§ THE DRIFT
Nine months ago, Anthropic published a format for teaching an AI agent a procedure: a folder, a file called SKILL.md, two required fields, and a body of Markdown. Microsoft wired it into VS Code within days. OpenAI added it to Codex. By this summer roughly forty products read the same files — Cursor, Copilot, Gemini CLI, Kiro, Goose, even Databricks and Snowflake. By March there were more than 490,000 skills across the major marketplaces.
I went looking through them for the ones that do my old job.
There are skills for writing commit messages, generating migration scripts, reviewing pull requests, drafting PRDs, running postmortems. Excellent ones. Some with tens of thousands of installs. What I couldn't find was one that asks whether the thing your status report calls Done is actually done.
That absence isn't random. It's the same absence PulseBoard exists because of. Nearly every popular skill operates inside one tool, on one artifact, for one person — write this file, review this diff, generate this doc. The work I spent fifteen years doing lives between the tools: the status doc says one thing, Jira says another, GitHub says a third, and someone has to sit down on a Friday and decide which is true.
The tooling world has spent nine months making agents better at producing work. It has spent almost none making them better at reconciling what the work actually says.
There's a second gap underneath the first. The benchmarks that arrived this summer are unkind: SkillsBench, the first peer-reviewed measurement of the format, found curated skills raise task pass rates from roughly 34% to 50% — real, but nowhere near the confidence with which skills get marketed. Snyk's audit of nearly 4,000 skills found more than a third carried at least one security flaw. A broader scan of 22,511 skills turned up over 140,000 issues.
So the ecosystem is now large, standardized, portable — and almost entirely unverified. A skill can have flawless frontmatter, a clean description, perfect structure, and still confidently tell you something false.
That's a familiar shape to anyone who has run a program. It's what a green dashboard is: well-formatted, widely trusted, structurally valid, and wrong.
So when I packaged my own reconciliation checks as skills this week, I built them to be scored. They ship with a synthetic eight-feature program, nine planted problems, and an answer key. Run them, compare, decide whether to trust them. That should be the default and it isn't.
§ THREE SIGNALS
1. Agent Plugins 1.0 lands — August 6
A vendor-neutral package format carrying skills and MCP servers in one directory, supported by VS Code, Cursor, Copilot, ChatGPT/Codex and Kiro. The spec explicitly puts trust and permissions out of scope. Portability is solved; whether the thing inside is any good is now formally somebody else's problem.
2. SkillsBench publishes the first paired benchmark
Curated skills lift average pass rates roughly 16 points, with per-configuration gains ranging from about 4 to 26. The spread matters more than the average — skill quality varies enormously, and until now nobody was measuring it.
3. Snyk's ToxicSkills audit finds flaws in over a third of skills scanned
Prompt injection in 36% of nearly 4,000 skills. The format's great virtue — anyone can write one in an afternoon — is also why the median skill is unaudited. Treat an installed skill like an installed dependency, because that's what it is.
§ DRIFT OF THE WEEK
A feature reported Done in the weekly status. Ticket closed, PR merged, one line in the doc: code complete.
Eight things have to be true before that word is accurate:
It covers what the PRD asked for
Every ticket is closed
The code is merged — to main, and not reverted
It's been tested since the last merge
It's in production
The serious bugs are closed
A human has walked the flow end to end
The teams downstream have been told
This one passed three. The PR had merged but never deployed. QA was still in progress. A Sev-2 had been filed the day before the status went out. And a second team couldn't test their feature, because it depended on this one being live — a dependency that appeared in neither PRD, discovered by a tester who happened to be the domain expert.
Nobody lied. The person writing the status saw a merged PR and wrote what their template called Done.
Three shapes recur, and only the first is the one people expect:
Reported Done, isn't. Usually check 3 or 5 — merged to a branch that never reached main, or merged and never deployed.
Reported on track, quietly stalled. A PR open longer than a sprint. Work is happening, so nobody re-reports, and the date moves once at the end instead of a little each week.
Reported at risk, already finished. Every check passes. The sign-off never reached the person writing the status, and a program carries a risk that closed a week ago.
Drift runs both directions. The pessimistic kind costs attention instead of dates, which is why it never gets caught.
Seen a drift like this? Reply to this email — best submission gets featured (anonymized) in a future issue.
§ THE RECONCILIATION
Take one feature currently reported as Done. Walk all eight checks. Open the deploy log, not the ticket. Ten minutes.
If you'd rather not do it by hand, the checks are now four free skills:
npx skills add pulseboard-ai/execution-drift-skills --skill reconcile-statusreconcile-status runs the eight checks per feature and shows the ticket, PR or message behind every answer. release-readiness compares a release against what was actually committed at planning — including the hotfix that rode along without anyone being told. reconcile-dependencies finds the dependencies nobody declared, read out of what your tickets and threads actually reference. resolve-blocker takes one problem and writes the memo: what's missing, how this was solved last time, who decides, two or three real options with what each costs, a deadline, and a default if nobody answers.
Three rules I'd defend to anyone who forks them. No scores — no percent-ready, no traffic lights; a verdict people can't audit is one they'll discount the first time it disagrees with them. Every answer cites its source — a finding without a ticket number or timestamp is a feeling with formatting. Manual means manual — two of the eight checks can't be verified from data, so the skill says so and names the check to run instead of guessing.
MIT, no network calls, no telemetry, free permanently.
Install it, run it against the sample, score it against the answer key before you point it at anything real. If it catches something in your program, reply and tell me what — that's the next issue.
In Case You Missed It
Know someone drowning in status reconciliation?
Execution Drift grows by referral — forward this email or share your link Here!
Stay coherent, M3 Mythreyi Chandoor · Building PulseBoard AI — the execution intelligence layer for engineering orgs · www.linkedin.com/in/mythreyic