Mined from the @thejoeywatts home timeline on 13 August 2026. Every post below is a real item from that feed. Every repo star count was fetched from the GitHub API during the same session.
python3 ledger_verify.py x-mine-ledger.jsonl to recompute every number on this page.api.github.com directly. Where a post's claim could not be checked, it is labelled unverified and the claim stays attributed to the poster.The first version of this page printed 56 as the AI-related count. That figure came from the looser keyword filter applied to the first 73 posts captured, and it was published beside 41, which came from the stricter filter applied to all 100. Two filters, two denominators, one line. Building the ledger and running both classifiers across all 100 rows gives the corrected 68. The verifier now refuses the old number: passing --expect strict=56 returns LEDGER: FAIL, published 56, ledger 41.
These eight define the pattern I mined for. Each is a harness fix, meaning enforcement lives in a hook, a gate, or a measurement rather than in a request.
| Fix | Shape | Where it lives |
|---|---|---|
| A hook that refuses to let the agent stop | enforcement moved out of the prompt | feedback_never_say_go_just_work_hard_gate |
| The completion evaluator reads the transcript, never the file | a checker blind to artifacts cannot catch a false claim | reference_goal_evaluator_judges_transcript_not_artifacts |
| Processed tokens are not paid tokens | 99% is cache read at 0.1x, an 8.5x gap | reference_transcript_token_cost_use_base_input_equivalents |
| One JSONL record per content block | 4,578 records for 2,372 real requests | reference_claude_code_jsonl_one_record_per_content_block |
| Independent tool calls go out in one block | measured 0.0% across 359 turns | feedback_batch_independent_tool_calls_hard_gate |
| Call the skill, never hand-roll it | 0 skill calls across 1,313 turns | feedback_invoke_skills_not_raw_paths_hard_gate |
| End long sessions, write state, clear | context hygiene as an operating rule | feedback_end_long_sessions_clear_hard_gate |
| Turn a repeated failure into a skill | codify the fix instead of re-explaining it | today's transcript, the impossible-guitar-parts pass |
Score is usefulness to your setup, 0 to 100. It drops when you already have the thing and it drops further when adopting it would break a standing policy.
| # | Finding | From | Maps to your fix | Verdict | Score |
|---|---|---|---|---|---|
| 1 | Invisible Unicode characters survive a word-level voice gate | @josesilesdata, @JespernissenSEO | anti-AI voice gate, text | new | 92 |
| 1b | C2PA provenance metadata inside generated image and PDF files | @josesilesdata | no gate covers this, file level | new | 58 |
| 2 | A 202k-star repository of Claude Code guidance, CLAUDE.md plus a skills directory | @charliejhills | your CLAUDE.md | new | 78 |
| 3 | Skills framework, 271k stars | @charliejhills | your 73 skills | new | 74 |
| 4 | Map the codebase into a dependency graph before the agent reads | @ai_explorer25 | ~/MAP.md, /ask | new | 66 |
| 5 | The agent pays full price to re-read its own transcript | @Roxx_0x | token truth | have | 64 |
| 6 | Harness built from zero, 74k stars | @charliejhills | reading | new | 60 |
| 7 | Configuration drift, nothing breaks and it just gets worse | @alex_prompter | your /drift skill | have | 55 |
| 8 | One file holding conventions, rejection rules, skills, escalation, contracts, logging | @Sprytixl | your CLAUDE.md | have | 52 |
| 9 | The agent read the same file four times in one run | @hanakoxbt | batching gate, session-audit | have | 50 |
| 10 | Karpathy's LLM wiki pattern as a Claude Code plugin | @0xkkai | Brandon-Vault | new | 45 |
| 11 | A writing harness that loads voice DNA and forbidden patterns, then checks the draft | @shannholmberg | /deslop | have | 42 |
| 12 | Drop-in memory layer, 63k stars | @N01ennn | your memory directory | have | 40 |
| 13 | Loop engineering, schedule then discover then build then verify | @ajay4ai, @arle0x | /loop, stop hooks | have | 38 |
| 14 | Obsidian as a memory layer an agent can read | @kocer_eth, @Nazik2053, @cyrilXBT | Brandon-Vault plus /ask | have | 35 |
| 15 | Run agents as a company with roles, a boss and budgets, 78k stars | @gippp69 | one-session policy | conflicts | 25 |
| 16 | Turn Claude Code into subagent graphs | @LunarResearcher | no-fanout policy | conflicts | 20 |
Claims model output carries invisible characters, C2PA metadata in files, and bias in token selection. A second post the same week, @JespernissenSEO on 12 Aug, points at the same tool independently.
Scope of this finding: text only. A Python text gate can inspect every codepoint in a string, so this one is inside its reach. The C2PA half of the same post is a separate problem and is filed as finding 1b below.
Why it scores highest: I read the gate rather than trusting the post. Before 13 August, ~/.claude/skills/_shared/anti_ai_voice_gate.py ran 362 lines and 23 pattern groups with no check for zero-width or invisible Unicode. The 132-tell corpus matches words. A zero-width space, a narrow no-break space, or a soft hyphen contains no word, so none of those groups catch it, and it stays in the file through a book chapter or an HTML page.
scan_invisible() now runs inside the gate at three tiers: severity 3 for zero-width characters, bidi controls, fillers, variation selectors, tag characters, and a stray BOM; severity 3 conditionally for U+200D and U+FE0F, judged by their neighbours so emoji sequences pass; severity 2 for the non-breaking space family, reported without failing.drwu-htmls/public/ scanned, 1 real hit and 0 false positives. The hit is a live U+200B at line 257 of tame-impala-charts.bestseller_format.py, and every HTML pushed to GitHub Pages.Filed separately because the tool that would catch it is a different tool. C2PA is provenance metadata written into image and PDF containers. A text gate reads a decoded string and never sees a container, so no amount of work on anti_ai_voice_gate.py reaches it. Catching this needs file-level inspection of the bytes, closer to exiftool or a C2PA reader than to a regex pass.
Described at the repository level, from its actual tree. GitHub's own repo description calls it a single CLAUDE.md derived from Karpathy's observations on LLM coding pitfalls, and the repository holds more than that file. Its top level carries 8 entries: CLAUDE.md at 2,357 bytes, CURSOR.md at 1,955, EXAMPLES.md at 14,838, README.md, README.zh.md, a skills/ directory, a .claude-plugin/ directory, and .cursor/. Whole repository size is 20KB.
The post's framing is that the agent guesses at structure until something maps it. That is the same reason you built ~/MAP.md and the /ask front door. The added piece is ranking, a dependency graph with PageRank so the agent reads the load-bearing files first rather than the first file it finds.
~/Projects subtrees more than the whole tree.These arrived as discoveries in the feed. You wrote each of them down first, and in four cases you have a measured number behind yours. The posts give none.
"every turn the transcript walks back in and takes the whole window." The proposed fix is two markdown files and four operations, write, retrieve, consolidate, forget.
Yours is the stronger version. You measured the same effect yesterday and found the payment claim is wrong in one direction: 99% of that returning context is cache read at 0.1x, an 8.5x gap between processed and paid. His protocol is close to what MEMORY.md plus POINTERS.md already do, with consolidate and forget as the two operations you run by hand.
"forty one tool calls, no errors, correct answer. four of those calls fetched the identical four thousand lines, and three of them bought nothing."
Your session-audit skill already counts duplicate Read calls as one of its named metrics, and your batching gate came from the same class of finding, measured at 0.0% multi-call turns across 359 turns.
"A skill routes to the wrong place occasionally. A naming convention loosens. A scheduled task stops firing and nothing downstream complains for weeks."
You have a /drift skill installed. This post is outside confirmation that the failure mode is real and that silence is its signature.
Strip the framing and the six-slot list is a decent audit of a CLAUDE.md. Yours is thin in two of the six slots: escalation, meaning what the agent does when a gate blocks it, and what to log.
Load the voice files, draft, then check the draft against them. That is /deslop plus anti_ai_voice_gate.py. Yours runs as a script. The post describes a manual loop.
Schedule, discover, build, verify, repeat. You run /loop, cron heartbeats, and a Stop hook that blocks the exit until the stated condition holds. Both posts sell as new what your /goal hook enforced today.
Three separate posts in one day's feed on the same idea. Brandon-Vault plus /ask with local nomic-embed-text is further along than any of them, since yours returns file-path citations rather than a summary.
Roles, a boss, goals, and a budget per agent, sitting above Claude Code and Cursor. Adopting it would break your one-active-session rule and your no-fanout rule, both of which exist because your weekly allowance is finite. Listed for completeness, and I would leave it.
Same conflict. The one part that fits your rules is the delegation section, since a single delegated pass costs less than a fan-out.
Every row fetched from the GitHub API during this session. Star counts are as of 13 August 2026.
| Repo | Stars | What it is | Named by |
|---|---|---|---|
obra/superpowers | 271,717 | agentic skills framework and development methodology | @charliejhills |
multica-ai/andrej-karpathy-skills | 202,161 | Claude Code guidance repository, 8 top-level entries: CLAUDE.md, CURSOR.md, EXAMPLES.md, two READMEs, and skills, .claude-plugin, .cursor directories | @charliejhills |
paperclipai/paperclip | 78,034 | manage agents at work | @gippp69 |
shareAI-lab/learn-claude-code | 74,135 | a nano Claude Code style agent harness, built from zero | @charliejhills |
mem0ai/mem0 | 63,209 | universal memory layer for agents | @N01ennn |
repowise-dev/repowise | 5,728 | codebase intelligence, health scores, generated docs | @ai_explorer25 |
guillaumemeyer/watermarks-remover | 5,177 | strip multi-vendor AI provenance marks, Unicode hygiene | @josesilesdata |
ussumant/llm-wiki-compiler | 313 | Claude Code plugin compiling markdown into a wiki | @0xkkai |
ekadetov/llm-wiki | 107 | persistent knowledge bases in Obsidian | @0xkkai |
praneybehl/llm-wiki-plugin | 86 | Karpathy's LLM wiki pattern as a skill and plugin | @0xkkai |
zhayujie/CowAgent at 46,502 and Hmbown/CodeWhale at 40,748, and neither is confirmed as the one meant.A large share of this feed is written by engagement accounts using the same headline grammar, all caps, a leak claim, a star count. The technique underneath is often real. The attribution rarely survives a check. Everything above separates those two.
scan_invisible() and strip_invisible() are in anti_ai_voice_gate.py, measured at 1 real hit and 0 false positives across 1,041 pages. Remaining work is the one live page it caught, tame-impala-charts.python3 ~/.claude/skills/_shared/ledger_verify.py <ledger> then recomputes each published figure and exits non-zero on any disagreement, duplicate id, malformed hash, or missing field. This rule exists because the first version of this page published 56 with nothing on disk that could reproduce it.