spend
Skill0 starsChanged 3 months agoMITView on GitHub
Diagnose where a Loop-Dee-Loup session spent tokens and whether that spend produced validated progress.
2 other projects use a version of this.
Adopted by
Most recently taken up by
- AvoDzis/spendtheir own version, first seen Nov 1, 2026
- runxhq/runxtheir own version, first seen Oct 13, 2026
Other versions2 projects use a version of this.
Word for word · 0
None found.
Edited · 0
None found.
Wrote their own version · 2
- runxhq/runx★ 97same name
- AvoDzis/spend★ 0same name
How matches are found. Word for word: the exact same text. Edited: shares passages with this file (“99% same” means 99% of its five-word phrases also appear here; 10% or more counts). Their own version: similar meaning, or, weakest, the same name. We can’t tell which project had it first. Projects are counted among the 200 most-starred candidates, and mentions are links to the repository we found, up to 50 per site, so both counts are a floor.
In this repository12 other files it gives its agents.
- AGENTS.mdAGENTS.md
- CLAUDE.mdCLAUDE.md
- context-clearingSkill
- local-workerSkill
- model-checkSkill
- persona-makerSkill
- retroSkill
- script-makerSkill
- siftSkill
- skill-makerSkill
- stage1-classifier-hardeningSkill
- telemetry-batterySkill
The file
We split this skill into the rules it gives an agent
10 rules. About 0 are common, 3 are distinctive, and 7 are very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈3
Distinctive · estimate
This author's own choices. Worth a look.
7
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Very specific: line 23
Read the printed record's three sections — measured, derived, unknown — and keep that separation in the final report.
- Very specific: line 26
Only when the record is missing, empty, or missing a field this analysis actually needs, fall back to the platform's own /usage or /context …
- Very specific: line 34
If measured.statusline_sample_count is 0, treat cost_usd_total, context_window_size, last_context_used_percentage, and last_token_usage as …
- Very specific: line 52
Never copy sensitive prompts, reasoning transcripts, or repository content into the report — neither the telemetry record nor …
- Very specific: line 82
Every /spend report renders one verdict per material claim it makes, not just one verdict for the whole session.
- Very specific: line 87
Before rendering a verdict for a claim that concerns token/cost allocation specifically, first check whether …
- Very specific: line 103
If tools/telemetry/sufficiency.mjs does not exist — an installed consumer repository per docs/consumer-contract.md that has not received …
- Distinctive: line 175
Do not produce a composite efficiency score, grade, ranking, target token count, or a "good/bad" threshold for commits, reviews, sessions, …
- Distinctive: line 196
Do not pad the report to fill every category — a clean session can produce a very short one.
- Distinctive: line 197
Never recommend weakening a safety, verification, review, audit, or founder-authority gate merely to improve a measured token number.
≈ “Common” is an estimate: at least 1 in 5 of the 350,495 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
Spend
Optimize validated progress per token, not minimum tokens. A session can be expensive and appropriate, or cheap and wasteful — token count alone never decides which.
This skill is the judgment layer on top of deterministic telemetry (tools/telemetry/, see
its README). It must not spend model reasoning reconstructing facts that mechanism already
measured or derived. It also must not treat that mechanism's output as proof of quality:
telemetry establishes what happened, never whether it was good.
Evidence order
- Reduce the current session's telemetry first, if not already reduced:
node tools/telemetry/reduce.mjs <session_id>(the running session's id is available from the statusLine payload'ssession_id, or from.claude/telemetry/sessions/if only one session is present). This is a deterministic script — running it costs no model tokens beyond issuing the command. - Read the printed record's three sections —
measured,derived,unknown— and keep that separation in the final report. Do not present ajudgedconclusion as if the record itself proved it. - Only when the record is missing, empty, or missing a field this analysis actually needs,
fall back to the platform's own
/usageor/contextreport for that specific gap — never as a routine replacement for the record, and never by parsing the transcript yourself to recompute a totalreduce.mjs(or/usage//context) already gives you deterministically.tools/telemetry/not existing in this repository at all (e.g. a consumer repository this hasn't been installed into yet, perdocs/consumer-contract.md) is the same case as an empty record — fall back for everything, and say so plainly rather than treating it as an error to work around. Ifmeasured.statusline_sample_countis0, treatcost_usd_total,context_window_size,last_context_used_percentage, andlast_token_usageas unavailable immediately and go straight to the/usage//contextfallback for them — do not spend a step probing session logs to rediscover this; seetools/telemetry/README.md's "statusLine's confirmed non-interactive gap" for why this is a known, evidenced condition rather than a one-off.measured.token_usage_main_total/measured.token_usage_subagent_by_agent_typeare a separate mechanism (recovered from the session's own transcript at SessionEnd/PreCompact, seetools/telemetry/transcript.mjs) and follow their own gate:measured.transcript_usage_sample_count === 0means they're unavailable this session — most commonly because the session hasn't reached SessionEnd or a compaction yet, or ran on a Claude Code build too old to supplytranscript_pathin hook payloads — there is no/usage//contextfallback for per-model/per-subagent-type token attribution specifically; report it as unavailable rather than reconstructing it from the transcript by hand. - Separately, inspect the smallest necessary durable repository state — the active issue's acceptance criteria, PR state, review/audit verdicts — to determine what outcome the session actually validated. Telemetry never establishes this; do not infer it from commit count, diff size, issue closure, or token consumption.
Never copy sensitive prompts, reasoning transcripts, or repository content into the report —
neither the telemetry record nor /usage//context output should contain any, but check
before including anything verbatim.
What telemetry can and cannot tell you
It can: session-total cost and duration (where statusLine fired); context-window usage over
the session, including the peak before a compaction, not just the last sample (where statusLine
fired); lines added/removed; structural subagent/compaction events (that one happened, its
type/trigger, when); and — from the session's own transcript, which works in this repository's
normal non-interactive execution mode where statusLine does not — total input/output/cache-read/
cache-creation tokens, broken down by model for the main thread alone
(measured.token_usage_main_by_model), by subagent type (measured.token_usage_subagent_by_agent_type),
and merged session-wide across main and subagent threads (measured.token_usage_session_by_model
— use this one, not token_usage_main_by_model alone, for "which models did this session's
tokens actually go to", since a subagent can run on a different model than the main thread),
plus the deterministic main-vs-subagent share of the session's token total
(derived.token_usage_grand_total, derived.token_usage_main_share_of_total).
It cannot: attribute tokens or cost to a specific skill invocation (no Claude Code interface
this collector uses exposes a skill-invocation boundary the way it does for subagents), break
tokens down per individual turn rather than per model/agent-type aggregate, compute monetary
cost from token counts (no pricing table — cost remains available only where statusLine's
cost_usd_total fired), or measure rate-limit consumption. tools/telemetry/reduce.mjs's
unknown list names each of these explicitly every run, and no other command in this workflow
closes that gap. If a conclusion needs that granularity, report it as unavailable rather than
approximating it from something else.
Evidence-sufficiency verdicts (CLEAN / NOT CLEAN / INCONCLUSIVE)
Every /spend report renders one verdict per material claim it makes, not just one verdict for
the whole session. A claim is any conclusion of the form "X was/wasn't a problem" or "X was/
wasn't appropriately allocated" — narrower structural observations ("no repeated compaction was
observed") are claims too, just claims with a lower evidence bar.
Before rendering a verdict for a claim that concerns token/cost allocation specifically, first
check whether tools/telemetry/sufficiency.mjs exists in this repository (a plain file check —
no script needed to answer that). If it exists, run the deterministic gate instead of
eyeballing the record:
node tools/telemetry/sufficiency.mjs <session_id> <claim_type>tools/telemetry/sufficiency.mjs's CLAIM_REQUIREMENTS names the claim types this gate covers
(token_allocation, monetary_cost_total, monetary_cost_by_model, compaction_frequency,
subagent_invocation_pattern) and the exact record fields each one requires. It returns
SUFFICIENT or INSUFFICIENT plus the specific fields that are missing — never estimate this
by hand, and never add a new ad hoc completeness rule inside this skill; extend
CLAIM_REQUIREMENTS instead when a new class of claim needs its own evidence bar.
If tools/telemetry/sufficiency.mjs does not exist — an installed consumer repository per
docs/consumer-contract.md that has not received tools/telemetry/ — this is the same case
step 3 of the evidence order above already names for individual fields: fall back for
everything rather than leaving the verdict undefined, and never promote a claim to CLEAN merely
because the deterministic gate itself is unavailable. Apply this fixed mapping instead of the
script, per claim type, using only the platform's own /usage//context report and this
session's own directly observable structure — never by parsing the transcript by hand to
reconstruct what the gate would have checked:
monetary_cost_total— SUFFICIENT only if/usageor/contextitself reports a total session cost figure; INSUFFICIENT otherwise.monetary_cost_by_model— always INSUFFICIENT. Neither/usage//contextnor any consumer fallback source breaks cost down by model; this matches the source-repository gate, which is always INSUFFICIENT here too (no local pricing table — see the README's "What it deliberately still cannot measure").token_allocation— always INSUFFICIENT. Whole-session main-vs-subagent token allocation by model requires the transcript-derived measurementtools/telemetry/transcript.mjsperforms; no consumer fallback source reconstructs it, and this skill must not attempt to by hand.compaction_frequency— INSUFFICIENT for the full claim ("how often, and why"). A narrower claim — "no compaction is visible anywhere in this session's own context" — may be SUFFICIENT when no compaction marker appears in the visible conversation, but do not promote that narrower observation into a whole-session frequency verdict: a compaction that already dropped early context is, by definition, no longer visible to check for.subagent_invocation_pattern— SUFFICIENT for the narrower claim "which subagents this session directly dispatched, of what type" when the full session is visible with no compaction (a directly observable structural fact, not hook telemetry). INSUFFICIENT if a compaction has occurred, since it may have dropped earlier subagent dispatches from visible context.
State plainly in the report which of the two paths (script or fallback) produced each verdict, so the evidence boundary stays explicit either way.
Two things token_allocation specifically guards against, beyond plain field presence: a
transcript_usage snapshot taken at PreCompact rather than SessionEnd only reflects usage
accumulated up to that point, not the whole session, so the claim also requires
measured.token_usage_is_session_complete === true — never render a whole-session allocation
verdict from a partial mid-session snapshot. And monetary_cost_by_model is always
INSUFFICIENT today (measured.cost_usd_by_model is always null) — this collector has no
local pricing table, so per-model cost is a distinct, currently-unanswerable claim from
monetary_cost_total; never let the total's SUFFICIENT verdict stand in for a per-model
question.
- CLEAN — the evidence needed for this claim is
SUFFICIENT, and nothing in it supports a material defect or recurring inefficiency. - NOT CLEAN — the evidence needed for this claim is
SUFFICIENT, and it supports one or more material defects or recurring inefficiencies requiring correction. - INCONCLUSIVE / INSUFFICIENT EVIDENCE —
assessSufficiency(or the equivalent reasoning for a claim type it doesn't cover) returnsINSUFFICIENT: material evidence this specific claim needs is unavailable. Never round this up to CLEAN. Say plainly which fields are missing (frommissingFields) and why (usuallymeasured.transcript_usage_sample_count === 0ormeasured.statusline_sample_count === 0— see the evidence order above).
Missing evidence for one claim does not make every claim in the report inconclusive. A
session with zero transcript_usage events can still render a CLEAN or NOT CLEAN verdict on
compaction_frequency or subagent_invocation_pattern — those claims' required fields
(measured.compaction_events, measured.subagent_start_events) come from hooks that fire
regardless — while its token_allocation claim must render INCONCLUSIVE. Keep each claim's
verdict scoped to the evidence that specific claim actually needed; do not let a strong
structural finding ("no review churn, no repeated founder interruption") get silently promoted
into a broader economic claim ("expenditure was appropriately scoped") that the same evidence
never supported. State the evidence boundary explicitly in the report instead.
Judgment
Using the record plus the outcome evidence from step 4 above, classify spend into: fixed
startup/instruction payload; issue and repository authority; tool input/output; repeated reads
or rediscovery; chat narration and status repetition; founder decision dialogue;
implementation/reasoning; verification and review; rework; handoff and durable-state
maintenance. Attribute what the evidence actually supports — a large subagent total from
derived.subagent_invocation_count/subagent_type_counts is not by itself evidence of waste,
and a high-token session that safely closed a difficult validated outcome is not either.
Do not produce a composite efficiency score, grade, ranking, target token count, or a "good/bad" threshold for commits, reviews, sessions, subagents, founder interactions, or lines changed. None of those quantities has an inherently desirable direction — more review spend may have prevented a defect; more sessions may have preserved cleaner context boundaries; a founder interruption may have been a mandatory decision, not overhead.
Report:
- validated outcome (from repository/issue/PR state, not from telemetry);
- compact measured session economics (from the record's
measuredsection); - material deterministic observations (from
derived); - spend that appears necessary for correctness or verification;
- avoidable waste, only with exact evidence — evidenced, not merely hypothesized from a large number;
- confidence/evidence limitations, including anything in the record's
unknownlist that would have mattered; - the CLEAN / NOT CLEAN / INCONCLUSIVE verdict for each material claim made, per the Evidence-sufficiency section above — not one blended verdict for the whole report;
- zero to three smallest justified corrections — it is a legitimate result to recommend none;
- what, if anything, should be measured in a subsequent fresh session.
Do not pad the report to fill every category — a clean session can produce a very short one. Never recommend weakening a safety, verification, review, audit, or founder-authority gate merely to improve a measured token number.
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

