Target Workflow
Release (workflows/release.md) — the only non-excluded agentic (LLM-driven) workflow not in scope for the "Token in name" exclusion and not optimized in the last 14 days with substantial analysis. Selected because its LLM prompt is the largest per-run payload relative to its output; the two prior entries in the optimization log audited only 1–2 runs each with minimal findings.
Analysis Period & Runs Analyzed
| Period |
Runs |
Conclusions |
| 2026-05-14 → 2026-05-26 (12 days) |
2 |
2 success, 0 failure |
Note: Token instrumentation returned 0 values for all workflows in the audit data for this period. Token estimates below are derived from prompt character counts and typical LLM encoding ratios (~0.25 tokens/char for English prose). Actual values may differ; estimates are conservative.
Token Profile
| Metric |
Value |
| Runs analyzed |
2 |
| Avg agent job duration |
~2 min 13 sec |
| Prompt body (chars) |
~7,600 |
| Estimated prompt tokens |
~1,900 |
| Estimated output tokens |
~400 |
| Total estimated tokens/run |
~2,300 |
| Cache efficiency |
Unknown (no telemetry) |
References:
Ranked Recommendations
1. Remove the "Selection rubric" section (est. –125 tokens/run)
The ## What to Write section already instructs the agent to focus on user-facing changes and exclude internal ones. The "Selection rubric for whether a change belongs in the release notes or changelog" (~500 chars) restates this guidance in a different form without adding new constraints.
Action: Delete the "Selection rubric..." paragraph block from the prompt. The three include/omit/collapse bullets are already covered by earlier instructions and the "Exclude internal-only changes" section that follows immediately after.
2. Merge "Anti-patterns to avoid" + "Writing quality constraints" into a single short list (est. –150 tokens/run)
These two sections (~1,100 combined chars) have significant overlap. "Avoid repeating the same fact in both ## Highlights and ## Upgrade Notes" is also covered by the structure instructions. "Keep bullets parallel and scannable" duplicates "use plain English and strong verbs."
Action: Replace both sections with a single **Avoid:** bullet list and a one-line style note:
**Avoid:** rewriting commit subjects as bullets; listing changed files as user value;
vague filler ("various fixes"); overstating maintenance changes; contributor attribution
in the release body; repeating the same fact in both Highlights and Upgrade Notes.
**Style:** plain English, strong verbs, outcome-first bullets, concise over exhaustive.
Estimated reduction: ~700 chars → ~200 chars, saving ~125 tokens per invocation.
3. Condense the semver guidance section (est. –65 tokens/run)
The four-bullet semver list (major / minor / patch / 0.y.z) restates the general principle ("infer from user-visible diff, not from commit prefixes") four times with minor variation. The final bullet restates the major-release bullet.
Action: Collapse to two bullets:
Major: breaking changes, removals, migration; Minor: new capabilities, deprecations; Patch: fixes, polish, reliability.0.y.z: note pre-1.0 status; call out breakage plainly.
Estimated reduction: ~400 chars → ~150 chars, saving ~65 tokens per run.
4. Remove "Treat the public surface" paragraph (est. –50 tokens/run)
The paragraph "Treat the public surface of this project as the workflow behavior users depend on: workflow names, supported triggers and inputs..." (~200 chars) restates what is already implied by the user-facing framing throughout ## What to Write.
Action: Delete this paragraph. The subsequent anti-patterns and style sections already scope the content appropriately.
Summary
| # |
Recommendation |
Est. savings/run |
| 1 |
Remove "Selection rubric" section |
~125 tokens |
| 2 |
Merge anti-patterns + style constraints |
~150 tokens |
| 3 |
Condense semver guidance bullets |
~65 tokens |
| 4 |
Remove "Treat the public surface" paragraph |
~50 tokens |
|
Total |
~390 tokens/run |
Combined, these changes reduce the prompt body by approximately 20% (from ~7,600 to ~6,100 chars) with no behavioral change — all preserved intent is already stated elsewhere in the prompt.
Caveats
- Token telemetry returned 0 for all workflows in the current audit window; all savings estimates are derived from prompt character counts and approximate encoding ratios, not measured values.
- The Release workflow runs infrequently (≤2 runs in the 12-day window), so aggregate monthly savings are modest (~800 tokens/month). However, reducing instruction redundancy also lowers the risk of conflicting guidance being applied inconsistently.
- No structural (sub-agent) changes are recommended: the single LLM task — synthesizing a changelog from diverse inputs (commits, file diffs, prior releases, workflow sources) — requires cross-referencing multiple heterogeneous sources and strategic synthesis. This is not well-suited to a smaller extractive model.
Generated by Agentic Workflow Token Usage Optimizer · ● 7.8M · ◷
Target Workflow
Release (
workflows/release.md) — the only non-excluded agentic (LLM-driven) workflow not in scope for the "Token in name" exclusion and not optimized in the last 14 days with substantial analysis. Selected because its LLM prompt is the largest per-run payload relative to its output; the two prior entries in the optimization log audited only 1–2 runs each with minimal findings.Analysis Period & Runs Analyzed
Token Profile
References:
Ranked Recommendations
1. Remove the "Selection rubric" section (est. –125 tokens/run)
The
## What to Writesection already instructs the agent to focus on user-facing changes and exclude internal ones. The "Selection rubric for whether a change belongs in the release notes or changelog" (~500 chars) restates this guidance in a different form without adding new constraints.Action: Delete the "Selection rubric..." paragraph block from the prompt. The three include/omit/collapse bullets are already covered by earlier instructions and the "Exclude internal-only changes" section that follows immediately after.
2. Merge "Anti-patterns to avoid" + "Writing quality constraints" into a single short list (est. –150 tokens/run)
These two sections (~1,100 combined chars) have significant overlap. "Avoid repeating the same fact in both
## Highlightsand## Upgrade Notes" is also covered by the structure instructions. "Keep bullets parallel and scannable" duplicates "use plain English and strong verbs."Action: Replace both sections with a single
**Avoid:**bullet list and a one-line style note:Estimated reduction: ~700 chars → ~200 chars, saving ~125 tokens per invocation.
3. Condense the semver guidance section (est. –65 tokens/run)
The four-bullet semver list (major / minor / patch /
0.y.z) restates the general principle ("infer from user-visible diff, not from commit prefixes") four times with minor variation. The final bullet restates the major-release bullet.Action: Collapse to two bullets:
Major: breaking changes, removals, migration;Minor: new capabilities, deprecations;Patch: fixes, polish, reliability.0.y.z: note pre-1.0 status; call out breakage plainly.Estimated reduction: ~400 chars → ~150 chars, saving ~65 tokens per run.
4. Remove "Treat the public surface" paragraph (est. –50 tokens/run)
The paragraph "Treat the public surface of this project as the workflow behavior users depend on: workflow names, supported triggers and inputs..." (~200 chars) restates what is already implied by the user-facing framing throughout
## What to Write.Action: Delete this paragraph. The subsequent anti-patterns and style sections already scope the content appropriately.
Summary
Combined, these changes reduce the prompt body by approximately 20% (from ~7,600 to ~6,100 chars) with no behavioral change — all preserved intent is already stated elsewhere in the prompt.
Caveats