AFPS Workflow Patterns review
Cross-session analysis of how agentic-skills (AFPS) workflows are used — activity trends, skill chains, reactive vs. proactive balance, and automation opportunities.
Generated 2026-06-09 • Source: Claude history (633 messages / 274 sessions) • Codex: 0 direct AFPS sessions (skills consumed from downstream projects)
Overview Stats
Copied!
633
Claude Messages
274
Claude Sessions
65 days
Active Period
~9.7/day
Avg Messages
Source comparison: All AFPS work runs through Claude Code. Zero Codex sessions target the agentic-skills CWD directly — Codex consumes AFPS skills from downstream project directories (2,849 Codex prompts reference skill-related terms). The AFPS repo is a Claude-first development environment.
Token & cost: Claude history.jsonl does not include token usage or cost metadata. Cost is unavailable from this data source.
Monthly Activity Trend
Copied!
Claim: AFPS activity is accelerating — June already exceeds full-month totals in 9 days (277 vs. 221 in May), driven by card-pack animation debugging and convention hardening.
Evidence: Monthly message counts from Claude history.jsonl. June top skills: /investigate(29), /ship-end(10), /pack(5), /exec(5).
Inference: The June spike is investigation-heavy — reactive work on card-pack animations is pulling volume up. Proactive convention work (excalidraw, alignment pages, narration) also contributes.
Skill Usage Breakdown
Copied!
| Skill | Count | % of Skill Calls | Role |
|---|---|---|---|
| /ship | 43 | 22% | Primary shipping workflow |
| /investigate | 33 | 17% | Bug hunting & root-cause analysis |
| /analyze-sessions | 22 | 11% | Meta — understanding own workflow |
| /ship-end | 21 | 11% | Session wrap-up & push |
| /run | 9 | 5% | Dev server & preview |
| /pack | 8 | 4% | Pack management |
| /exec | 5 | 3% | Plan step execution |
| /reconcile-dev-docs | 4 | 2% | Doc drift correction |
| /targeted-skill-builder | 3 | 2% | Skill creation from gap |
| /devtool-docs-audit | 3 | 2% | Docs quality audit |
| Others (20 skills) | 42 | 21% | Long tail — 1-2 uses each |
Claim: The AFPS workflow is dominated by a ship→investigate→analyze cycle. /ship and /ship-end together account for 33% of all skill calls — shipping is the most frequent structured activity.
Evidence: 43 /ship + 21 /ship-end = 64 shipping calls out of ~193 total skill invocations (excluding /clear and /resume).
Session Type Classification
Copied!
Claim: Sessions are short and focused. 81% (222/274) are 1-3 messages. Only 3% (8) exceed 7 messages. The workflow pattern is: open session → do one thing → ship or close.
Evidence: 141 sessions with 1 message, 81 with 2-3, 45 with 4-7, 7 with 8-15, 1 with 16+.
Inference: This is a rapid-fire cadence — not deep multi-turn exploration. You treat Claude Code sessions as focused task slots, with /ship or /ship-end as the closing gate.
Workflow Sequences & Chains
Copied!
Most sessions use 0-1 skills — chaining is rare because sessions are short. Observed multi-skill chains:
| Transition | Count | Pattern |
|---|---|---|
| /analyze-sessions → /ship-end | 2 | Analyze then close the session |
| /analyze-sessions → /analyze-sessions | 2 | Iterative refinement of analysis |
| /investigate → /analyze-sessions | 1 | Bug → check if pattern is recurring |
| /analyze-sessions → /targeted-skill-builder | 1 | Find gap → build skill for it |
| /session-triage → /targeted-skill-builder | 1 | Diagnose incident → build prevention |
| /benchmark-test-skill → /benchmark-agent-review | 1 | Test then review benchmark |
| /spec-interview → /roadmap | 1 | Spec → plan |
| /spec-drift → /reconcile-dev-docs | 1 | Detect drift → fix docs |
Claim: The AFPS workflow is implicitly sequential but not explicitly chained. You rely on skill recommendations ("Recommended next command") rather than pre-defined pipelines. The routing works but is manual — you choose the next step each time.
Recurring Themes
Copied!
| Theme | Messages | Example Prompts |
|---|---|---|
| Skill creation/editing | 113 | "verify all codex skills say not to ship to feature branch", "why do we have both agentic-skills and claude-skills" |
| Pack management | 43 | "which workshops would be of value", "they should live in the creator media pack" |
| Planning/roadmap | 37 | "plan-phases should be renamed plan-phase", "gap once done with plan interview" |
| Alignment pages | 36 | "create test environment for alignment-loop pack", "html is perfect, not asking for template" |
| Investigation/debug | 29 | "card sinks back into the card pack", "long pause on drawer close" |
| Docs/convention | 25 | "should that be in claude/agents md for all cases", "excalidraw convention for agents" |
| Shipping/deploy | 18 | "verify all codex skills", "create skill that wraps manual blockers" |
| Session analysis | 10 | "did we not have a conversation about this before?", "why is that not in the history?" |
| Benchmarking | 8 | "update benchmarks and add test setups for new skills" |
| Skillmap/excalidraw | 6 | "excalidraw graphic explaining packs and skills", "handwritten styling for borders only" |
Repeated prompts
- "can you run the install script?" — 3 occurrences
- "can you run the dev server?" — 3 occurrences
- "the sync skill should flag if the version number..." — 2 occurrences
- "can we add to the alignment page that we can flag..." — 2 occurrences
Reactive vs. Proactive Balance
Copied!
58%
Proactive (create/build/plan)
41%
Reactive (investigate/fix/debug)
Claim: Overall balance is healthy — more proactive than reactive. However, June has flipped: investigation dominates (29 /investigate calls in 9 days vs. 33 total in 65 days). The card-pack animation work created a reactive debt spiral.
June Investigation Spike
Copied!
88% of all /investigate calls (29/33) occurred in June. The dominant driver was card-pack animation debugging — a sustained multi-session bug chase from Jun 3–5:
- "animation performance regression" (Jun 3)
- "long pause on drawer close after cards consolidate" (Jun 4)
- "card sinks back into the card pack" (Jun 4)
- "two distinctly animating cards on drawer close" (Jun 4)
- "card hovering above card pack in final step" (Jun 4)
- "pulling card out of pack inconsistently triggers drawer" (Jun 5)
Claim: The card-pack animation work created a debugging marathon. Each /investigate found a fix but revealed the next visual bug — a cascading chain of 12+ sequential investigations over 2 days. This is the single largest reactive time sink in AFPS history.
Inference: Visual animation work is uniquely hard for the current AFPS workflow — each bug requires human visual verification, the agent can't see the result, and fixes cascade. A visual regression test or screenshot diffing step could break this pattern.
Evidence Matrix
Copied!
| Claim | Source | Evidence | Confidence | Assumption |
|---|---|---|---|---|
| Activity accelerating | Claude history.jsonl | 135 → 221 → 277 (9d) monthly messages | High | June pace sustained (31 msgs/day) |
| Ship-dominated workflow | Claude history.jsonl | 64 shipping skill calls (33% of total) | High | None |
| Short focused sessions | Claude history.jsonl | 81% of sessions ≤ 3 messages | High | None |
| Claude-first for AFPS | Codex rollouts | 0 Codex sessions in agentic-skills CWD | High | Codex cwd metadata is complete |
| Animation reactive spiral | Claude history.jsonl | 12+ sequential /investigate calls Jun 3–5 on card animations | High | None |
| Manual skill routing | Skill transition data | Max chain frequency is 2; most sessions single-skill | Medium | Some chains lost across session boundaries |
| Convention maturation phase | Recent prompts | Excalidraw convention, narration, alignment page improvements all in last 3 days | Medium | Conventions will stabilize |
Confidence & Assumptions
Copied!
- Evidence-backed: Message counts, skill counts, session lengths, monthly trends, session type distribution.
- Provisional: Reactive/proactive classification (keyword-based, may misclassify some messages). Session type categories overlap when a session touches multiple themes.
- Missing: Token usage and cost data (not in Claude history.jsonl). Cross-session workflow chains (a skill recommended at the end of session A may be invoked in session B — this linkage is lost). Codex AFPS work is invisible from CWD matching alone.
- Would change conclusions: If token/cost data showed that /investigate sessions consume 5× more tokens than /ship sessions, the reactive cost would be even more significant than message counts suggest.
Recommendations
Copied!
| Pattern | Frequency | Type | Suggestion | Avoided Manual Prompts |
|---|---|---|---|---|
| "run install script" / "run dev server" | 6 | Standing instruction | Add to CLAUDE.md: auto-detect when install/dev-server is needed and run without prompting | ~6/month |
| Animation visual bug cascade | 12+ in 2 days | Skill / convention | Screenshot-diff verification step in /investigate for HTML prototype work; break reactive spiral | ~8-10 per visual feature |
| Session-end shipping | 64 calls | Standing instruction | Already covered by shipping contract — working well | N/A (working) |
| Analyze → build skill | 3 chains | Skill routing | Working as designed: /analyze-sessions → /targeted-skill-builder pipeline is functional | N/A (working) |
| Convention creation from ad-hoc decisions | ~8 | Convention | Gate in skill-building flow: "is this a one-off or a convention?" — if convention, route to docs | ~3-4/month |
Top 5 by impact
- Visual verification for prototypes — Screenshot/diff step in /investigate for HTML work. Breaks the "fix → new visual bug" spiral. Est. 8-10 prompts avoided per visual feature.
- Auto install/dev-server detection — Standing instruction to run install/dev without asking. Est. 6 prompts/month avoided.
- Convention-detection gate — When a decision looks reusable, auto-suggest writing a convention doc. Est. 3-4/month.
- Cross-session chain tracking — Persist "Recommended next command" across sessions so the next session picks up where the last left off. Reduces context loss.
- Codex AFPS visibility — Tag Codex sessions consuming AFPS skills with origin pack metadata so /analyze-sessions can track downstream adoption.
Approval Gates
Copied!
Gate: Evidence Coverage
Is the evidence sufficient for these workflow pattern conclusions?
Gate: Scope & Non-Goals
Does this analysis cover the AFPS workflow areas you wanted to understand?
Gate: Recommendation Priority
Which recommendation should be acted on first?
Gate: Artifact Destination & Proposed File Changes
Approve the output location for this analysis report.
Compile Section
4 required questions unanswered
Copied! Copied!