Overview
This report answers: which recurring natural-language asks or plain-text commands should be collected into reusable skills, and which pack should own each one.
| Source | Records | Sessions | Plain-text records |
|---|---|---|---|
| Claude | 9,804 | 2,943 | 6,722 |
| Codex | 1,905 | 595 | 1,863 |
The parsed date range is 2025-12-13T17:10:26Z through 2026-05-30T02:20:54Z. Top projects by message volume include metternich-engine, bismarck-ai, poke/monorepo, loadoutworks, and this agentic-skills repository.
Short answer: yes. The highest-value natural-language skill surfaces are plan execution, shipping, staging deploy validation, next-step status, agent instruction updates, test failure fixing, review remediation, and skill visibility repair.
Method
I parsed the full available compact history files at ~/.claude/history.jsonl and ~/.codex/history.jsonl. For Codex, I also scanned ~/.codex/sessions/**/*.jsonl for session metadata such as cwd, model, provider, and session id. I used the compact Codex history as the prompt-count source to avoid double counting rich rollout records.
Plain-text records exclude prompts that start with / or $. The analysis then combined exact normalized repeats, keyword clusters, project/source breakdowns, and manual review of representative examples. Counts labeled as "classifier" are evidence for magnitude, not exact semantic truth, because keywords such as "current" and "overlap" can be noisy.
Candidate Skills
| Rank | Potential Skill | Pack | Frequency Evidence | Recommendation |
|---|---|---|---|---|
| 1 | plan-implementation-runner |
exec-loop |
541 history matches for "Implement the following plan", "Implement the plan", or Codex continuation plan prompts. | Create as an explicit plain-text router or strengthen exec to claim pasted-plan implementation. |
| 2 | plain-text-ship |
exec-loop primary, gitops secondary |
580 classifier hits. Exact repeats include 167 "update docs, commit and push", 74 "commit this", 38 "commit this and push", and 25 "commit and push". | Prefer an alias/router over a brand-new workflow, because ship already owns most behavior. |
| 3 | staging-deploy-smoke |
release-ops |
127 classifier hits, including 48 exact "deploy to staging" records. | Extend deploy or add a staging-focused wrapper that builds, deploys, smoke-tests, and reports URLs. |
| 4 | what-now |
exec-loop or repo-maintenance |
109 status/next-step classifier hits. Exact repeats include 62 "continue", 33 "please continue", and 42 "what's next" variants. | Make a user-friendly status router over task docs, git state, recent history, and available packs. |
| 5 | agent-instructions-update |
agent-work-admin |
125 classifier hits around CLAUDE.md, AGENTS.md, progress notes, and instruction updates. |
Create a disciplined updater for durable agent conventions, progress notes, and lessons capture. |
| 6 | test-failure-fixer |
code-debug |
54 classifier hits around failing tests, test errors, and "fix tests" prompts. | Add a focused debugging skill that starts from failed validation output and ends with passing targeted tests. |
| 7 | review-fix-runner |
code-review |
45 classifier hits around taking review findings, issues, or audit reports and implementing fixes. | Add a remediation runner paired with expert-review, regression-check, and review report artifacts. |
| 8 | visual-polish-pass |
website-polish |
40 classifier hits around screenshots, layout issues, mobile/desktop checks, overlap, and visual polish. | Broaden the currently asset-focused website polish pack with a real UI screenshot/viewport pass. |
| 9 | task-doc-sync |
agent-work-admin |
33 classifier hits for updating task docs, marking complete, documenting results, and syncing roadmap/todo/history. | Prefer as a subroutine in exec and ship, or expose as a utility skill for cleanup-only turns. |
| 10 | skill-visibility-repair |
skill-dev |
31 classifier hits, concentrated in agentic-skills and claude-skills, about missing skills, pack install, refresh, and visibility. |
Add an operational repair skill for pack install status, managed copies, session reload caveats, and missing command diagnosis. |
| 11 | showcase-data-refresh |
website-polish or docs-health |
21 classifier hits around Skills Showcase, generated data, and validation refreshes after skill metadata changes. | Useful as a repo-local or pack-local utility because it bundles generated data, validation, and stale-row checks. |
| 12 | open-alignment-preview |
alignment-page-admin |
6 classifier hits, but recent examples show repeated WSL/PowerShell browser-open friction. | Small utility skill or script wrapper; high convenience despite low count. |
Evidence Matrix
| Claim | Evidence | Inference | Confidence | Decision Impact |
|---|---|---|---|---|
| Shipping is the largest plain-text automation gap. | Exact repeats: "update docs, commit and push" 167, "commit this" 74, "commit this and push" 38, "commit and push" 25. Classifier total: 580. | The workflow is stable: inspect diff, update docs if needed, commit, push, and report status. | High | Make plain-language shipping route reliably to ship or a thin alias. |
| Pasted plan execution deserves a first-class route. | History contains 541 matches for "Implement the following plan", "Implement the plan", or Codex continuation plans. | The user repeatedly gives a plan as the source of intent and expects implementation, validation, and shipping. | High | Clarify exec-loop ownership for pasted plans, especially after context transfers. |
| Staging deploys are frequent enough for a deployment submode. | 127 staging/build/deploy classifier hits; 48 exact "deploy to staging" records. | Deployment asks are not only "run deploy"; they often include build, smoke validation, URLs, and follow-up test instructions. | High | Extend release-ops/deploy with a staging-smoke contract or add a wrapper. |
| "What next?" and "continue" should resolve to a status router. | 109 status/next-step hits; exact repeats include "continue" 62, "please continue" 33, and "what's next" variants 42. | The agent should infer from task docs, roadmap, git state, and current context instead of asking the user to restate work. | High | Create or alias what-now to reduce repeated plain-text status prompts. |
| Agent instruction updates are recurring and risky enough for structure. | 125 classifier hits around CLAUDE.md, AGENTS.md, progress notes, and instruction updates. |
These edits affect future agent behavior, so they need provenance, lesson capture, and validation. | Medium high | Place in agent-work-admin, with provision-agentic-config awareness. |
| Skill visibility repair is a real agentic-skills maintenance lane. | 31 classifier hits, mostly in agentic-skills and claude-skills, plus recent task history around pack install and discovery repairs. |
Missing skill symptoms involve pack metadata, managed installs, current-session reload caveats, and runner differences. | Medium high | Put in skill-dev as an operational repair skill rather than a skill creation workflow. |
| Visual polish is present but the count is noisy. | 40 classifier hits around visual checks and screenshots, but some "overlap" matches were semantic false positives. | The workflow exists, but it needs a stricter parser before deciding whether to build a general skill. | Medium | Candidate for website-polish, not top priority until manually triaged. |
| Research freshness is cross-cutting but not clearly a standalone skill. | 203 loose classifier hits, but terms like "current" often referred to UI state or current work, not web freshness. | The behavior belongs in standing instructions and research skills more than in a standalone user-facing skill. | Low as counted | Do not build from this count alone; tighten existing research-admin contracts instead. |
Highest Impact
plain-text-shipinexec-loop: highest avoided manual prompts because shipping phrases appear constantly and already have stable behavior.plan-implementation-runnerinexec-loop: captures pasted-plan implementation and continuation plans, which are common after context transfer.staging-deploy-smokeinrelease-ops: staging deploys recur and often need smoke validation, not just deployment command execution.what-nowinexec-looporrepo-maintenance: converts "continue" and "what's next" into an evidence-backed next action.agent-instructions-updateinagent-work-admin: reduces risky ad hoc edits to persistent agent behavior files.
Lower Confidence Or Not New
| Pattern | Disposition | Reason |
|---|---|---|
freshness-verify |
Standing instruction or research-admin contract, not a priority standalone skill. | The keyword count is noisy because "current" often means current UI state or current task. |
benchmark-batch-runner |
Possible agentic-skills-bench enhancement, but not proven by compact plain-text history alone. |
Much benchmark work is already done through explicit skill commands, task docs, or continuation plans. |
repo-contributor-guide |
Do not create yet. | The canonical Codex "Generate a file named AGENTS.md" prompt appears, but only as a low-frequency explicit request. |
analyze-sessions templates |
Enhance existing analyze-sessions, not a new skill. |
Session analysis is already owned by session-analytics; add report templates if needed. |
Confidence And Assumption Register
| Item | Status | What Could Change It |
|---|---|---|
| Shipping, pasted-plan execution, staging deploys, and next-step status are the top priorities. | Evidence-backed | Unlikely to change unless rich-session dedupe removes a large number of compact prompts. |
Pack ownership for what-now is split between exec-loop and repo-maintenance. | Inference | Decide whether it should mutate or remain read-only. |
agent-instructions-update belongs in agent-work-admin. | Inference | Could move to repo-maintenance if scoped to AGENTS/CLAUDE file mechanics rather than workflow governance. |
| Classifier counts are exact enough to rank broad patterns. | Qualified | A semantic embedding pass or manual audit could refine medium and low-confidence rows. |
| Codex compact history is the correct prompt-count source. | Method choice | Would change if rollout user records contain missing prompts not present in compact history. |
Alternatives Considered
- Do nothing: rejected for high-frequency patterns because the same phrases keep recurring across projects and sources.
- Make every pattern a new skill: rejected because several should be aliases, subroutines, or existing skill enhancements.
- Route all plain text to
exec: too coarse for deploys, review remediation, and instruction updates where pack-specific gates matter. - Use only standing instructions: useful for freshness and task-doc habits, but insufficient for staged deploys, skill visibility repair, and review fixes.
Proposed File Changes If Approved Later
No skill files should change before approval of this report. Likely follow-up work would add or update skills under packs/exec-loop, packs/release-ops, packs/agent-work-admin, packs/code-debug, packs/code-review, packs/skill-dev, and possibly packs/website-polish. Each changed SKILL.md would require archive, version, changelog, focused tests, and generated showcase refresh when metadata changes.
Review Gates
Evidence Coverage Gate
Is the evidence sufficient to use this ranked list as the basis for follow-up skill work?
Candidate Verdict Gate
Which candidate set should move forward first?
Pack Ownership Gate
Do the proposed pack assignments look right?
Artifact Destination Gate
Where should the durable result of this analysis live?
Coverage Checkpoint Gate
Is the current source coverage acceptable?
Compile Answers
Answer every required gate question, then compile the YAML and send it back for follow-up implementation decisions.