autogap

The first Claude Code skill that diagnoses the top 3 gaps in any project — code or not.

Give autogap a project directory. It classifies the project type from file signals (not assumptions), infers the top-value objective the project is reaching for, and reports the 3 biggest gaps blocking that objective plus 3 macro-steps to close them. Then it hands you a numbered menu and waits. If you pick, it plans and executes autonomously.

It works on:

• source code — single package or monorepo
• documentation sets — research notes, spec folders, knowledge bases
• strategy / business — roadmaps, OKRs, decks, briefs
• creative / content — drafts, outlines, edit lists
• ops / infra — runbooks, playbooks, inventories
• hybrid — whatever you've actually got

Most agent skills assume software. autogap doesn't.

30-second example

$ cd research-project/
$ claude
> /autogap

# Project Objective Hypothesis
- Publish a peer-reviewed paper on X by Q3 2026
- Project type: research
- Confidence: High
- Signals: 3 draft sections, referee feedback letter, 47-entry BibTeX,
  methods/ folder with 2 Jupyter notebooks, reproducibility/ empty

# Top 3 Gaps
## Gap 1 — Methods section incomplete
  Why it matters: editors reject without reproducible methodology.
  Evidence: methods.md is 40% length of Background + has 3 TODO markers.
  Consequence: guaranteed R&R loop, likely 3-month delay.

## Gap 2 — No reproducibility kit
  Why it matters: target journal requires data + code on acceptance.
  Evidence: reproducibility/ directory empty; notebooks have no seeds.
  Consequence: acceptance delayed until kit is produced.

## Gap 3 — Discussion skirts disconfirmation
  Why it matters: weakest reviewer target; editor likely flags.
  Evidence: 0 mentions of limitations; no alternative-explanation section.
  Consequence: weakest part of the paper, most reviewer ammunition.

# 3 Macro-Steps
## Macro-Step 1 — Complete methods from lab notes (Low / High / High)
## Macro-Step 2 — Produce reproducibility kit  (Medium / High / Medium)
## Macro-Step 3 — Add disconfirmation to discussion (Low / Medium / High)

Choose what to implement:
1. Macro-Step 1 only
2. Macro-Step 2 only
3. Macro-Step 3 only
4. Macro-Steps 1 + 2
...

Pick a number. autogap plans, executes, verifies, and reports exactly what changed.

No user confirmation on every step. No fake completion. Stops cleanly on ambiguous irreversible moves.

Why it's different

Full thinking in docs/design-rationale.md.

Install

macOS / Linux:

curl -fsSL https://raw.githubusercontent.com/micpet7514088/autogap/main/install.sh | bash

Windows (PowerShell 7+):

Invoke-RestMethod https://raw.githubusercontent.com/micpet7514088/autogap/main/install.ps1 | Invoke-Expression

Restart Claude Code. From any project directory: claude → /autogap.

To update after a new release, clone the repo once and use the --update / -Update flag:

git clone https://github.com/micpet7514088/autogap.git ~/autogap
bash ~/autogap/install.sh --update     # macOS / Linux
# .\autogap\install.ps1 -Update        # Windows

Clone + symlink (updates propagate automatically on pull):

git clone https://github.com/micpet7514088/autogap.git ~/autogap
bash ~/autogap/install.sh

Project-level skill (per-project, no clone required):

mkdir -p .claude/skills/autogap
curl -sSL https://raw.githubusercontent.com/micpet7514088/autogap/main/skills/autogap/SKILL.md \
  -o .claude/skills/autogap/SKILL.md

File copy: Copy skills/autogap/SKILL.md into any directory Claude Code scans for skills.

Examples you can run

Three demo projects, one per major project type. Each one has an expected-output file so you can compare what autogap produces against what it should produce.

• examples/demo-01-code/ — Python CLI blocked on PyPI release → expected output
• examples/demo-02-docs/ — CIKM 2026 paper, 50 days to deadline → expected output
• examples/demo-03-strategy/ — Q3 product launch with top-3 LOI contingency → expected output

See examples/ for the full set and the run instructions.

Case studies

Real diagnoses on real projects. Some are anonymized; one is a fully public post-mortem.

• Accessibility audit on two production sites — 88/95 → 100/100 in one day
  Diagnosis, fix, deploy, certification. Includes a real drift incident + recovery. Fully public, real commit SHAs.
• NIS2 incident response audit on a European insurer
  Independent audit, 3 parallel workstreams, DR failover error caught in tabletop. Anonymized.
• AI Coach fine-tuning for a healthcare vertical — retention +17%
  Response-variance + semantic fix + within-session latency bug. Anonymized.

See docs/case-studies/ for the full set.

How it works, briefly

1. Classify project type from file signals (folder tree, file types, naming, metadata). No software default.
2. Infer the top-value objective the project is trying to reach. State confidence, name the strongest alternative hypothesis.
3. Scan breadth-first. Deepen on the richest paths. Sample ≥1 artifact per distinct family (code / docs / config / media). Run a disconfirmation pass before closing.
4. Rank gaps by blocker strength, not cosmetic polish. "Nice-to-have" is not a gap.
5. Present a 3-gap / 3-macro-step menu. Stop. No auto-execute.
6. On pick: plan tight, execute autonomously, verify proportional, report compactly with 3 next-steps.

Architecture choices, trade-offs, and the "why 3 (not 5, not 7)" question: docs/design-rationale.md.

Contributing

See CONTRIBUTING.md. Short version:

• Real use cases welcome, theoretical improvements discouraged.
• One change per PR, closing one concrete gap.
• No SaaS wrappers. No scope creep "just in case".

Issues and PRs: https://github.com/micpet7514088/autogap.

License

MIT — use freely, attribution appreciated.

Authors

Built by the team behind:

• Synergix — fractional IT direction and AI governance for European finance and insurtech mid-caps. 20+ years of enterprise delivery applied to AI programs.
• Arenia — AI coaching avatars for corporate training. Realistic, scalable practice with measurable KPI frameworks.

autogap was battle-tested on our own client work before we open-sourced it.