the AUDIT pillar

Real power-user signal on Claude Code waste, cost, model-routing, and workflow bloat, gathered via agent-reach (Twitter/X · OpenCLI backend) on 2026-07-06 and distilled for Mission Control's AUDIT pillar (VISION.md §2 — "the pillar the user named as the whole point"). Every quote below is a real tweet: handle · approx likes at capture · URL. Engagement is live and drifts ±a few; treated as order-of-magnitude signal, not precise. ~40 searches run; the frontier of existing tools (ccusage / CodexBar / OpenUsage) is mapped in §3.5.

One-line finding: The whole market shows people cost totals (ccusage: "you spent $1,100"). Nobody shows people waste with a cause — which skills, which model choices, which re-sent context, which cache misses turned that $1,100 into $1,100. That gap is the entire AUDIT pillar, and the most authoritative voice in the space (Armstrong, ~6.1k likes) says the fix is exactly what Mission Control is: visibility + better defaults + routing, not nags.

1. Cost / waste pain — what power users FEEL but can't SEE

The dominant, viscerally-felt problem: you pay full input price for your entire session context on every single message, and nothing shows you how much of that was avoidable.

• The "$20 hey" — the canonical waste story. @mardehaym · ~225 likes · ~105k views · <x.com/i/status/2073827260113506726>

  "A Reddit user lost $20 after typing 'hey' into Fable 5. … He pulled up the logs. 847,000 tokens on one message. Claude Code resends your entire session context with every message. System prompt, tools, full chat history. You pay for all of it as input. 'hey' carried 847,000 tokens of baggage with it. His total spend hit $336. Post got 1,400 upvotes on r/ClaudeAI."

  This is the single most-shared articulation of the felt-but-invisible waste. The user never sees the 847k until the bill lands. This is the mechanic the AUDIT screen exists to surface.

• The bill lands and it's mostly overhead you didn't notice. @theo · ~1.6k likes · <x.com/i/status/2064214943210324243>

  "10 days into my reactivated $200 Claude Code sub. According to ccusage, I've done over $1,100 in inference in that time. Most of my usage is just auditing work that 5.5 did."

  A pro with a tracking tool (ccusage) can see the total but frames his own spend as mostly low-value ("just auditing"). The total is visible; the waste breakdown is not.

• "You are probably wasting 80% of your context window." @DataChaz · ~880 likes · <x.com/i/status/2055929071733743693>

  "🚨 STOP BURNING YOUR TOKENS! If you use Claude Code, you are probably wasting 80% of your context window. I found 10 ace tools that will completely rescue your API bill."

• The waste is now so visible people trade screenshots of it. @meta_alchemist · ~410 likes · <x.com/i/status/2038992182812844261>

  "the problem with Claude Code leaking/wasting tokens is very apparent now — everyone's been sharing screenshots, so i created a full guide to limit/stop the waste … it's important to track your token usage via the right tools."

• The felt fix is context hygiene — but it's manual and un-instrumented. @brian_armstrong · ~6.1k likes · ~4.2M views · <x.com/i/status/2070670644577280109>

  "Keep Context Lean – Start fresh sessions when switching tasks. Scope file context narrowly. Disconnect unused tools. Don't just compact. The goal isn't fewer tokens used, it's fewer tokens wasted."

• The 5-hour/weekly limit is the daily pain surface. @zeyu1337 · ~20k likes · <x.com/i/status/2068284460089618602> ("start the clock for your 5 hour Claude Code usage limit while you fresh"); @johnlindquist · ~120 likes · <x.com/i/status/1947717007287161173> ("Finally hit the Claude Code MAX ($200) usage limit in a 5-hour span"); and a grassroots reply on theo's own thread: "even with this, it is burning like crazy. it is burning at impossible rate."

What they FEEL but can't SEE: the split between work tokens (the diff, the answer) and baggage tokens (re-sent history, tool schemas, skill descriptions, system prompt) — and, inside baggage, the fraction that was a cache miss (paid full price) vs a cache hit (paid ~10%). That split is sitting in every transcript's usage block and no tool renders it.

2. Model-routing mistakes — evidence people run the wrong model

Strong, repeated signal that model choice is the biggest lever people get wrong, and that the "just use the best model" reflex is now the expensive mistake.

• "Almost 100% overkill." @rewind02 (quoting AI builder Nate on Anthropic's own Fable docs) · ~25 likes · <x.com/i/status/2073139129055907930>

  "You probably only need to reach for Fable 5 five to fifteen percent of the time. If you're using it for everything, that is almost 100% overkill."

• The named playbook: right-size by task, cut the bill 60%. @PrajwalTomar_ · ~75 likes · <x.com/i/status/2029232665954828419> (this is the exact playbook in the user's own CLAUDE.md doctrine)

  "You're wasting money because you're using Opus for EVERYTHING. Stop. Here's how to match models to tasks and cut your bill by 60%: → Opus/Sonnet for strategy … → Codex for all coding tasks …"

• Phase-based routing is the power-user default. @cjzafir · ~2.8k likes · <x.com/i/status/2065104422762684745>

  "I'm burning 50% less weekly Claude Code limits now. Fable 5 high for planning · Codex 5.5 xhigh for execution · Fable 5 max for review."

  Echoed by @theo (8.3k likes · <x.com/i/status/2072481845363822914>): "I wrote up a big section in my CLAUDE.md on how to prioritize different models for different work when orchestrating workflows and subagents … Things that are unnecessarily token hungry (computer use, codebase analysis), I do with other models and report results back to Fable." And @shadcn's /improve (6.3k likes · <x.com/i/status/2064671802509410806>): *"use your most capable model to audit your codebase and write plans for cheaper models to execute later."*

• The routing intuition is often WRONG — "cheaper" ≠ cheaper. Two mistakes people make in the other direction:

  • Sonnet is not automatically the cheap choice. @theo · ~2.6k likes · <x.com/i/status/2072066764465393917>: "Sonnet 5 cost MORE than Opus 4.8 on the Artificial Analysis Intelligence Index." (also @theo ~3.6k · <x.com/i/status/2046284117155762506>: "Sonnet is still more expensive than GPT-5.4.")
  • Downgrading to an open model can raise effective cost via retries. @bridgemindai · ~550 likes · <x.com/i/status/2067925035168211016>: "You will prompt GLM 5.2 three or four times to accomplish what GPT 5.5 or Opus 4.8 nails in a single prompt."

  → Implication: a right-sizing audit must reason about cost-per-completed-task, not sticker price. Naive "downgrade everything" advice is itself a documented failure mode.

• The market is already reaching for automated routing. @ryancarson (1.1k likes · <x.com/i/status/2070876856317010406>): "I'm hitting $15-20k/month in token spend … switching to GLM 5.2 as default and only using frontier models for harder tasks." And Armstrong's thesis, verbatim: *"Ultimately, humans shouldn't be choosing models — AI can automate this task."* @cyrilXBT (120 likes · <x.com/i/status/2071948870712476109>) lists the model selector as a feature "most people have never opened": *"Match the model to the task instead of overpaying for simple things."*

Evidence people run the wrong model: the "Opus/Fable for everything" reflex is named as the #1 waste; the counter-reflex ("Sonnet is cheap, GLM is cheap") is documented as also wrong. Nobody can currently answer "for this session, did I overpay?" from their own logs. Mission Control can.

3. Skill / MCP / config bloat — the prompt tax in the wild

• The setup, not the model, is the lever — and the setup is where bloat hides. @ArchiveExplorer · ~3.7k likes · ~800k views · <x.com/i/status/2073136973162872897>

  "everyone's still picking between opus and sonnet like the model is the ceiling. it isn't. … CLAUDE.md → hooks → verifier subagent → skills → mcp → memory. you write the folder once, the folder runs the model. — skills/ - 33 muscle memories …"

  This is the thesis anchor for the whole app: SAME MODEL — DIFFERENT RESULT, decided by the .claude/ folder. Note "33 skills" as a deliberate loadout — a foil to a 108-skill catalog where 93 never fire. The folder is the product surface; the AUDIT pillar grades it.

• MCP tool bloat is both a cost tax AND a quality tax. @TheTuringPost · ~100 likes · <x.com/i/status/2073528308411814054>

  "Don't use too much tools, because tool-selection accuracy drops below 90% after 10–30 tools. See MCP servers as [context you pay for on every call]."

  Corroborated by @ericzakariasson (Cursor) · ~530 likes · <x.com/i/status/2066570396183548350>: "cli > mcp for stuff the model already knows. git, gh, npm, docker, file ops … costs almost nothing in context. [mcp] costs context." And the blunt skeptic @aidenybai · ~1.6k likes · <x.com/i/status/1958546886513590697>: "does anyone actually use MCPs in Cursor / Claude Code?" Every idle MCP server is schema tokens injected into every session — and past ~10-30 tools it degrades routing, not just cost.

• "Default settings" is where money leaks. @RoundtableSpace · ~390 likes · <x.com/i/status/2073168657799405769>: "If you keep running it on default settings, you're burning cash." The prompt tax is invisible precisely because it's the default.

• Effort level is a furnace nobody audits. @theo · ~8.3k likes · <x.com/i/status/2072481845363822914>

  "I only use Fable on 'high' effort for now. xhigh is token hungry. max/extra is a furnace with worse outputs than lower options imo."

  A Chinese reply on the same thread crystallizes it: "high 档够用… xhigh 确实烧得快" ("high is enough; xhigh really does burn fast"). Effort is a persisted settings.json default that silently multiplies every token — and no tool surfaces "you're running above what your work needs."

• Skills are exploding in count with no usage feedback loop. The ecosystem is pushing more skills hard — @sairahul1 (~1.5k likes · <x.com/i/status/2057019152460865961>): "DON'T USE CLAUDE WITHOUT SKILLS" (×7); Addy Osmani's agent-skills at 68,925 stars (@Granite0x ~1.75k · <x.com/i/status/2073540259208700311>); a "Find Skills" skill so you can search your own skills (@angeldot_ ~1.4k · <x.com/i/status/2070620788592369881>). The install pressure is enormous and the deletion / usage-audit side is entirely absent. Nobody tracks which of their installed skills ever fired — the exact gap the Skill Ledger fills.

The prompt-tax problem in the wild: every skill description, every MCP tool schema, every hook, and the effort default ride in every session — and the whole culture optimizes for adding them. The subtraction tooling does not exist.

3.5 The tooling frontier — who exists, and the gap they leave

Evidence: @0xSero (250 likes · <x.com/i/status/2047079658244476954>): "Install codexbar or ccusage … track token usage … export all your agentic traces." @theo donated $500 to ccusage's creator (160 likes · <x.com/i/status/2060496307530461473>) — it's the de-facto standard. @csaba_kissi on OpenUsage (90 likes · <x.com/i/status/2038591279001370876>); @ErickSky (65 likes · <x.com/i/status/2063044332622381084>): "exactamente cómo te desangran la billetera" ("exactly how they bleed your wallet").

The frontier stops at COST TOTALS and CREDIT BALANCES. Not one tool answers why the number is what it is, or what to change. No dead-skill ledger, no model-mismatch finding, no cache-miss attribution, no subagent-doctrine check. That "why + what to change" is Mission Control's entire white space.

4. What an AUDIT should REVEAL — ranked by "would a power user act on it?"

Ranked by (felt pain × invisibility today × one-click actionability):

1. Wasted-context / cache-miss dollars, per session. "Of your $1,100, $X was re-sent context you paid full input price for — here are the 8 sessions that did it." This is the "$20 hey" mechanic made visible and priced. Highest pain, totally invisible, directly fixable (fresh session / narrow scope / don't-just-compact). Armstrong names cache misses as "the easiest way to drive your cost up" (5%→60% hit rate cut his spend ~half).
2. Model-mismatch overspend. "This ran on Fable but its shape (few messages, one file, no Agent calls) says Sonnet work — est. overspend $X." The "100% overkill" finding, priced and per-session.
3. The dead-skill / prompt-tax ledger. "93 of 108 skills never fired; they cost ≈N tokens in every one of your 5,248 sessions." Turns "tidy your skills" into a priced decision — and there is no competing tool for this.
4. Effort-furnace flag. "Your persisted default is xhigh; doctrine says high. Est. multiplier on N sessions." theo: "max/extra is a furnace with worse outputs."
5. Subagent doctrine violations. "This orchestration spawned 14 subagents; 2 silently inherited Fable." The fork inheritance trap the user's CLAUDE.md warns about, caught red-handed.
6. MCP/tool overload. "3 MCP servers add ≈N schema tokens to every session; you're past the 10-30-tool line where selection accuracy drops below 90%." Cost tax + quality tax in one row.
7. Honest coverage math. The pxpipe "99%" lesson generalized: every hero stat carries its denominator (coverage ≠ depth ≠ scope). A number that can't survive a stranger's first question poisons trust in the whole dashboard.

The unifying move for all seven: cost → attributed waste → the one action that fixes it.

5. Futuristic audit asks — what nobody ships

• Auto right-sizing (the holy grail everyone gestures at, nobody ships). Armstrong: "humans shouldn't be choosing models — AI can automate this task." Every power-user "playbook" (Prajwal, cjzafir, theo, shadcn /improve) is a manual routing ritual encoded in prose in a CLAUDE.md. The unshipped product: a system that watches your actual sessions and recommends the routing, then bakes it into a launchable recipe. Mission Control's AUDIT→LAUNCH loop is exactly this, minus the auto-execute (which it correctly refuses to fake).
• Tasteful waste surfacing, not nags. Armstrong is emphatic — "Not with friction and spend alerts … Better Visibility — usage visible." The entire market ships either blunt caps (the 5-hour wall everyone resents) or nothing. Nobody ships calm, causal visibility: the number, the cause, the trend — no red banners. This is literally VISION.md's Armstrong-principle north star.
• Cost-per-completed-task, not cost-per-token. The GLM "3-4 prompts to do one" finding means the honest unit is task cost. No tool computes it. A retry-aware waste number would be genuinely novel.
• The setup as the audited artifact. ArchiveExplorer's "the folder runs the model" — nobody grades your .claude/ folder as a system (skills × MCP × hooks × effort × memory) with a legibility/health score. Everyone sells you more folder; nobody audits the folder you have.
• Context-baggage forecast before you send. The "$20 hey" horror is post-hoc. Pre-send "this message will carry 847k tokens of context ($8)" is unbuilt and would print money in trust.

6. Actionable implications for the AUDIT screen

Design doctrine, validated by the loudest authoritative voice (Armstrong ~6.1k likes): visibility + better defaults + routing + caching, never friction/alerts/caps. This is not our opinion — it's the consensus of the people spending $15-20k/mo. Ship calm causal numbers, not nags. (VISION.md §"Not a nag" is directly market-confirmed.)

Skill Ledger (VISION 2.2): the market is all install-pressure ("DON'T USE CLAUDE WITHOUT SKILLS", a skill to find your skills), and there is zero subtraction tooling — we'd be first. Price the dead skills in tokens × sessions (the ArchiveExplorer "33 muscle memories" foil makes a 108-with-93-dead catalog land as obvious bloat). Keep the "explicit invocations" honesty label.

Doctrine / model-mismatch audit (VISION 2.3): the "100% overkill" (rewind02/Nate) and phase-based routing (cjzafir/theo/Prajwal) tweets are the user's own doctrine, in the wild — so grading against it is legible to any power user, not just this one. Two must-haves the research forces: (a) reason in cost-per-completed-task so we don't repeat the naive-downgrade mistake (GLM retry tax); (b) present findings as review candidates with evidence, never verdicts (heuristic on transcript shape). The subagent Fable-inheritance row is the sharpest single artifact — it's the exact fork trap, caught.

Credit-era burn governor (VISION 2.5): the Jul-7 credit cliff is the topic on X right now (trq212, claudecode_lab, PrajwalTomar all counting down). Lead with the absolute daily $ + trajectory, tier-split, no alarms — Armstrong's "make usage visible, the goal is fewer tokens wasted." Label it API-rate-equivalent, never a fake bill.

Which OTHER on-disk findings would land hardest

(Beyond the two the user already computed: 93/108 skills never explicit-fired; 90 subagent runs touched Fable.) Ranked by expected "holy shit" per unit of implementation, all computable from ~/.claude transcripts + configs + ~/.pxpipe/events.jsonl:

1. ★ Cache-hit rate + cache-miss dollars (the single most valuable — see below). Each transcript usage block carries cache_read_input_tokens vs cache_creation_input_tokens vs plain input_tokens. Compute effective hit rate per session and $ paid at full price that a warm cache would have billed at ~10%. Armstrong: cache misses are "the easiest way to drive cost up"; 5%→60% ≈ half the bill. This is the "$20 hey" made auditable and is invisible in every existing tool.
2. Baggage-vs-work ratio + the heaviest-context offenders. Per message: input tokens (baggage: re-sent history + tools + skills + system) vs output tokens (work). Surface "your top-10 sessions averaged ~Nk input tokens/message; ~M% was re-sent context." Names the specific sessions that are the user's personal "$20 hey." (VISION already has cost-per-message math — extend it.)
3. Effort-furnace exposure. Read persisted effortLevel from settings.json; count sessions run above high and estimate the token multiplier. theo's "furnace" quote gives it a name a stranger gets instantly. Trivially computable, high pain.
4. Model×task-shape overspend (broader than the 90 subagent runs). Across all sessions, the fraction of Fable/Opus token spend that went to sessions whose shape says Sonnet (few messages, ≤1 file edited, no Agent/Task calls). Turns "90 subagent runs" into "$X of your frontier spend looks like Sonnet work." usageByTier + isSubagent already exist.
5. MCP/tool always-on tax + the 10-30 accuracy cliff. 3 MCP servers × their tool-schema token footprint × session count = a concrete "you pay N tokens/session for tools." Flag if total tool count crosses the ~10-30 line (TheTuringPost) where selection accuracy degrades — cost and quality in one finding.
6. Compaction / abandoned-heavy-session waste. Count /compact events (cache-reset cost, per Armstrong "don't just compact") and sessions that ballooned then went idle mid-context (paid peak weight for turns that led nowhere).
7. pxpipe coverage honesty (VISION 2.1). 298/445 compressed = 67% coverage; the "99%" is depth-over-compressed-only. First credibility fix — every other number inherits its trust.

The single most valuable audit-finding to surface: cache-miss / re-sent-context dollars per session — "of your $X spend, $Y was context you re-sent and paid full input price for; a warm cache would have billed it at ~10%. Here are the sessions that leaked the most." It wins because it is (a) the exact mechanic behind the most-viral pain story ("$20 hey," 847k tokens on "hey"), (b) felt by everyone ("burning tokens", "wasting 80% of your context") yet rendered by no existing tool — ccusage/CodexBar/OpenUsage stop at the total, (c) endorsed as THE lever by the most authoritative voice in the space (Armstrong: cache misses = "the easiest way to drive cost up", 5%→60% ≈ half the bill), and (d) the data is already in every transcript's usage block, so it is shippable now. It converts an inert "$1,100" into an actionable, causal, per-session waste number — which is the whole point of AUDIT.

Method notes / honesty

• Backend: agent-reach → opencli twitter search "<q>" -f yaml and opencli twitter thread <id>; ~40 queries across cost/waste, model-routing, config/skill/MCP bloat, orchestration cost, observability, and existing tools. Full raw YAML retained in scratchpad for re-verification.
• Noise filtered, not cited: several searches were dominated by unrelated viral content — a recurring "hidden spyware in Claude Code" hoax (@IntCyberDigest ~17.6k), Karpathy "vibe coding" /idea-file threads, the Claude Science launch, and Fable-jailbreak drama. These out-rank on likes but are off-topic for waste/routing and were excluded.
• Failed/empty command: opencli twitter search "claude code model routing" (slug x12) timed out (>2 min, no result) inside a 12-query batch; the other 11 in that batch returned. Not retried — adjacent queries (x06 claude code right model, m05 claude code model selection, m12 intent) already covered the topic. All other ~39 searches returned 15 results each.
• Engagement counts are live and drifted ±1-3 between reads (e.g. mardehaym 228→224/225); treat as order-of-magnitude. Every URL is a real, resolvable tweet as of 2026-07-06.