Releases: sananthanarayan/skilldrop
Release list
v0.11.3 — enforce README coverage, and finish the RFC status pass (0.11.3)
Two guards and a bookkeeping fix.
validate.py already checked that README's skill count matched disk. A count is not coverage —
it can pass while a skill, pack or agent has no documentation at all, because it checks the
number rather than the thing the number describes. Now every skill on disk needs a README row,
every pack in packs.json needs documenting, and every agent in agents/ needs a row. Verified by
breaking a row on purpose and watching it fail.
That guard immediately found a real gap: security-reviewer has been shipping undocumented since
it was added. Now in the reviewer agents table.
AGENTS.md says mark an RFC accepted before building and implemented when it merges. Eleven
had shipped without the second half ever happening, so the status field had quietly stopped
meaning anything. Now 24 implemented, 2 draft — the template, and 0010, which is honestly unbuilt.
npx skilldrop-cli@0.11.3 install --pack solution-architectv0.9.3 — add agent-adoption-stage (RFC-0024) and release 0.9.3
RFC-0023's four skills assess organisational preconditions. None answers what an
engineering leader actually asks once tools are in hand: how far has our practice got,
and what is the single next unlock?
agent-adoption-stage places a team on a 0-4 ladder (gated / assisted ~1 agent /
parallel ~10 / supervised autonomy ~100 / intent-steered ~1000+) from observables —
agents in flight, who writes the code, what still gets reviewed — never from licence
counts. The load-bearing idea is that the bottleneck MOVES: attention -> review
throughput -> trust in the loop -> finding work to automate, so the fix that unlocked
the last step does nothing for the next.
It forces exactly one stage (a range is a refusal) and exactly one unlock (ten is a
backlog), places on the median engineer rather than the power user, names the guardrail
to RETIRE as well as the one to add (reading every diff is right at stage 1 and
arithmetically impossible at stage 3), and refuses the named trap of scaling agent
count before the loop has earned trust.
Adapted from Boris Cherny's "Steps of AI Adoption" (2026-07-16) with two required
changes: stated as capabilities rather than product names (skilldrop's routing layer is
provider-neutral by design, and a SKU list dates within a quarter), and the source is
credited rather than absorbed.
Its real value is routing: each unlock points at the skilldrop skill that already
implements it — pre-merge-review + gate.py for the trusted verification loop, the
reviewer panel for automated review, subagent-design for parallel isolation,
agent-loop-design for loops, agent-budget for token efficiency. skilldrop already
shipped the rungs; this adds the map.
ai-readiness-assessment gains an explicit cross-reference so the two axes stay distinct
(a team can be governance-ready and stuck at stage 1). README 55 -> 56, og.png
re-rendered by the count guard.
npx skilldrop-cli@0.9.3 install --pack solution-architectv0.9.2 — add four AI-adoption skills (RFC-0023) and release 0.9.2
The ai-engineering pack was strong on building AI systems (agent-loop-design,
subagent-design, agent-budget, agent-threat-model, llm-eval-harness) and on measuring
usage after the fact (ai-usage-report), but had nothing for the stage before any of
that — deciding whether, where and how to adopt at all. Four artifacts no existing
skill produces:
- ai-readiness-assessment — six fixed dimensions scored 0-4, each with an evidence
line or [insufficient evidence]; gaps ranked by what blocks first, not by lowest
score; a posture sentence that takes a position. - ai-use-case-triage — value x feasibility x risk with the weights shown; candidates
restated as a workflow and a person; a first slice with an observable kill condition;
a mandatory not-yet list where each rejection carries its promotion condition. - ai-adoption-rollout — the human counterpart to migration-plan: cohorts smallest
first, per-cohort enablement, champions with protected hours, an adoption gate per
wave (observable + threshold + bake time + rollback), and a comms plan that must
state what will NOT change about headcount and evaluation. - ai-usage-policy — three data tiers with examples from the org, prohibitions that
always name a permitted alternative, a review matrix keyed to consequence rather
than to model names, and an exception route with a turnaround time.
All standard tier, all in ai-engineering (triage also in product-manager), each with
evals whose no-trigger rows point at the siblings they collide with. Registered in
model-routing.json, packs.json and the README (51 -> 55).
Declined: a fifth AI-ROI skill. It fails the fit check's own duplication test —
business-case prices the investment, success-metrics defines the measure and decision
rule, ai-usage-report reports actual usage. A wrapper over those three would be filler.
Recorded in RFC-0023 so the idea leaves a trace.
The og.svg and README count guards both caught the stale 51 — og.png re-rendered.
npx skilldrop-cli@0.9.2 install --pack solution-architectv0.9.1 — backfill evals for the sonar pair + ops lifecycle (32 -> 38 of 51)
Next targeted round (#4b), same discipline.
- sonar pair (sonar-onboard, sonar-review): a clean two-way collision — scaffolding
Sonar into a repo vs. running a scan and reporting the gate — each also bounded
against devils-advocate and pre-merge-review. - ops lifecycle (runbook-generator, incident-comms, postmortem-generator,
observability-plan): before / during / after, plus the subtler runbook-vs-observability
boundary — one responds to alerts, the other designs them. postmortem is also
bounded against decision-log, since both ingest a raw Slack thread.
Assertions come from each skill's own Quality bar — gate is the source of truth, findings
scoped to changed files, roles not names in causal sentences, next-update time on every
incident message, SLIs as good-over-valid events never averages.
Rides in the unpublished 0.9.1.
npx skilldrop-cli@0.9.1 install --pack solution-architectv0.9.0 — release 0.9.0 — Claude 5 model map, Agent Skills standard, review-panel install
Minor bump (new --panel CLI flag). Ships to users:
- #1 model-routing refreshed to the Claude 5 family (claude-sonnet-5 / claude-opus-5;
light stays claude-haiku-4-5). - #2 README declares conformance to the Agent Skills open standard (agentskills.io).
- #6
skilldrop install --panel review— one-command install of the 3 reviewer
subagents + the pre-merge-review orchestrator, with per-tool parallel dispatch
documented (RFC-0020).
Regenerates the version-sourced .claude-plugin manifests to match.
npx skilldrop-cli@0.9.0 install --pack solution-architectv0.8.4 — release 0.8.4 — opt-in blocking gate hook (RFC-0019)
Makes pre-merge-review's gate enforceable as a portable harness without skilldrop
becoming one: enforcement moves into the substrate every tool shares (git + CI),
not the skill's prose.
- gate.py --install-hook / --uninstall-hook: writes a marker-fenced, worktree-safe
blocking git pre-commit hook that runs the gate and exit-1s on RED, so git blocks
the commit for any tool (Claude, Cursor, Codex, Aider). Bypassable with
--no-verify; CI required-checks are the un-bypassable backstop. Functionally
tested: RED blocks, --no-verify bypasses, uninstall is clean, GREEN commits. - pre-merge-review gains an "Enforce as a harness (opt-in)" section: Layer 1 git
hook, Layer 2 CI required checks, Layer 3 per-tool native (Claude Code hook,
Aider --test-cmd). - RFC-0019: narrows the opt-in exception to RFC-0006's reminder-only hooks —
a deterministic gate installed by explicit command, never on a plain install.
npx skilldrop-cli@0.8.4 install --pack solution-architectv0.8.3 — release 0.8.3 — ship the security-reviewer agent + review-panel loop
Bumps skilldrop-cli 0.8.2 -> 0.8.3 so users get the new security-reviewer agent
(skilldrop install --agent security-reviewer) and feature-implement-loop's
three-reviewer panel (RFC-0017) on their next update. Regenerates the version-sourced
.claude-plugin manifests to match.
npx skilldrop-cli@0.8.3 install --pack solution-architectv0.8.2 — release 0.8.2 — ship RFC-0015/0016 checks + heavy-tier oracles
Bumps skilldrop-cli 0.8.1 -> 0.8.2 so users get the 5 new worked-example oracles
(devils-advocate, doc-critique, tech-comparison-matrix, strategy-analysis,
business-case) on their next skilldrop update. Regenerates the version-sourced
.claude-plugin manifests to match.
npx skilldrop-cli@0.8.2 install --pack solution-architectv0.8.1 — close RFC-0009: ship agents-md-generator — v0.8.1
All three open questions settled and the skill built.
- Audit stays in this skill rather than becoming a doc-critique rubric.
The audit's value is verifying commands against the repo; doc-critique
takes documents, so it would have to trust every command on faith —
the one thing this skill exists not to do. - AGENTS.md always; satellites opt-in and delta-only. A satellite
carries only what is specific to its tool plus a pointer for the rest
— the shape this repo's own CLAUDE.md already has. Copying shared
content into a satellite is refused: four copies of a command list is
four chances to be wrong, and the drift is invisible. - The ~150-line cap is hard on generate, a warning on audit. A
generated file has no excuse; an existing one may be long for good
reason, as this repo's AGENTS.md is.
The skill is built around the evidence rule: every command traces to a
CI step, Makefile target or manifest key, or it appears as
[missing: ...] rather than a guess. A dependency is not a command —
jest in devDependencies does not mean npm test exists. Convention lines
must pass a counterfactual test, and virtue instructions are stripped
rather than softened, with a strip-list in reference.md.
50 skills.
npx skilldrop-cli@0.8.1 install --pack solution-architectv0.8.0 — add the Antigravity agent target — v0.8.0; no target refuses now
RFC-0013 is rewritten, not amended: its premise was false. It designed a
plugin-bundle emitter on the belief that Antigravity had no agents
directory. Its own docs list two — ~/.gemini/config/agents/.md
global, .agents/agents/.md workspace — with plugin bundles as a
third option, not the only one. Keeping that draft would have sent
someone down a bundle build to deliver two markdown files.
That is three for three. Kiro, Codex and Antigravity were each filed as
blocked on an unconfirmed fact; all three facts were published and I had
not looked. The survey generalised from the first layout it found each
time. Recorded in the RFC so the pattern is visible rather than repeated.
The emitter is markdown with rewritten frontmatter: subagent: true is
added unconditionally (without it the agent exists but invoke_subagent
cannot reach it — a silent failure), tools map Read/Grep/Bash to
view_file/grep_search/run_command, and model passes through when it is
one of inherit|flash|pro, which the shipped agents already are.
Glob is named and dropped, not mapped to list_dir. They are different
operations and a wrong tool is invisible where a missing one is not.
Frontmatter verified well-formed: quoted scalars parse, embedded quotes
and em-dashes survive, body is the system prompt, nothing leaks across
the delimiter.
Every surveyed tool with an agent format now installs by command. Only
Cursor is absent, and only because it has no such format.
npx skilldrop-cli@0.8.0 install --pack solution-architect