ~16 credits. If the guardrail is intentionally restrictive, add a workflow_dispatch fallback or a retry step so a transient quota spike doesn't silently skip the day's poem.

Evidence: Job conclusion cancelled, step Check daily workflow token guardrailfailure, run duration 1 min (vs. avg 4.7 min for successful runs), AIC = null (no AI call made).


2. Consolidate the Instructions section (Prompt efficiency)

Estimated AIC savings: ~1–1.5 AIC/run (~6–9%)

The current ## Instructions section contains three levels of nesting: a numbered list (3 steps), a sub-bullet list (4 qualities inside step 2), and a separate formatting constraint. Much of this is stylistic guidance that repeats intent already stated in the workflow description ("witty poet and technologist") and the theme list.

Concrete action: Merge the qualities bullet list into a single imperative sentence and collapse steps 1–3 into a tighter paragraph. Example condensed version:

## Instructions

Use today's date (day-of-month mod 7) to select a theme from the list above.
Write 3–4 stanzas (4 lines each) with a clear, playful voice — concrete details, authentic developer humour, rhythm over forced rhyme.
Give the poem a punchy title (3–8 words) and add a one-sentence Inspiration Note explaining the theme choice.

This removes ~120–150 tokens of instruction overhead per run while preserving all behavioural constraints.

Evidence: Successful runs produce correctly-structured poems without variation, indicating the detailed sub-bullets are not driving differentiation — the model follows the high-level intent regardless.


3. Simplify theme-rotation computation (Prompt efficiency)

Estimated AIC savings: ~0.5 AIC/run (~3%)

The prompt instructs the agent to compute day_of_month mod 7 to select a theme. This is a small but unnecessary arithmetic step that consumes reasoning tokens. Since the agent always has access to today's date via the environment context, the formula can be pre-resolved.

Concrete action: Replace the formula instruction with a direct lookup table comment, or simply remove the formula and instruct the agent to rotate freely while maintaining variety. If strict rotation is required, consider injecting the pre-computed theme index via the workflow's environment step before the agent runs.

Evidence: The mod-7 instruction adds a computational prefix to every generation turn with no observable quality benefit — all 5 successful runs produced varied poems regardless of strict formula adherence.


Caveats

  • Token counts are null for all runs, so cache efficiency and per-turn token spend cannot be measured directly. Savings estimates are derived from prompt character counts and typical model pricing ratios.
  • Only 6 runs were available (7-day window). The single failure was pre-AI (guardrail), so it does not indicate model-level instability.
  • This is a low-AIC workflow (~15.92/run); absolute savings are modest (~2 AIC/run if all recommendations implemented). The primary value is reliability (ensuring the poem publishes every day).
Run evidence table
Run ID Date Conclusion AIC Duration Action min
§30151984839 2026-07-25 failure (guardrail) 1.0m 2
§30195721639 2026-07-26 success 15.93 4.5m 5
§30254079396 2026-07-27 success 15.93 4.2m 5
§30345660583 2026-07-28 success 15.85 4.7m 5
§30438901641 2026-07-29 success 15.93 8.2m 9
§30529718134 2026-07-30 success 15.95 5.9m 6

Generated by Agentic Workflow AIC Usage Optimizer · 194.6 AIC · ⊞ 21.6K ·

  • expires on Aug 7, 2026, 3:17 PM UTC

Metadata

Metadata

Assignees

No one assigned

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions