Implement
Apply evidence-based implementation rhythms—read/search/edit ratios, verification cadence, and debugging spiral guards—from large-scale tracked agent sessions.
Overview
Implement is an agent skill most often used in Build (also Ship testing, Operate iterate) that applies quantitative implementation and verification benchmarks from large tracked coding sessions.
Install
npx skills add https://github.com/hyperb1iss/hyperskills --skill implementWhat is this skill?
- Quantitative benchmarks from 21,321 tracked operations across 64+ projects (Sibyl and related studies)
- Sweet spot: 2–3 edits between verifications; ~48.7 changes and ~23.2 verifications per commit in sampled data
- Typecheck:Lint:Test ratio guidance at roughly 4.6:1.9:1
- Debugging metrics: ~35% verification failure rate, 42% quick fixes vs 42% six-plus-iteration spirals
- Mature-codebase edit:write ratio ~9:1 and 31% of sessions with 10+ reads before first edit
- 0.8 reads and 0.5 searches per code change (Sibyl)
- 2–3 edits between verifications (sweet spot)
Adoption & trust: 545 installs on skills.sh; 13 GitHub stars; 3/3 security scanners passed (skills.sh audits).
What problem does it solve?
Your agent sessions feel random—too few reads, noisy commits, or endless fix spirals—and you lack numbers for a sane verify-edit rhythm.
Who is it for?
Indie developers who want data-backed cadence for agentic implementation on mature or fast-moving repos.
Skip if: Greenfield one-file prototypes where strict verify-every-two-edits overhead outweighs benefit, or teams that only need language-specific API docs.
When should I use this skill?
When planning or running an implementation session and you want benchmark-backed read, edit, and verification cadence.
What do I get? / Deliverables
You align edits, verifications, and exploration steps with measured benchmarks so commits stay test-backed and spirals shrink.
- Session cadence plan aligned to benchmarks
- Verification checklist tuned to project maturity
Recommended Skills
Journey fit
Spans multiple journey phases - primary shelf plus alternate fits below.
Build/pm is the canonical shelf because the skill encodes how to plan and execute implementation loops, not a single integration or launch tactic. Benchmarks inform task ordering, when to verify, and batch size—planning discipline for coding agents and humans alike.
Where it fits
Before a feature branch, you set a rule: explore with reads until baseline met, then verify every 2–3 edits.
Pre-release you rebalance checks toward the documented typecheck:lint:test mix instead of only running tests.
After a six-iteration bug spiral, you reset session context using quick-fix vs spiral benchmarks.
How it compares
Use as evidence-based session choreography instead of ad-hoc “just keep prompting until green.”
Common Questions / FAQ
Who is implement for?
Solo builders and tech leads optimizing how often their coding agent reads, edits, and runs checks during real feature work.
When should I use implement?
During build pm when scoping an implementation loop; during ship testing when tightening verify ratios before release; during operate iterate when post-morteming a debugging spiral.
Is implement safe to install?
It is read-only methodology with cited internal benchmarks—review the Security Audits panel on this page; it does not execute repo changes by itself.
SKILL.md
READMESKILL.md - Implement
# Implementation Benchmarks & Archetypes Quantitative data from 21,321 tracked operations across 64+ projects. ## Quantitative Benchmarks | Metric | Value | Source | | ----------------------------------------- | --------- | ------------------------ | | Reads per code change | 0.8 | Sibyl (21k ops) | | Searches per code change | 0.5 | Sibyl | | Edits between verifications (sweet spot) | 2-3 | Sibyl + v2 | | Changes per commit | 48.7 | Sibyl | | Verifications per commit | 23.2 | Sibyl | | Typecheck:Lint:Test ratio | 4.6:1.9:1 | Sibyl | | Edit:Write ratio (mature codebase) | 9:1 | v2 | | Verification failure rate | ~35% | v2 | | Post-fix verification rate | 27% | Debugging (612 sessions) | | Quick fix rate (1-2 iterations) | 42% | Debugging | | Spiral rate (6+ iterations) | 42% | Debugging | | Context overflow in spirals | 99% | Debugging | | Sessions with tests | 90% | Sibyl | | Exploration before first edit (10+ reads) | 31% | v2 | | Agent survivorship in research swarms | 67-100% | Config/skills | --- ## Implementation Archetypes ### The Quick Fix ``` Read error -> Grep for pattern -> Read 2-3 files -> Edit 1-2 files -> Verify -> Commit ``` - **Budget:** 1-2 cycles, 5-20 edits - **Verification:** One typecheck + one test run - **When:** Bug with clear error, config change, typo fix ### The Feature Build ``` Orient (read existing patterns) -> Plan (task list) -> Implement layer-by-layer -> Verify per layer (typecheck + lint) -> Tests -> Commit ``` - **Budget:** 5-15 cycles, 50-200 edits - **Verification:** Typecheck per layer, tests at end - **When:** New endpoint, new UI component, new service integration ### The Research-First Build ``` Dispatch 3-7 research agents -> Synthesize findings -> Write spec/plan -> Implement in dependency order -> Verify -> Cross-model review -> Commit ``` - **Budget:** 10-30 cycles, 200-500 edits - **Verification:** Per-wave verification + final cross-model review - **When:** Unfamiliar domain, greenfield feature, technology evaluation needed ### The Parallel Epic ``` Research swarm -> Task graph with dependencies -> Wave dispatch (3-7 agents per wave) -> Collect outputs -> Integration -> Full verification suite -> Commit ``` - **Budget:** 30-100+ cycles, 500-1000+ edits - **Verification:** Per-agent verification + integration verification + full suite - **When:** Multi-system feature, major refactor, new project inception --- ## Context Engineering Budget | Item | Tokens | | --------------------------- | ----------------- | | Baseline system + CLAUDE.md | ~20k | | Usable budget (200k window) | ~180k | | Target utilization | 40-60% | | Context rot threshold | ~15-20 iterations | ### Context Preservation Strategies - **Subagent delegation:** Research and verbose operations in separate context windows - **Just-in-time loading:** Glob/Grep to discover, Read on demand (don't pre-load everything) - **`/clear` between unrelated tasks:** Kitchen-sink sessions kill performance - **After compaction:** Always preserve modified file list and test commands --- ## The Ralph Loop (Edit-Verify-Fix Cycle) 346 cycles detected across 27/30 Sibyl sessions: | Profile | Pattern | Frequency | | ---------------- | ---------------------------------- | -------------- | | **Quick fix** | 3 changes -> LINT