Claude Last-24h Usage Feedback
Review of the pasted Claude usage panel against local Claude history, rich transcript logs, subagent metadata, and current session metadata.
Overview Stats
The local analysis window is anchored to the run timestamp: 2026-06-09T04:37:33Z through 2026-06-10T04:37:33Z. Claude compact history provided user prompt counts; rich per-project transcript JSONL files provided assistant usage, model, subagent, and context metadata.
Dashboard Confirmation
Bottom line: the pasted panel is directionally confirmed. The exact percentages are not locally reproducible, but most rows have a close local analogue. The clearest confirmed rows are the subagent-type rows: Explore, general-purpose, and Plan match local metadata almost exactly.
| Pasted panel claim | Local evidence | Verdict | What it means |
|---|---|---|---|
| 73% came from subagent-heavy sessions | Sessions with any subagent account for 67.6% of raw token volume. With a cache-read weight of 0.1, that becomes 75.0%. A stricter "sidechain >= main" definition ranges from 3.9% raw to 64.0% when cache reads are excluded. | Directionally confirmed, exact definition unknown | The panel likely uses a billable-ish or provider-side weighting, not raw transcript tokens. |
| 68% was while 4+ sessions ran in parallel | Current ~/.claude/sessions/*.json windows show 4+ open sessions during most local usage, but this method overcounts because open session windows remain active while idle. Depending on raw/no-cache-read weighting, it reports 97.5-99.2%. | Parallelism confirmed, exact 68% not reproducible | The warning is credible, but local session files are too coarse to mirror the dashboard. |
| 24% was at >150k context | Raw transcript tokens from records with context >150k are 35.7%. With a 0.1 cache-read weight, the share is 22.7%. | Confirmed as plausible | Long-context turns are a real driver. The exact share depends on cache-read weighting. |
| 17% came from subagents under workflow-subagent | Local agentType: workflow-subagent accounts for 14.3% raw usage and 16.3% if cache reads are weighted at 0.5. | Confirmed as close | The workflow research harness is a measurable usage driver. |
| deep-research 14% | No local subagent metadata uses agentType: deep-research. The two workflow fan-outs were deep-research-style landing-page research workflows and account for 14.3% raw usage as workflow-subagent. | Likely dashboard-only classification | Treat this as a provider label layered over workflow fan-out behavior. |
| Explore 4%, general-purpose 2%, Plan 2% | Local raw token shares: Explore 4.4%, general-purpose 2.8%, Plan 1.9%. | Confirmed | These rows align with local subagent metadata. |
Local Command Evidence
| Command in last 24h | Count | Representative prompt |
|---|---|---|
/pack | 4 | /pack install youtube-video-prelaunch-audit, /pack install icon-fixer, /pack install ship-end |
/clear | 4 | Context reset across content, gblock-party-redux, and agentic-skills sessions. |
/resume | 2 | Two resumes in gblock-party-redux. |
/status | 2 | Status checks in personal landing and gblock-party-redux sessions. |
/youtube-video-prelaunch-audit | 1 | Audit request for a YouTube video URL. |
/user-flow-map | 1 | Product flow mapping for personal workstation. |
/session-triage | 1 | Brief-me model loading/backgrounding question. |
/analyze-sessions | 1 | GitHub-based project status audit skill discussion. |
/ship-end | 1 | Shipping the personal landing session. |
Token And Cost Check
Claude rich transcript usage fields are per-assistant response records, not cumulative snapshots. The local total below is therefore the sum of assistant response usage in the selected window. Local logs do not expose direct invoice, subscription, credit, CCU, or dashboard dollar fields, so actual billed cost remains unavailable. The cost figures below are estimated API-equivalent costs using Anthropic's published Claude API pricing retrieved on 2026-06-10.
| Token class | Total | Share of raw total |
|---|---|---|
| Input tokens | 1,565,183 | 0.8% |
| Cache creation input tokens | 14,664,878 | 7.1% |
| Cache read input tokens | 188,181,853 | 90.9% |
| Output tokens | 2,571,938 | 1.2% |
| Raw total used for local shares | 206,983,852 | 100.0% |
Estimated Cost Formula
Formula: input_tokens * base input rate + 5m cache writes * 5m cache-write rate + 1h cache writes * 1h cache-write rate + cache reads * cache-hit rate + output_tokens * output rate. The usage window had 8,953,191 5-minute cache-write tokens, 5,711,687 1-hour cache-write tokens, and no unknown cache-creation bucket. Server-side web search requests were zero in the parsed usage records, so no web-search surcharge was added. Managed-agent runtime, subscription-plan effects, private discounts, dashboard-specific weighting, and provider-side adjustments are not locally visible and are excluded.
| Cost line item | Tokens | Estimated cost | Notes |
|---|---|---|---|
| Base input tokens | 1,565,183 | $15.62 | Fable 5 at $10/MTok; Opus 4.6 at $5/MTok. |
| 5-minute cache-write tokens | 8,953,191 | $103.67 | Fable 5 at $12.50/MTok; Opus 4.6 at $6.25/MTok. |
| 1-hour cache-write tokens | 5,711,687 | $101.69 | Fable 5 at $20/MTok; Opus 4.6 at $10/MTok. |
| Cache-read tokens | 188,181,853 | $174.74 | Fable 5 at $1/MTok; Opus 4.6 at $0.50/MTok. |
| Output tokens | 2,571,938 | $122.17 | Fable 5 at $50/MTok; Opus 4.6 at $25/MTok. |
| Total estimate | 206,983,852 | $517.90 | Transcript-usage estimate, not actual billing. |
| Model | Input | Cache creation | Cache read | Output | Estimated cost |
|---|---|---|---|---|---|
claude-fable-5 | 1,558,325 | 12,091,610 | 161,303,988 | 2,315,025 | $477.21 |
claude-opus-4-6 | 6,858 | 2,573,268 | 26,877,865 | 256,913 | $40.68 |
<synthetic> | 0 | 0 | 0 | 0 | $0.00 |
| Model | Raw token volume | Raw share | Estimated cost share |
|---|---|---|---|
claude-fable-5 | 177,268,948 | 85.6% | 92.1% |
claude-opus-4-6 | 29,714,904 | 14.4% | 7.9% |
<synthetic> | 0 | 0.0% | 0.0% |
Top Projects By Raw Usage
| Project | Raw token volume | Share | Estimated API-equivalent cost |
|---|---|---|---|
/home/georgeqle/projects/static-web/prod/personal-landing | 71,990,040 | 34.8% | $218.40 |
/home/georgeqle/projects/sandbox/tools/agentic-coding-os | 44,950,370 | 21.7% | $80.85 |
/home/georgeqle/projects/web/dev/gblock-party-redux | 29,714,904 | 14.4% | $40.68 |
/home/georgeqle/projects/static-web/prod/lexcorp/war-room | 23,187,555 | 11.2% | $55.28 |
/home/georgeqle/projects/content | 19,050,023 | 9.2% | $65.85 |
/home/georgeqle/projects/tools/dev/agentic-skills | 11,627,332 | 5.6% | $35.41 |
Subagent Evidence
There were 228 subagent JSONL files with activity in the window. The largest usage driver was workflow research fan-out. Only two workflow groups accounted for all workflow-subagent raw tokens.
| Subagent type | Raw token volume | Raw share | No-cache-read share | Usage records |
|---|---|---|---|---|
workflow-subagent | 29,505,297 | 14.3% | 39.0% | 2,279 |
Explore | 9,168,914 | 4.4% | 11.9% | 346 |
general-purpose | 5,835,559 | 2.8% | 2.7% | 160 |
Plan | 3,933,143 | 1.9% | 4.0% | 124 |
| Workflow group | Raw token volume | Share | Observed question |
|---|---|---|---|
wf_32bc8310-743 | 18,187,409 | 8.8% | YouTube-first creator personal landing pages; CTA strategy, funnel direction, social proof, and whether website-to-YouTube subscribe is a useful primary conversion surface. |
wf_f8c8bb92-504 | 11,317,888 | 5.5% | Popular and high-converting build-in-public indie hacker / tech creator personal landing pages; above-fold content, CTA strategy, social proof, transparency widgets, and conversion lessons. |
Subagent Prompt Terms
Term hits in subagent prompts show a research-heavy pattern: search 767, research 426, review 319, github 210, test 167, plan 130, validate 76, implement 54, audit 50, explore 43.
Recent Workflow Patterns
The dominant pattern is not a single bad session. It is a combination of multiple active sessions, long-context work, and research fan-out. The last-24h top sessions are enough to explain the dashboard warning even when exact percentages differ.
| Session | Project | Raw total | Subagent raw | Main raw | Distinct subagent paths |
|---|---|---|---|---|---|
7ee461db... | personal-landing | 71,990,040 | 31,961,115 | 40,028,925 | 4 |
cb049e9f... | agentic-coding-os | 43,175,147 | 0 | 43,175,147 | 0 |
452ab39d... | lexcorp/war-room | 23,187,555 | 2,258,651 | 20,928,904 | 6 |
ec08938f... | gblock-party-redux | 18,421,733 | 6,780,688 | 11,641,045 | 6 |
786176ee... | content | 13,456,144 | 1,944,182 | 11,511,962 | 2 |
bb655678... | gblock-party-redux | 6,258,798 | 3,653,920 | 2,604,878 | 3 |
Current Session Metadata
Open Claude session metadata showed several simultaneous busy/waiting sessions during the run: content, agentic-coding-os, personal-landing, lexcorp/war-room, gblock-party-redux, and omega-war were active or recently updated around 2026-06-10T04:40Z. This supports the parallelism warning even though the local files cannot reproduce the provider's exact 68% share.
Skill Recommendations
The best improvement is not "avoid all subagents." Subagents are doing useful work. The fix is to make fan-out deliberate, budgeted, and mode-aware.
| Rank | Recommendation | Type | Evidence | Validation expectation |
|---|---|---|---|---|
| 1 | Add budget/depth controls to the workflow-subagent or deep-research harness: --quick, --standard, --deep, source caps, extractor caps, and adversarial verifier caps. Require a preview before spawning more than 5 subagents. | Skill improvement | Workflow-subagent accounts for 14.3% raw usage; two workflow groups spawned 228 active subagent files in the window. | Focused benchmark or fixture test proving default mode caps fan-out and deep mode documents why extra agents are needed. |
| 2 | Adjust Claude-facing subagent guidance from "throw more compute" to "budgeted delegation": independent lanes only, default cap of 2-4 exploratory subagents, ask/preview for broad research fan-out, and summarize cost/context risk before launch. | Standing instruction / skill contract | Pasted panel says 73% subagent-heavy; local any-subagent share is 67.6% raw and 75.0% with a 0.1 cache-read weight. | Layer1 contract scan on provision-agentic-config output and generated AGENTS/CLAUDE blocks. |
| 3 | Create or promote a personal project-portfolio-status skill for your GitHub/project audit habit: scan selected GitHub repos, local project roots, recent commits/PRs, deploy clues, and produce a brief portfolio status memo. This should remain personal because it depends on your repo map and priorities. | New personal skill | Last-24h /analyze-sessions prompt explicitly said the GitHub audit pattern is personal rather than generally useful. Subagent prompt terms also include github 210 times. | Fixture with 2-3 fake repos and expected status memo sections; no GitHub Actions. |
| 4 | Improve analyze-sessions with a Claude rich-transcript reconciliation mode: parse ~/.claude/projects/**/*.jsonl, meta agent types, workflow journals, session files, and report provider-panel-compatible weighting variants. | Skill improvement | The current skill mentions Claude compact history but not rich Claude transcript parsing. This run needed rich transcripts to confirm subagent/context rows. | Layer1/parser fixture covering isSidechain, agentType, workflow-subagent, context >150k, and stale stats-cache.json. |
| 5 | Add a context/session checkpoint helper: before heavy research or when context exceeds 150k, recommend /compact; when switching projects, recommend /clear; when 4+ sessions are active, suggest queueing or pausing idle sessions. | Plugin/hook or standing instruction | High-context local share is 22.7-35.7% depending on cache weighting; parallelism is materially present in current session metadata. | Non-mutating check that reports active sessions, context-risk signals, and suggested action without blocking simple tasks. |
Highest-Impact Automations
| Automation | Manual prompts avoided | Why it matters |
|---|---|---|
| Research fan-out budget gate | Dozens of source extraction and adversarial verifier subagents per heavy research run | Directly attacks workflow-subagent and dashboard deep-research usage without removing the research workflow. |
| Personal GitHub portfolio status skill | Repeated custom instructions for repo/project status audits | Keeps personal portfolio intelligence out of general-purpose skills while making your own workflow fast. |
| Analyze-sessions Claude rich transcript parser | Manual reconciliation work like this report | Turns provider-panel validation into a repeatable mode with transparent caveats. |
| Context/session checkpoint | Manual reminders to compact, clear, or queue sessions | Addresses the two non-skill warnings: high context and many active sessions. |
| Cheaper-model defaults for simple subagents | Manual model-selection decisions | The pasted panel explicitly recommends this; local subagent types show many Explore/Plan/general-purpose lanes suitable for cheaper models where the runner supports it. |
Evidence Matrix And Assumptions
| Claim | Evidence | Inference | Confidence | Assumption status | Decision impact |
|---|---|---|---|---|---|
| The pasted usage panel is plausible but not exactly reproducible from local logs. | Local raw and cache-weighted shares bracket several pasted numbers; exact dashboard accounting formula is absent. | The panel is likely using provider-side weighting and classifications. | High | Provider dashboard is treated as an external calculation, not a local source. | Use local evidence for direction and recommendations, not exact percentage disputes. |
| Subagent usage is a real driver. | 67.6% raw token volume in sessions with any subagent; 48.4M raw sidechain tokens; 228 active subagent files. | Subagents should become budgeted and mode-aware. | High | "Subagent-heavy" remains dashboard-defined. | Approve research fan-out budget controls before broad new research. |
| Workflow research fan-out is the most concrete optimization target. | Two workflow groups account for 29.5M raw tokens and 14.3% raw usage. | The high-leverage fix is source/search/extractor/verifier caps. | High | Deep-research is inferred from workflow behavior and dashboard label. | Improve workflow/deep-research before changing every skill. |
| Explore, general-purpose, and Plan dashboard rows are confirmed. | Local raw shares are 4.4%, 2.8%, and 1.9% respectively. | Local metadata aligns with pasted rows. | High | Raw token volume is the comparison basis. | These lanes are candidates for cheaper model defaults. |
| The GitHub project status audit habit is personal-skill shaped. | User prompt in the window explicitly says the GitHub audit pattern is personal rather than generally useful. | A personal skill or pack is better than generalizing it into the shared catalog. | High | Need final scope: current repo only vs portfolio-wide. | Approve a personal skill only after naming target repositories and output cadence. |
| High context is material. | >150k context share is 35.7% raw and 22.7% under a 0.1 cache-read weight. | The pasted 24% is plausible under cache-discounted accounting. | Medium | Threshold uses local context calculation: input + cache creation + cache read. | Add compact/clear checkpoints for long sessions. |
Coverage Gaps
- The provider dashboard formula is not available in local logs.
~/.claude/stats-cache.jsonis stale for this question; itslastComputedDateis2026-04-17.- Claude session files are coarse active-window metadata, so they overcount parallelism when idle sessions remain open.
- No local
agentType: deep-researchexists; that label is inferred from dashboard classification and workflow research behavior. - Costs are unavailable because no direct cost fields or verified pricing table were used.
Review Gates
Answer these gates only after reviewing the evidence above. Feedback-only YAML can be compiled from any section feedback control before every required question is answered.
Research Completeness
Is the local evidence sufficient to accept this as a directional confirmation of the pasted usage panel?
Recommended Path
Which improvement direction should be treated as the next approved path?
Assumptions/Confidence
How should the dashboard/local mismatch be handled in future reports?
Artifact Destination And Proposed File Changes
Is this review artifact and its provenance/task tracking scope acceptable?
Post-Approval Route
After final approval, what should happen next?
Compile YAML
Use feedback YAML for emphasis requests, concerns, or clarification before answering all gates. Use final answers YAML only when every required gate is answered.